Agent Harness & Coding Tooling
DeepSeek Harness (dsh), DeepSeek TUI, and Reasonix: plugin-based agent architecture built on Cordis, session logs, sandboxing, subagent bridging, and competition with Claude Code/Codex.
22%
Best tweets about DeepSeek
50 Best Tweets About DeepSeek (2026)
Explore the best tweets about DeepSeek, including model releases, reasoning, coding, benchmarks, deployment, and open-model discussions. Updated every two weeks.
Technical DeepSeek evaluations, releases, deployment experience, and comparisons supported by concrete evidence.
- Creators
- 38
- Updated
Original Xholic analysis
What 50 top DeepSeek posts reveal
Across 50 posts from 38 creators, this conversation centers on technical DeepSeek evaluation rather than pure hype: 94% supportive stance, 92% positive sentiment, and 82% of posts carrying media. DeepSeek's own V4 Preview announcement (2047516922263285776) anchors attention at roughly 95x the median engagement score. Discussion clusters around open-source infrastructure releases, API pricing economics, agent tooling, local deployment, and head-to-head benchmarking against Claude and GPT-class models. Only 3 of 50 posts are critical, and the two critical posts covered by the evidence both target pricing changes rather than model capabilities (2088677834910748815, 2088924161959927819).
- Dominant tone
- Positive
- Median score
- 40.9
- Leading format
- Announcement
- Recent posts
- 40%
92% of posts
All-time engagement
68% of posts
Published in 90 days
Conversation map
The themes creators return to
Benchmarks & Head-to-Head Evaluations
Concrete evaluations of DeepSeek models against Claude, GPT, GLM, and each other — leaderboard placements (Design Arena, AI Intelligence Index), agentic coding scores, blind tests, and hands-on build comparisons.
14%
Local & DIY Deployment
Running DeepSeek models on consumer hardware — Macs, single GPUs, CPUs — with quantization, offloading, throughput numbers (tok/s), and local inference engines like ds4.
14%
New Model Releases & Specs
Announcements of DeepSeek model launches (V4, V4.1-Flash, OCR, variants) with parameter counts, context windows, modality, license, and headline capabilities.
14%
API Pricing & Cost Economics
Token pricing, price changes (cuts and hikes), cost comparisons vs Western labs, cache-hit pricing, and the cost-efficiency argument for Chinese labs (vertical integration, subsidized energy, funding structure).
12%
Architecture & Efficiency Research
Technical breakdowns of DeepSeek's model and systems innovations: sparse attention (CSA, HCA, DSA), MLA/KV-cache compression, MoE routing, Engram memory, FP4/FP8 training, Muon optimizer, speculative decoding (DSpark), and DeepGEMM kernels.
10%
Open-Source Infrastructure Releases
DeepSeek's open-sourcing of non-model artifacts — weights under MIT, DeepSpec, DeepGEMM, speculative decoding repos — and commentary on open-source philosophy and compounding between labs (e.g., Kimi cross-pollination).
8%
Funding, Valuation & Company Strategy
DeepSeek's first outside fundraising rounds, investor terms, valuation, Liang Wenfeng's research-first stance, talent poaching, and competitive pressure from ByteDance and US labs.
6%
Tone and stance
SentimentPositive leads
Author postureSupportive leads
Performance benchmark
- Median likes
- 175
- Median reposts
- 17
- Median replies
- 12
- Median views
- 15.2K
Posts with media make up 82% of this collection. Their median all-time score is 46.1, compared with 14.1 for text-only posts.
Format mix
- Announcement68% · score 22.9
- List18% · score 34.0
- Opinion10% · score 112.1
- Tutorial4% · score 114.1
Consensus and debate
Where creators agree, and where they do not
Shared view
DeepSeek's own announcements dominate engagement
The two top engagement outliers are official DeepSeek posts: 2047516922263285776 (DeepSeek-V4 Preview, allTimeScore 3891.45, 95.08x the median) and 2097930608790167907 (V4.1-Flash, 2441.66, 59.65x). The deepseek_ai account's medianAllTimeScore of 3166.56 leads all 38 creators, and ANNOUNCEMENT is the dominant format at 68% share.
Shared view
Open-source infrastructure earns the highest theme-level engagement
The Open-Source Infrastructure Releases theme shows the highest theme medianAllTimeScore at 191.8. 2070928299744972814 covers DSpark speculative decoding plus the DeepSpec training framework at 535 (13.07x median), and 2105633301625282924 covers MIT-licensed DeepGEMM kernels for Hopper and Blackwell GPUs at 262.96 (6.42x median).
Shared view
Pricing economics drives high engagement
API Pricing & Cost Economics carries a theme medianAllTimeScore of 96.64, the second-highest after open-source releases. 2048675097687294264 reports a 90% API price cut and Claude Code compatibility; 2049072308900200757 pegs V4 Pro at ~$0.87 per 1M tokens versus Opus 4.7's $25 (~28x cheaper); 2050607438433567021 explains the cost gap through vertical integration and subsidized energy.
Shared view
Agent harness tooling is the largest thematic cluster
Agent Harness & Coding Tooling is the biggest theme at 22% share (11 tweets). 2099800913456365613 describes dsh's everything-is-a-plugin design and ~210,000 stars in developer preview; 2089664601478844825 counts 148,000 stars in 4 days; 2088317043741491699 covers dsh's plugin architecture as a Claude Code/Codex competitor, and 2083927091507314698 documents ds4, a local MIT-licensed inference engine from the Redis作者
Open debate
Critical posts concentrate on price hikes, not model quality
Only 3 of 50 posts (6%) are critical. 2088677834910748815 flags V4 Flash output rising from $0.28 to $1.32 per million tokens at peak hours; 2088924161959927819 questions a 6x–12x cache-hit input price increase shipping the same week as the DeepSeek Harness release, noting cached input is exactly what agent loops consume.
Open debate
Blind tests show capability does not equal preference
2073462000961597529 ran a labeled-blind comparison in which DeepSeek V4 Pro won 4 of 4 against Claude Opus 4.8, yet the author still chose Claude. Separately, 2084995338122002633 found V4 Flash outbuilding V4 Pro across 50+ side-by-side builds while acknowledging inconsistent physics and controls. Sentiment overall is 92% positive with a single MIXED post.
