Kimi K3 Runs on One CPU With 8 GB of RAM

Start here ↓

A 2.78T-parameter model ran on one CPU with 8.24 GB of RAM.

kimi-k3-in-c runs Kimi K3, all 2.78 trillion parameters, on one CPU, and the engine that does it is 176 KB of portable C99, with no BLAS, framework, or GPU.

That sounds fake until you see the tradeoff. The model still needs a 1.56 TB checkpoint on disk. Also, it is painfully slow: 26.5 seconds per token at 8 GB, 19.8 at 64 GB, 5.6 once 128 GB holds the whole thing.

Nobody is serving traffic from this. That is not the point. It turns the “you need a cluster for this” assumption into a concrete systems question: which bytes must be in memory, which bytes can sit on disk, and how much speed are you willing to trade for access?

If you like tiny, readable systems code, this is the rabbit hole to dig.

FareedKhan-dev/kimi-k3-in-c

🚀 Shipped

Warp launched Skill Doctor, a way to improve agent skills from past sessions. Skill Doctor reads old Claude Code, Codex, and Warp transcripts, scores where the agent did well or wasted time, then proposes diffs to your skill files. It’s going from “write a better prompt” to “learn from the mess your agent already made.”
Skill Doctor: score and improve your agent skills | Warp

Tencent dropped a 770B open model that might actually be servable. Hy4-preview is a 770B-parameter MoE with 49B active parameters per token and a 1M-token context window. That active-parameter count is the interesting part: it puts the model in the “huge on paper, less terrifying to run” category. Tencent released it under Apache 2.0, so teams can test it without immediately hitting a commercial-use wall.
tencent/Hy4-preview · Hugging Face

Firecrawl made web search and scraping usable before signup, no API key required. You get 1,000 free credits per month, across MCP, CLI, and REST API. Specially great for prototypes, workshops, and agent demos.
Introducing Firecrawl Keyless

GLM-5.3 is now fine-tunable through Thinking Machine’s Tinker API. Tinker is Thinking Machines’ training API for LoRA fine-tuning, SFT, RL, DPO, and distillation. The model is currently the strongest open-weights model on coding evals like Terminal-Bench 3.0 and DeepSWE 1.1, and great for teams with their own datasets.
tinker-docs.thinkingmachines.ai

fal is giving free MiniMax H3 Max video generations (five videos a day). The model can generate a 5-second 768p clip in under 3 seconds, with free daily generations and no signup. If you work with video models, this is worth testing with your own prompts instead of trusting the cherry-picked examples.
MiniMax H3 Max

🧠 Worth Knowing

OpenAI is winding down model access for Cursor after SpaceX’s acquisition starting November 12, 2026. OpenAI says the cancellation window comes from a change-of-control clause, and that future models including Astra will not be provided. Cursor CEO Michael Truell responded that OpenAI models serve about 5% of Cursor user traffic and that talks are still ongoing.
openai.com

Sometimes quiet failures happen when relevant context gets treated like authority, not prompt injection. A file, snippet, or prior chat can help an agent understand a task, but it should not automatically get to decide what is true or allowed. The post shows three failures: sibling files steering code repairs, copied evidence being counted as independent sources, and conversational trust being mistaken for permission. The fix is provenance: track source, lineage, trust, and authorization before the agent acts.
AI Agent Context Security Needs Provenance

Warp has run 10M Claude Code sessions inside Warp, with 400K+ sessions per week and 800K monthly developers. The interesting part is Warp’s loop: users give feedback where work happens, then a separate agent turns that feedback into small skill edits. That is much better than letting every correction vanish when the session ends.
How Warp builds self-improving agents on Claude

TurnBench measures the part of voice agents that gets edited out of demos. Sesame’s TurnBench evaluates when a voice system should speak, wait, or treat speech as an interruption. The benchmark uses 30 hours of dual-channel human conversations across 154 dialogues and 106 actors. The leaderboard shows some systems detect more turn endings but interrupt too often. While others stay careful but respond slowly.
TurnBench

🔧 Clone and Run

Clone & Run of the Day
OpenKB is a vectorless RAG that turns raw documents into a wiki-style knowledge base and retrieves through PageIndex reasoning instead of embedding search. Not every knowledge system has to start with vector search. If your current RAG stack is brittle on source structure, this is worth trying on a small corpus.
OpenKB

only-cli turns websites into tiny CLIs for agents so an agent can use it in hundreds of tokens instead of reading full pages. It also gets past blocks that stop naive fetchers on some sites. If page views are eating your context window, turning messy web UI into a small command surface is one of the cleaner ideas to test.
only-cli 