What performs
Patterns behind standout posts
Local deployment evidence is concrete and quantified
Local & DIY Deployment shows a theme medianAllTimeScore of 55.099, well above the overall median of 40.93. 2106373861818224907 reports 2000 tok/s prefill and 29 tok/s decode for prose (41 tok/s for code) on a local V4.1-Flash build; 2036154648348991985 runs DeepSeek-V3 (671B parameters, 355 GB of weights) at 1.4 tok/s on a MacBook Pro with a single M4 Max; 2106079556436214251 documents CPU inference at 20 tok/s.
Benchmarks are strong but caveats appear in-thread
2111390565635969305 reports V4.1-Flash-1017 beating Opus 5 on DeepSWE (77% vs 74%), AutomationBench (60.5% vs 50.3%), and Apex Bench (72.1% vs 65.8%), with MIT-licensed weights. 2098489852027375784 cites a 39-position jump to 6th on Design Arena. 2083995230416830738 cautions that some comparisons use DeepSeek's own harness and internal datasets.
Architecture research themes skew technical and cited
Architecture & Efficiency Research holds a 10% share with a theme medianAllTimeScore of 62.135, above the overall median. 2011468534887469448 describes the Engram O(1) memory module with BBH +5.0 and long-context retrieval rising from 84% to 97%; 2048566818118545887 breaks down V4's CSA/HCA hybrid attention, FP4 training, and MoE routing; 2040654624093253707 covers the HISA hierarchical sparse attention upgrade.
A fresh, creator-diverse conversation
Posts span 2026-01-14 to 2026-10-17 with a median age of 110.3 days; 40% fall within 90 days. 38 creators produced 50 posts, with the top five voices holding a 20% placement share. At the format level, TUTORIAL posts (4% share) carry a medianAllTimeScore of 114.11 and OPINION posts 112.086, both well above the ANNOUNCEMENT median of 22.95.
Statistical standouts
- View standout post 1Score 3891.4 · 95.08× median
- View standout post 2Score 2441.7 · 59.65× median
- View standout post 3Score 535.0 · 13.07× median
- View standout post 4Score 379.1 · 9.26× median
- View standout post 5Score 263.0 · 6.42× median
Creator landscape
Who shapes this conversation
The five most represented creators account for 20% of the selected posts.
1. 0xSero
@0xSero
2 posts
2. David Ondrej
@DavidOndrej1
2 posts
3. DeepSeek
@deepseek_ai
2 posts
4. Hasan Toor
@hasantoxr
2 posts
5. ℏεsam
@Hesamation
2 posts
6. Nav Toor
@heynavtoor
2 posts
Since the previous snapshot
What changed since Oct 15, 2026
- 80% of the selected posts remained.
- The creator count changed by -2.
- The leading sentiment remained stable.
How this analysis was made
Themes, sentiment, stance, and post format are classified per tweet. All counts, shares, medians, creator concentration, freshness, and performance comparisons are then calculated directly from the published snapshot.
Xholic's all-time score compares engagement while accounting for reach, post age, and creator consistency. It is used for relative comparisons within this collection.
This report analyzes the exact 50-post snapshot shown below. AI identifies editorial categories and drafts explanations; all statistics are calculated from the snapshot, and every narrative claim is checked against cited posts before publication.
Best DeepSeek tweets
Top DeepSeek tweets from 38 creators
Ranked 01–50
- 01
@deepseek_ai ·
🚀 DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length. 🔹 DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models. 🔹 DeepSeek-V4-Flash: 284B total / 13B active params. Your fast, efficient, and economical choice. Try it now at https://t.co/GCdiMzk1Dl via Expert Mode / Instant Mode. API is updated & available today! 📄 Tech Report: https://t.co/drlDrxkYtp 🤗 Open Weights: https://t.co/T13Y8i7SDM 1/n
- 1.7KReplies
- 7.5KReposts
- 45.5KLikes
- 10.1MViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 02
@deepseek_ai ·
🚀 Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient. 🔹 Introducing the smallest model in our new architecture family, with native visual understanding. 🔹 Designed for greater capability, faster inference, higher throughput, and scaling to larger models. 1/6
- 1.1KReplies
- 3.1KReposts
- 28.8KLikes
- 7.2MViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 03
@Yuchenj_UW ·
DeepSeek is the GOAT. 🐳 They just published DSpark, a new speculative decoding method that boosts throughput by 51% to 400%. They also open-sourced DeepSpec, the training framework behind it. This is the real open AI.
- 105Replies
- 458Reposts
- 3.7KLikes
- 368.5KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 04
@0xSero ·
Deepseek-V4.1-Flash-2-Sparks is out! It took nearly a month to get this done right - 2000 tok/s prefill - 29 tok/s decode prose - 41 tok/s decode code - 2 million kV cache pool - 262k context window - 8 concurrency - vision - 92% top token agreement - 0.07 KLD - load with pi
- 59Replies
- 84Reposts
- 1.9KLikes
- 134.8KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 05
@simplifyinAI ·
DeepSeek has open-sourced a library that speeds up the GPU math behind AI models. It's called DeepGEMM. It optimizes the matrix calculations that power large language models, helping developers get more performance from NVIDIA Hopper and Blackwell GPUs. → Supports FP8, FP4, and BF16 calculations → Optimizes workloads for mixture-of-experts models → Overlaps GPU communication and computation → Compiles kernels at runtime, with no CUDA compilation during installation DeepSeek reports performance that matches or beats expert-tuned libraries across a range of matrix sizes. Free. Open Source. MIT licensed.
- 15Replies
- 57Reposts
- 537Likes
- 23.5KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 06
@trikcode ·
DeepSeek cut their API prices by 90% and it plugs directly into Claude Code. US labs are charging more every quarter. Chinese labs are charging less every week. one of these companies understands developers.
- 68Replies
- 81Reposts
- 1.2KLikes
- 21.6KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 07
@_avichawla ·
Fine-tune DeepSeek-OCR on your own language! (100% local) Most vision models treat documents as massive sequences of tokens, making long-context processing expensive and slow. DeepSeek-OCR uses context optical compression to convert 2D layouts into vision tokens, enabling efficient processing of complex documents. It is a 3B-parameter vision model that achieves 97% precision while using 10x fewer vision tokens than text-based LLMs. In fact, you can easily fine-tune it for your specific use case on a single GPU. I used Unsloth to run this experiment on Persian text and saw an 88.26% improvement in character error rate. ↳ Base model: 149% character error rate (CER) ↳ Fine-tuned model: 60% CER (57% more accurate) ↳ Training time: 60 steps on a single GPU Persian was just the test case. You can swap in your own dataset for any language, document type, or specific domain you're working with. I've shared the complete guide in the next tweet, which includes the code, notebooks, and environment setup ready to run with a single click. Everything is 100% open-source!
Watch video- 15Replies
- 81Reposts
- 489Likes
- 35.9KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 08
@DavidOndrej1 ·
DeepSeek V4 just dropped and it's 40x cheaper than GPT-5.5 Pro 1.6 trillion parameters, trained on limited GPUs, still matches the top labs I built 4 full apps with it to see if the hype is real
Watch video- 66Replies
- 56Reposts
- 1KLikes
- 109.8KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 09
@alex_verem ·
DeepSeek is about to end the entire coding agent industry with this repo. It's called DeepSeek Harness, dsh for short. One command in the terminal gives you a local web app with an agent that runs tools, edits files, and keeps a session going. That part isn't new. Coding agents are everywhere. What's different is the rule the whole thing is built on. Everything is a plugin. That doesn't mean it supports plugins. It means every part of it is one. The model adapter is a plugin. The tool registry is a plugin. The session log is a plugin. The loop that actually runs the agent is a plugin. There's no protected core, so any piece can be swapped from a config file without touching the rest. In practice the model becomes a setting. Point it at DeepSeek, point it at a local model, point it at a competitor's API, the harness doesn't care. The sandbox works the same way. Point it at a remote machine and the shell, the terminal, and the language server follow it. The second rule is the one I'd steal. Anything the model gets to see has to be written to the session log first, and the code checks that while it runs. There's no hidden context and no injected instruction that doesn't show up in the record. Fork a session, replay it, audit it, all from the same log. It's a developer preview and they say so in capital letters. Breaking changes are coming, and there's a safety notice you're supposed to read before running it. It sits around 210,000 stars and nearly 15,000 commits, which for a preview is a lot of people betting on the design. DeepSeek helped make open weights normal. Now it wants the agent around the model to be every bit as replaceable.
Watch video- 30Replies
- 38Reposts
- 393Likes
- 46.9KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 10
@hasantoxr ·
I'm uninstalling Cursor and every other AI coding tool after finding this. It's called DeepSeek TUI. A full coding agent that runs in your terminal on DeepSeek's API. One command to install. No browser. No IDE plugin. No subscription. npm i -g deepseek-tui Here's what it does: → Edits files across your entire codebase → Executes shell commands → Browses the web with web run → Handles git operations → Resumes sessions after you close the terminal → Connects to any MCP server → Spins up as an HTTP API with deepseek-tui serve --http Three modes depending on how much control you want: → Plan: see the full execution plan before the agent touches anything → Agent: interactive multi-step tool use → YOLO: auto-approves every tool call in a trusted workspace DeepSeek's API is already one of the cheapest frontier models available. Now you have a full coding agent on top of it. Built in Rust. 29 releases. Active development. 100% Opensource. MIT License.
- 48Replies
- 43Reposts
- 320Likes
- 27.2KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post.
- 11
@Hesamation ·
DeepSeek-V4 uses the Muon optimizer with Kimi's recipe to scale it for LLM training. meanwhile, Kimi K2 (and K2.6) uses DeepSeek-V3's architectural techniques (ultra-sparse MoE + MLA). open-source AI labs are compounding on each other's research, and it's the way it should be.
- 25Replies
- 86Reposts
- 830Likes
- 33.9KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 12
@LiorOnAI ·
DeepSeek is back. They just figured out how to make AI models way smarter without adding compute. They built Engram, a memory module that retrieves information instantly with O(1) lookup. 𝗛𝗼𝘄 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀 Think of it as giving your model a lookup table. Instead of burning compute simulating memory retrieval, it just grabs what it needs. Early layers stop wasting effort on pattern matching. The model goes deeper on actual reasoning. 𝗪𝗵𝗮𝘁 𝘆𝗼𝘂 𝗴𝗲𝘁 Huge wins across benchmarks on the same compute: > BBH +5.0 > MMLU +3.4 > HumanEval +3.0 𝗧𝗵𝗲 𝗯𝗶𝗴 𝗱𝗲𝗮𝗹 Long-context retrieval jumped from 84% to 97%. Lookups handle local patterns, so attention focuses on global context.
- 26Replies
- 72Reposts
- 491Likes
- 29.6KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 13
@Hesamation ·
you don’t realize how CHEAP DeepSeek is until you use it all day and pay the price of a bag of chips. DeepSeek V4 Pro: ~$0.87 / 1M tokens Opus 4.7: $25 / 1M tokens that's ~28x cheaper. Opus is stronger on benchmarks, but at this price gap, stronger has to be A LOT stronger.
- 52Replies
- 59Reposts
- 753Likes
- 36.6KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 14
@trikcode ·
Chinese models were supposed to be cheaper DeepSeek was supposed to kill OpenAI's pricing. but starting tomorrow, DeepSeek V4 Flash output goes from $0.28 to $1.32 per million tokens at peak hours and then V4 Pro lands at $3.96. I still can't believe this..
- 137Replies
- 23Reposts
- 956Likes
- 105.6KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 15
@DesignArena ·
BREAKING: DeepSeek‑V4.1‑Flash takes 6th overall on Design Arena with an Elo of 1347! This marks a 39-position jump over the next-highest DeepSeek model - and DeepSeek’s return to a top-10 placement on Design Arena. The model ranks in the same performance band as Claude Fable 5.1 on real-world frontend design tasks. Congratulations to the @deepseek_ai team!
- 40Replies
- 44Reposts
- 719Likes
- 59.9KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 16
@TimJayas ·
DeepSeek just EMBARRASSED Claude Opus 4.7 Just switched to DeepSeek V4 Pro for a few days Cost: DeepSeek V4 Pro = $2.02 Claude Opus 4.7 = $265.21 Same quality for most of the medium tasks with no noticeable difference in output I know it’s hosted in China with cheap electricity but at this point western labs are getting cooked on price
- 78Replies
- 44Reposts
- 566Likes
- 52.3KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 17
@quxiaoyin ·
Why is @deepseek_ai 100x cheaper than @AnthropicAI? China is vertically integrated to be cheap. → Cheap model: token-optimized, aggressive caching, less GPU per query → Cheap chips: Huawei silicon, no Nvidia tax → Cheap energy: subsidized power, state-scale grid → Cheap talent: top researchers at a fraction of US salaries → Cheap economics: DeepSeek funded by trading profits, inference doesn’t need to make money The only thing they lag is performance. But as they get “good enough”, being cheap matters a lot, and being frontier keeps getting harder.