Fuxi is a terminal coding agent that makes model routing part of the workflow instead of a hidden backend choice. It can edit files, run shell commands, use tools, and switch across LLM providers while tracking cost. The useful thing to inspect is its routing behavior: which steps go to cheaper models, which need stronger ones, and how much the run costs as it works.
Fuxi

Repomix is for the step right before you paste a whole repo into an agent. It packs a codebase into a single agent-readable bundle with file structure, selected source content, AST-aware compression, XML-style output, and Secretlint scanning. That gives the model enough project context without dumping every raw file into the prompt.
Repomix

Google’s Chrome DevTools MCP gives frontend agents the browser state they have to guess from screenshots or logs. It lets coding agents inspect Chrome DevTools directly: console errors, network requests, DOM state, screenshots, and performance traces. That means an agent can edit the code, reload the app, inspect the real browser, and keep debugging.
chrome-devtools-mcp

OpenBot is an open-source alternative to Grok Bot. These are AI coworkers you can hand real work to and trust with access. Each gets a computer of its own: a real browser with its own logins, its own files, and only the tools you grant. Every action is decided before it happens and recorded after.
OpenBot

GenOffice is a free, open-source office suite for macOS, Windows & Linux with AI agents built in. It can edit Word docs, spreadsheets, presentations, PDFs, and Markdown, which makes it more practical than tools that only handle clean text. Worth a look if your automation work involves messy office files.
genoffice

geo-seo-claude audits how a site appears across ChatGPT, Claude, Perplexity, and Google AI Overviews. It checks citation scores, AI crawler access, schema markup, brand authority, and platform-specific optimization, then produces reports. “Are AI answers citing us?” is still fuzzy for most teams, and this turns it into a concrete agent job.
geo-seo-claude

Awesome LLM Apps is a curated collection of 100+ AI Agents, Agent skills, and RAG apps. It covers models from OpenAI, Anthropic, Google, and open-source models like GLM, DeepSeek, and Qwen that you can run locally on your computer. (Now accepting GitHub sponsorships)

📊 By the Number

Number of the Day
Claude Code’s permanent limit bump is smaller than the temporary one users have right now. They are increasing their standard weekly Claude Code limits permanently by 25% for Pro, Max, Team, and seat-based Enterprise plans starting September 14. Until then, the current 50% temporary increase remains. Compared with today’s temporary level, the permanent level is about 17% lower. So if Claude Code feels unusually roomy this week, do not use that as your long-term baseline.
ClaudeDevs

OpenAI is resetting paid Codex and ChatGPT Work usage after fixing several token-burn bugs across image compaction, memory workers, goals, automations, subagents, computer history, rolling summaries, and MCP result handling. You may see usage go 10% to 50% further depending on workflow.
Tibo

Terminal-Bench 4.0 makes cost part of the benchmark, not a side conversation. After publishing a new version of the benchmark’s dataset and leaderboard, Terminal-Bench 4.0 now reports resolution rate, tokens, and cost together. Opus 5 leads at 51.8% resolution with 6.5B tokens and about $6.0K. GLM-5.3 reaches 41.8% with 8.7B tokens and about $2.7K. Fable 5 sits at 44.5%, with overlapping error bars versus GLM-5.3, but higher reported cost.
Terminal-Bench 4.0

Anthropic is giving scientists a cheaper Claude path, with eligibility limits. Anthropic opened 10,000 free standard Claude seats for scientists for one year. Premium seats with 5x usage limits are $15 per month. The program is gated to verified principal investigators or equivalents at academic or nonprofit research institutions.
Expanding our support for scientists

That's all for today. Come back tomorrow for the next batch of AI tools, model drops, agent repos, and weird benchmarks worth your time.

If you found one thing to try, share the issue with someone who ships.

Unwind AI - X | LinkedIn | Threads

Awesome LLM Apps | Sponsor Us

Stop Paying for 10 Tools. One AI Does It All.

Most e-commerce sellers are running their store across 6 to 10 separate tools — and spending more time managing software than growing their business. StoreClaw replaces your entire stack with one autonomous AI engine that monitors competitors, optimizes listings, automates marketing, and tracks real profit across Shopify, Amazon, and beyond.

It doesn't wait for you to ask. It runs 24/7 in the background, so you wake up to a full dashboard instead of a list of things you forgot to check.

Connect your store, and StoreClaw gets to work — no prompts, no complex setup, no six-app stack.

Free to start. No credit card required.

Try it for FREE today