- 35Replies
- 45Reposts
- 373Likes
- 19.4KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 18
@TheTuringPost ·
DeepSeek-V4 is a full-stack redesign of LLMs around long context + efficiency Here are some of the changes: - Hybrid attention: Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA) for long-context efficiency - 1M-token context becomes ~3–10× cheaper in memory and compute - New residual connections (mHC) + Muon optimizer help to train trillion-scale models without constant collapse - FP4 quantization is used during training (not only inference) - More efficient Mixture-of-Experts model with 1.6T params, but only ~49B active per token is achieved via improved roughing + light sequence-level regularization + hash routing in early layers Post-training recipe is also different - train specialists (math, code, agents) - merge via distillation (OPD) - cleaner and more scalable than RLHF All of this is about making models smarter per FLOP + per token
- 7Replies
- 45Reposts
- 237Likes
- 12KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 19
@HuggingPapers ·
DeepSeek Sparse Attention gets a hierarchical upgrade HISA replaces the flat token scan with a two-stage block-then-token filtering pipeline, eliminating the indexing bottleneck at 64K context without any additional training.
- 4Replies
- 43Reposts
- 333Likes
- 55.4KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 20
@danpacary ·
Update: DeepSeek-V3 running at 1.4 tok/s on my MacBook Pro. 671 billion parameters. 355 GB of weights. One M4 Max, no cloud. Here's some of the system...
- 22Replies
- 18Reposts
- 359Likes
- 126.5KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 21
@_xjdr ·
i wanted to understand DeepSeekv4.1 Flash a bit better, so i tried to implement it from scratch in a very hostile environment. I was able to get it to run at between 1 tok/s and 10 tok/s bs=1 ctx=1M+ at full precision (which is a combo of bf16, fp8 and mxfp4) depending on the expert cache hits on a single 4090, 64GB of CPU ram and 1TB of nvme . this architecture slaps and sadly is optimal for grace blackwell type deployments . im incredibly impressed with the deepseek teams continued engineering work within their constraints.
- 9Replies
- 8Reposts
- 277Likes
- 12.5KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 22
@RoundtableSpace ·
DeepSeek just open-sourced a full coding agent harness for free that directly replaces Claude Code's $200 a month plan. → Everything is a plugin, model adapters, tools, session logs and the agent loop itself, all swappable from config → Built on Cordis, DeepSeek's own plugin runtime → One command to spin up a local Web UI: npx @deepseek-ai/dsh web → Ships with real architecture docs and an AGENTS.md written for agents to read the codebase themselves → TypeScript, MIT, developer preview with breaking changes expected https://t.co/OYDwiQF6G2
- 27Replies
- 29Reposts
- 294Likes
- 65.1KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 23
@DavidOndrej1 ·
After 14 months it's finally here... Deepseek v4 It's cheap and goes head to head with GPT and Opus. In this 29 min video i'll show you everything you need to know about it: timestamps: 00:00 Introduction: DeepSeek V4 Arrives 00:49 Specs: 1.6 Trillion Parameters & Architecture 01:55 The GPU Story: China's Hardware Limitations 03:09 Benchmarks vs GPT-5.5 and Claude Opus 4.7 04:39 Agentic Coding & Terminal Performance 06:29 Pricing: 40x Cheaper than GPT-5.5 Pro 09:05 5 Major Weaknesses of DeepSeek V4 10:43 Censorship & Tool-Calling Constraints 11:57 Setup: Using DeepSeek V4 via OpenCode 14:33 Build 1: Interactive Architecture Explainer 16:11 Build 2: SVG Plant Animation Demo 17:02 Build 3: Full Arcade Karting Game 18:10 Build 4: Exoplanet Data Visualization 25:24 Final Results & Performance Review 28:50 Cost Analysis: Parallel Agents for Pennies
Watch video- 9Replies
- 10Reposts
- 136Likes
- 8.4KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 24
@HuggingPapers ·
DeepSeek just released DeepSeek-V4.1-Flash A 552B MoE multimodal model with 1M-token context that compresses the KV cache to 890 bytes per token, slashing deployment costs for long-context agents.
- 7Replies
- 24Reposts
- 247Likes
- 16.8KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 25
@kimmonismus ·
DeepSeek is moving ahead with a $10.29 billion financing round. Good news: Liang Wenfeng, however, remains focused on building open-source AI models rather than chasing short-term commercialization. Via Bloomberg
- 25Replies
- 25Reposts
- 371Likes
- 14.2KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 26
@0xSero ·
How I got deepseek running on cpu at 20 tok/s I'm learning as well so lots wrong, correct me if u catch it. we go through: - low bit quantisation - cpu offloading - exl3/trellis - costs
Watch video- 12Replies
- 6Reposts
- 125Likes
- 8.8KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 27
@heynavtoor ·
Every month, developers pay to run three models: DeepSeek V4 Flash: $0.14 in, $0.28 out per million tokens DeepSeek V4 Pro: $0.435 in, $0.87 out per million tokens GLM 5.2: $1.40 in, $4.40 out per million tokens GLM Coding Plan: $18 to $160 a month Cheap. Not zero. And every prompt goes to a server on the other side of the world. Now meet ds4. An inference engine that runs DeepSeek V4 Flash, DeepSeek V4 Pro, and GLM 5.2 directly on your own machine. No API key. No bill. No data leaving your laptop. 19,691 stars. 1,743 forks. MIT license. Pure C. 87 days old. The builder is Salvatore Sanfilippo. The same guy who wrote Redis. He got tired of pinging a server every time he wanted an answer and shipped his own local runtime. What it runs on: Apple Metal on Macs with 96 GB or more NVIDIA CUDA, single or multi-GPU, including DGX Spark AMD ROCm on Strix Halo SSD streaming for smaller machines Two-Mac RDMA tensor parallelism 100% offline Speed on a MacBook Pro M5 Max 128 GB, short prompt: 87 t/s prefill. 34 t/s generation. That is frontier open-model speed on a laptop you already own. Cost to run it. Zero. Forever. Your model. Your machine. Your data. MIT licensed. Pure C. Yours forever. (Link in the comments)
- 7Replies
- 16Reposts
- 96Likes
- 15.2KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 28
@DeRonin_ ·
DeepSeek just dropped a 5-page paper + free GitHub repo that makes any LLM respond 80% faster it's called speculative decoding. in plain english: Guess → Check → Keep → Repeat > Guess: a small fast model predicts the next few words > Check: the big smart model checks all guesses at once > Keep: lock in the right ones, fix the first wrong one > Repeat: do it again, 2-4x faster than normal DeepSeek's version (DSpark) is the first to combine two approaches that always had a tradeoff: getting both speed AND accuracy at the same time production results on their flagship model: 50% more requests handled, responses up to 80% faster how you can use it: > just using AI apps: expect them to get noticeably faster soon > building with AI: integrate this into your stack for instant 2x speedups > running your own models: clone the repo today and apply it to whatever model you run every other AI lab has been guarding this. now it's free on GitHub, MIT licensed this completely changed how I'm thinking about AI products this week one of the most powerful releases this month repo + papers below ↓
- 35Replies
- 8Reposts
- 122Likes
- 16.6KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 29
@kimmonismus ·
The company that proved you don't need billions to build world-class AI is now asking for money. DeepSeek is raising outside capital for the first time. The target: at least $300 million at a valuation north of $10 billion. Until now, founder Liang Wenfeng funded everything through his hedge fund High-Flyer Capital Management, famously rejecting offers from China's biggest VCs and tech giants. So what changed? For one, key researchers have been poached: Luo Fuli, a core V3 contributor, left for Xiaomi, and Guo Daya jumped to ByteDance for significantly higher pay. Meanwhile, the much anticipated V4 model has been delayed multiple times as engineers work to make it natively compatible with Huawei's Ascend chips, a move that could prove China's ability to build frontier AI without Nvidia. The competitive pressure is enormous. ByteDance's Doubao already surpassed DeepSeek in monthly active users, and tech giants on both sides of the Pacific are outspending startups at record pace. Even the efficiency pioneer needs ammunition now.
- 22Replies
- 21Reposts
- 190Likes
- 16KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 30
@ChrisGPT ·
Holy shit, DeepSeek just released V4.1-Flash-1017, and it’s competitive with Claude Opus 5 on several agentic benchmarks. It scores 77% on DeepSWE compared to Opus 5’s 74%, 60.5% on AutomationBench against 50.3%, and 72.1% on Apex Bench against 65.8%. Opus 5 still leads on several other benchmarks, but this is a pretty massive improvement over the previous V4.1 Flash. And yes, DeepSeek has already released the model weights under the MIT license!!
- 10Replies
- 6Reposts
- 160Likes
- 7.7KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 31
@hasantoxr ·
DeepSeek just did it again. They open-sourced their own agent harness that goes straight against Claude Code and Codex. Every coding harness today ships as a monolith. The prompt, the tools, the memory, the UI, all baked into a single binary. If you want to swap the model router or add a new tool, you're patching someone else's core. And when the maintainer changes the architecture, your fork breaks. DeepSeek Harness (dsh) inverts that. The core is tiny. Every capability loads at runtime as a plugin. Model connectors, tool schemas, the memory layer, and the web UI, all pluggable. It's built on Cordis, a runtime described in a paper called "A Programming Paradigm for Spatiotemporal Composability." Plugins load, swap, and unload while the agent is running. No restart. The repo ships with .claude, .agents, CLAUDE .md, and AGENTS .md folders. That's aimed squarely at anyone stuck inside Anthropic or OpenAI's harness who wants a way out. One command runs the full web UI locally: Check it out on GitHub: https://t.co/1xSz3OS8YH
- 13Replies
- 10Reposts
- 103Likes
- 15.2KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 32
@techNmak ·
125K+ GitHub stars in 3 days. DeepSeek Harness. It's an open-source agent harness. "Harness" here is simply the software around a model that lets it do work => give it access to tools, files and a shell, maintain sessions, apply permissions and sandboxing, and handle the loop between model responses and tool results. DeepSeek Harness is built on Cordis, and almost every major part of the system is a plugin. The model adapter, tool registry, session log and even the default agent loop are all replaceable pieces. A running agent is assembled through profiles and bundles, so you can change the model provider, add tools or introduce new policies without rewriting the rest of the runtime. Sessions are handled as an append-only event log. User messages, model output, tool calls and tool results are recorded as events, and the history shown back to the model is derived from that log. Forking, resume, transcripts, telemetry and persistence also build from the same event stream. Code Mode lets Harness expose tools through a code runtime. The model can write a small program that calls several tools, uses loops or runs independent operations concurrently. Those calls still pass through the normal Harness tool-execution pipeline and its policies. There is a lot more in the repo => subagents, workflows, MCP tools, skills, approvals, persistent terminals, background jobs, LSP-based code navigation, context compaction, goals, scheduled follow-ups and a Web UI. Not every configuration enables all of these, profiles decide which capabilities are assembled into a particular Harness setup. GitHub Repo: https://t.co/oj58ZcJ5bJ
- 8Replies
- 7Reposts
- 54Likes
- 4KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 33
@mark_k ·
DeepSeek V4 is live. Fully open weights, native 1M-token context, and a hybrid sparse attention architecture that makes it actually usable at scale. Efficient, powerful, and priced to shift the conversation. A notable step forward for open AI. @deepseek_ai
- 7Replies
- 3Reposts
- 136Likes
- 5.2KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 34
@mhdfaran ·
DeepSeek is coming for Claude Code. They just released DeepSeek Harness, an open-source agent system that can actually work inside your codebase. It can read files, edit code, run commands, plan tasks, and delegate work. But the wild part is the architecture: Everything is a plugin. Models, tools, agent loops, sandboxes, even the session system can be replaced. Fully open source. MIT licensed.
Watch video- 10Replies
- 28Reposts
- 65Likes
- 10.7KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 35
@CodeByPoonam ·
JUST IN: DeepSeek just raised $7.4 billion. its first ever outside funding. and investors got no voting rights. here's what investors agreed to: → capital invested through a limited partnership controlled by founder Liang Wenfeng → no direct stake in the company → zero voting rights → 5-year lock-up period → China's National AI Investment Fund participated anyway translation: give us $7.4 billion. we'll do what we want with it. see you in 5 years. and investors said yes. why? because DeepSeek has the leverage. the lab that shocked the world in January with R1 — a model that matched GPT-4 at a fraction of the cost — has been the most talked about AI story of 2026. the numbers: → $7.4 billion raised → $50 billion valuation → first outside funding in company history → Tencent and CATL among top backers what they'll use it for: → AI infrastructure expansion → R&D capabilities → employee equity → commercialization Liang Wenfeng previously said investors struggled to align with DeepSeek's long-term research horizons. rising compute costs and intensifying competition finally pushed them to raise. but they raised on their terms. entirely.
- 25Replies
- 15Reposts
- 108Likes
- 15.2KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 36
@VaibhavSisinty ·
DeepSeek just made their AI 5x faster. Without changing the model at all. It is called DSpark. And it is one of the most impressive pieces of system engineering I have come across. Here is the thing. The biggest bottleneck in running an AI model is not compute. It is memory. Reading the model weights from memory takes so long that generating 10 tokens in parallel costs almost the same as generating one. DSpark exploits this completely. Instead of generating one token at a time it drafts multiple tokens using a cheap smaller model first. Then verifies all of them through the main model in one pass. If the draft is right it skips ahead. If it is wrong it falls back. Most tokens are easy to predict. So most of the time it skips ahead. But what makes DSpark genuinely different is how it combines two approaches that nobody had successfully merged before. DFlash is fast at early token positions. Eagle is more accurate at longer sequences. DSpark runs both together and beats either approach on its own. And it does not draft a fixed number of tokens. It drafts more when the GPU is free and less when the server is busy. All on GPU. No CPU slowdowns. To be frank, this is not a model breakthrough. It is an infrastructure breakthrough. DeepSeek keeps finding ways to do more with less. That is the real moat.
- 7Replies
- 10Reposts
- 53Likes
- 4.1KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 37
@rohanpaul_ai ·
DeepSeek Harness reached 122K+ GitHub stars in 3 days, one of the fastest to rise in GitHub's history. It makes models, tools, loops, storage, scheduling and even the UI swappable plugins. its an orchestration layer for coding agent, runs as a local web app on a configurable port, and exposes every component as a swappable plug-in: shell access, file editing, web search, skills, sessions, model choice, reasoning effort, and permission scope, all editable in a YAML config. sub-agent support is the really special part. You can wire Claude Code or Codex in as plug-ins and let Harness route subtasks to whichever agent suits each step. It's MIT licensed, so you can add other providers or point it at a self-hosted model and never touch DeepSeek's API. However, their release timing for this is strange. They raised the price of cache-hit input by 6-fold to 12-fold in the same week it shipped an agent framework, and cached input is exactly what agent loops consume, since every step replays the same system prompt and accumulated history. Whoever set the API pricing either didn't coordinate with the harness group or deliberately priced the new workload higher. My read is the 2nd, because caching subsidies made sense when agents were rare and become the largest unpriced cost once a framework makes them routine.
- 3Replies
- 7Reposts
- 27Likes
- 2.7KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 38
@heynavtoor ·
148,000 stars in 4 days. DeepSeek dropped the open source rival to Claude Code. That is roughly one star every 2.5 seconds since the repo went public on August 13. Not 148,000 downloads. Not 148,000 signups. 148,000 people who went to a GitHub page and clicked the star button. The repo is called DeepSeek Harness. dsh for short. Released by the same Chinese AI lab in Hangzhou that sent Nvidia's stock tumbling in January 2025 with R1. Here is what dsh replaces: Claude Code. $20 a month on Pro. $100 on Max 5x. $200 on Max 20x. Locked to Anthropic models. DeepSeek Harness. $0. MIT licensed. Read every line. Fork it. Ship it in a product. Sell it. No permission needed. The tagline is three words: Everything is a Plugin. The model adapter is a plugin. The tool registry is a plugin. The sandbox is a plugin. The session log is a plugin. The agent loop itself is a plugin. Nothing is hardwired. Here is what dsh ships with today: - 4 built-in agent presets. standard, code, minimal, cordis. Pick one. Start building. - Web UI at http://127.0.0.1:3080 with one command: npx @deepseek-ai/dsh web - Full session logs. Fork a session. Replay a session. Every action captured. - OS-level sandboxing on all three platforms. Landlock on Linux. Seatbelt on macOS. ACL on Windows. Your agent cannot escape the box. - Subagent bridge to Claude Code and Codex. Both packages ship inside dsh. Your competitors become your plugins. - Full MCP support. Plug in any MCP server. The lead maintainer is Tianyi Cui with over 5,000 commits. Backed by DeepSeek's engineering team. Powered by the Cordis plugin runtime. Here is the interesting part. Anthropic ships Claude Code but has not open sourced the harness. Cursor is closed source. On June 16, 2026, SpaceX agreed to buy Cursor's parent company Anysphere for $60 billion in stock, the largest venture-backed startup acquisition in history, for scaffolding no one outside the company can read. DeepSeek is the first major AI lab to publish the full agent harness under a real open source license. One honest note. DeepSeek Harness is still in developer preview. Version 0.1. Release candidate 7 shipped this week. The README says there will be compatibility-breaking changes. If you need stability today, wait. If you want to see where AI tooling is going tomorrow, install it tonight. The scaffolding matters as much as the model. The people building agents in 2027 will not be the ones who trained the biggest LLM. They will be the ones who owned the harness. DeepSeek gave that harness away for free. Your model. Your tools. Your agent. (Link in the comments)
- 4Replies
- 13Reposts
- 30Likes
- 6KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 39
@quxiaoyin ·
@AnthropicAI’s business model just got disrupted by @deepseek_ai v4 pricing. Anthropic’s all about subscription locked-in: – Claude Code Max: $200/mo – Same usage via API: $5,000 – You still get rate limited; New limits added every few hours (not daily, not weekly — hours). If exceeding you will pay api pricing which can cost $100+ an hour. – Caps on parallel threads - can only use max in max not Hermes/openclaw – SDK requires the API, so devs eat the $5K DeepSeek: - do whatever. No subscription. 100x cheaper. No rate limit. No parallel limit. No usage limit. Whatever. The problem is, can Claude catch up? Yes Claude’s smarter still but most coding tasks DeepSeek is enough. Then what?
- 8Replies
- 5Reposts
- 43Likes
- 13.9KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 40
@The_AI_Investor ·
DeepSeek did it again. The new V4 Flash is not really a small model. It has 284B total parameters, but only activates 13B for each token. That makes it much cheaper and faster to run than its total size suggests, while still supporting a 1M-token context window. It is not beating the best frontier models overall, but it is getting surprisingly close. On DeepSeek’s published agent benchmarks, V4 Flash scored 82.7 versus Claude Opus 4.8’s 85.0 on Terminal Bench, and 25.2 versus 25.7 on Agents’ Last Exam. It still trails Opus more clearly on harder repository and cybersecurity tasks. Some results also use DeepSeek’s own agent harness and internal datasets, so comparisons should be treated carefully. The crazy part is the price: just $0.14 per million input tokens and $0.28 per million output tokens. For comparison, GPT 5.6 Sol costs $5 and $30, while Claude Opus 4.8 costs $5 and $25. That makes V4 Flash around 36 times cheaper on input and roughly 90 to 107 times cheaper on output. It is still too large for a normal laptop. DeepSeek’s official serving example uses four GB300 GPUs. But for AI clouds and large-scale agent workloads, this level of performance at this price could put even more pressure on frontier-model pricing.
- 5Replies
- 7Reposts
- 41Likes
- 11.7KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 41
@TheGeorgePu ·
I ran a blind test. DeepSeek V4 Pro vs Claude Opus 4.8. Same prompts, no labels, separate agents judging. Python. Writing in my voice. A bug-fix with trick questions. 4 out of 4, DeepSeek won. And I'm still choosing Claude. Capable isn't the same as the one you want to keep using.
Watch video- 9Replies
- 2Reposts
- 22Likes
- 1.8KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 42
@sabir_huss50540 ·
🚨BREAKING: Someone built a DeepSeek-native coding agent designed so your token bill stops climbing the longer you leave it running. It's called Reasonix. Most agents quietly break the model's cache every turn. The context shifts, the prefix invalidates, and the cost creeps up the longer the session lives. Reasonix is built around the opposite bet: keep DeepSeek's prefix cache warm so token cost stays flat across long sessions. Here's what it actually does: → Tuned around DeepSeek's prefix cache so token costs stay low over long sessions → One config file swaps in any OpenAI-compatible model — no hardcoded models, no new code → Runs two models together: an executor and a separate planner, each in its own cache-stable session → Plugins run as MCP-compatible subprocesses over stdio JSON-RPC → A single static Go binary that cross-compiles to six targets, with one dependency → Same local engine across CLI, desktop app, and VS Code Most coding agents are built to impress you in a two-minute demo. This one is built to survive a workday. 27.8k stars. 1.8k forks. MIT License. 100% Open Source. (Link in the comments)
- 7Replies
- 2Reposts
- 10Likes
- 890Views
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 43
@pankajkumar_dev ·
DeepSeek-V4 Drops: Open-Source Push Toward Cheaper, Long-Context AI - DeepSeek-V4-Pro is a 1.6T MoE model (49B active) trained on 33T tokens and released under a permissive MIT license - Efficiency gains: supports 1M-token context with 9.8× lower FLOPs and 9.5–13.7× smaller KV cache vs V3.2 - Benchmarks (vs GPT-5.4 & others): APEX (90.2% vs 85.9%), SWE Verified (80.6% on par with top models), Terminal Bench 2.0 (67.9% vs 65.4%), Codeforces (3060 rating, near top human level) - Head-to-head vs Anthropic Claude Opus 4.6: overall 53% win vs 37% loss (10% tie), stronger in task completion (98.3 vs 96.7) and content quality (83.3 vs 78.0) - Uses hybrid Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA) instead of quadratic attention - Enables long-context inference on smaller GPU clusters that previously couldn’t handle it - Variants: Pro (1.6T) and Flash (284B) with low-cost API pricing - Overall direction points toward efficient, open, agent-ready models rather than just scaling brute compute DeepSeek-V4 mainly stands out for making long context + agent workflows cheaper and more accessible without massive infrastructure.
- 1Replies
- 2Reposts
- 17Likes
- 1.1KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 44
@nvk ·
Here is a version of DeepSeek V4 Flash on a 128GB Mac MP w/ profiles for Claude Code & Codex https://t.co/176ez4kgcr claude-ds4 codex-ds4 "ok": prefill 28.96 t/s, gen 8.93 t/s 256-token: prefill 61.21 t/s, gen 38.57 t/s Warm weights: prefill 76.92 t/s, gen 37.81 t/s
- 3Replies
- 2Reposts
- 10Likes
- 2.4KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 45
@RoundtableSpace ·
DeepSeek V4 Flash scores 52 on the AI Intelligence Index for $72 to run the full test suite. GPT-5.6, Opus 5 and Fable 5 score the same and cost 10x more.
Watch video- 11Replies
- 2Reposts
- 51Likes
- 41.6KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 46
@_vmlops ·
DEEPSEEK JUST OPEN SOURCED AN AGENT HARNESS deepseek-harness (dsh) everything is a plugin ▫️ built on cordis, a runtime for spatiotemporal composability ▫️ ships a web ui, run instantly via npx @deepseek-ai/dsh web or build from source with pnpm ▫️ still in developer preview breaking changes expected open source agent tooling from deepseek keeps shipping https://t.co/ycbEoFw8UI
- 1Replies
- 3Reposts
- 4Likes
- 810Views
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 47
@AISystemGuy ·
Engineering update 🛠️ We’re adding native support for DeepSeek’s Multi-Head Latent Attention (MLA) and DeepSeek Sparse Attention (DSA) directly into our LLM inference runtime. The goal: break through the memory-bandwidth wall. A technical deep dive 🧵 1/ MLA: compressing the KV cache MQA and GQA reduce KV-cache size, but at the cost of representation capacity. MLA takes a different approach: Keys and Values are projected into a shared, low-dimensional latent space, and that compressed representation is cached instead of the full KV matrices. In simplified terms: d_c ≪ d_h × n_h This can dramatically reduce long-context memory pressure. 2/ The RoPE challenge Rotary Positional Embeddings are position-dependent, so they cannot be fully baked into the compressed latent vector. We split Keys and Queries into: • A compressed latent component for semantic context • A smaller position-aware component for RoPE Our CUDA kernels then combine these components mid-kernel. 3/ Reordering the math The updated kernels perform key matrix multiplications out of order, reconstructing only what is needed during attention. This avoids unnecessary intermediate tensors and extra memory transactions—critical when decode performance is limited by memory bandwidth. 4/ Adding DSA routing We’re layering DeepSeek Sparse Attention on top of the latent cache. A lightweight “lightning indexer” transforms and quantizes each incoming query, then scores it against the historical token cache. 5/ Dynamic top-K attention The runtime scheduler uses those scores to select the most relevant attention blocks. Low-affinity token connections are masked out, while only the top-K paths are computed. The result is a much smaller runtime compute graph. 6/ Why this matters MLA reduces the size of the data we need to move. DSA reduces how much of that data we need to process. Together, they help shift long-context inference away from being memory-bandwidth-bound and back toward a more compute-efficient execution profile. Early kernel tests are showing: • Major reductions in activation-memory overhead • Significant token-throughput gains • Stronger scaling on long-context sequences Benchmarks coming soon.
- 2Replies
- 1Reposts
- 4Likes
- 150Views
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 48
@rtodi ·
Interesting development in the AI space: China's DeepSeek is getting ready to launch its new V4 model — a capable multimodal AI that works with text, images, and video — and it's optimized to run on Huawei's latest domestic chips. This reflects China's ongoing push for tech self-reliance amid U.S. export restrictions. Huawei's new Ascend 950PR chip shows solid progress, with claims of strong inference performance (reportedly up to nearly 3x that of the restricted Nvidia H20 in some tasks), good memory, and competitive pricing. Big Chinese tech firms have placed large orders. That said, it still trails Nvidia's latest unrestricted high-end chipsets in raw power and overall ecosystem maturity. A notable step in building a parallel AI ecosystem.
- 0Replies
- 1Reposts
- 8Likes
- 1.9KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 49
@JulianGoldieSEO ·
DEEPSEEK JUST OPEN-SOURCED THE LAYER THAT ACTUALLY MAKES AI AGENTS USEFUL And the biggest feature isn’t the model. It’s what you can replace. What DeepSeek Harness changes: → The model is a plugin → Memory is a plugin → Tools and skills are plugins → The sandbox and file system are plugins → Even the think → act → check → repeat agent loop can be swapped Why that matters: ✓ MCP standardized how agents connect to tools. Harness takes that idea across the entire agent stack. ✓ It reads AGENTS.md and CLAUDE.md, so it can inherit instructions already used by Claude Code and other agents. ✓ It runs locally through a browser UI after a terminal command instead of forcing everything through a terminal interface. ✓ It launched under the MIT license, so developers can modify it or build products on top. The important lesson: Stop collecting AI tools. Build one agent system where Claude Code, Hermes, OpenClaw, DeepSeek Harness and whatever ships next can be swapped in without rebuilding everything. That’s a much better way to survive how fast this space is moving.
Watch video- 0Replies
- 0Reposts
- 2Likes
- 1KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post. - 50
@JulianGoldieSEO ·
DEEPSEEK V4 FLASH JUST BEAT ITS OWN PRO MODEL. I ran 50+ builds side by side—and the difference was impossible to ignore. What Flash 0731 did better: → Built a working 3D flight simulator while V4 Pro completely failed → Produced smoother controls, stronger graphics and cleaner game UI → Created usable coding projects despite being designed as a fast agent model Where it still failed: ✓ One build let you see through the floor ✓ The pool game was genuinely terrible ✓ Complex physics and controls were still inconsistent The bigger opportunity: ✔ Run the latest version free through OpenCode ✔ Connect it to Hermes using the OpenCode skill ✔ Delegate coding tasks automatically inside an agent workflow DeepSeek V4 Flash is not better than Fable 5 or GPT-5.6 Sol. But for a free, fast and lightweight agent coder, it can already outperform DeepSeek V4 Pro on real builds.
Watch video- 0Replies
- 1Reposts
- 1Likes
- 1.4KViews
Rewrite this post in your own voice and angle.See the hook, structure, and reusable template behind this post.
Best DeepSeek tweets
Write posts like these, in your own voice
Xholic studies what works in your niche, drafts posts in your voice and schedules them for the hours your audience is online.
Start 7-day free trialTry the live demoTake a 2-minute tour of a sample account in the real app. No sign-up or login needed.
No card needed · Nothing to cancel
Browse all tweet collectionsLive demo: no sign-up or login needed · takes about 2 minutes
Keep exploring
From the blog
- Best AI Tweet Generators for X in 2026The 10 best AI tweet generators for X compared by use case, how they work and the prompts that get results, plus a free tool to try now.
- 10 Top AI Tools for Creators in 2026Compare the top AI tools for creators in 2026 and learn how to build a focused creator stack around ideation, visuals, video, audio, repurposing, and X growth.
- Xholic AI Chrome Extension: Grow on X Without Leaving Your Feed (2026)The Xholic AI Chrome extension writes replies in your voice, remixes posts, shows analytics for any profile and keeps your mentions in one place, all inside X. Here's what it does.
Guides
- How to Write a Tweet With an AI Tweet GeneratorTurn a rough idea into a tweet worth posting: write a specific prompt, pick a tone and length, compare three drafts and make the best one sound like you.
- How to Write a Twitter Thread With an AI Thread GeneratorTurn a topic into a hook-first X thread: choose 4 to 10 posts and a tone, check the hook, edit each post and publish the thread in X's composer.
More topics like this
- Best tweets about GrokRead the best tweets about Grok and xAI, including model updates, features, benchmarks, developer access, and user experiments. Updated every two weeks.
- Best tweets about OpenAIExplore the best tweets about OpenAI, from model and API releases to research, developer tools, and real-world applications. Updated every two weeks.
- Best tweets about Vector DatabasesBrowse the best tweets about vector databases, covering embeddings, indexing, retrieval, Pinecone, Weaviate, Qdrant, Milvus, benchmarks, and architecture.
- Best tweets about Multimodal AIBrowse the best tweets about multimodal AI, including vision, audio, video, model architectures, benchmarks, applications, and developer experiments.