Kimi K3: first open 3T model 🔓, Torvalds tells AI critics to fork it 🍴, Claude logs in blind 🔑

Kimi released K3, a 2.8 trillion parameter open model with native vision and a 1 million token context window. Built on Kimi Delta Attention and Attention Residuals, it activates 16 of 896 experts and claims 2.5 times better scaling efficiency than K2. It is live on Kimi.com and the API today, with full weights by July 27, though Kimi concedes it trails Claude Fable 5 and GPT 5.6 Sol.

Xi Jinping endorsed open-source AI at a Shanghai conference, casting China as a champion of openness and implicitly criticising US moves to protect its AI lead. Dignitaries from 29 countries signed up a day earlier to create a China-led body promoting global AI cooperation, and Moonshot AI released open-weight Kimi K3 hours before the speech. Stanford's Graham Webster called the new governance body more rhetorical than substantive so far.

Google is months behind on Gemini 3.5 Pro, its most powerful flagship model, because it has taken time to improve capabilities, particularly coding. Ten current and former employees told Bloomberg the delay has frustrated engineers, researchers and managers who fear Google is losing its edge as Anthropic and OpenAI ship models exceeding Gemini's. Sources blame multiple layers of stakeholders preparing releases across a portfolio spanning search, maps and YouTube.

This spring Musk put Michael Nicolls in charge of xAI with one agenda, catching up with Anthropic's Claude chatbot. Nicolls told staff the near-term goals were to match Claude's performance, and Bloomberg saw internal projects naming Claude directly plus multiple xAI Slack channels named after it. He inherited a mess: Musk merged xAI into SpaceX this year, dozens of employees left including many co-founders, and daily operations were chaotic.

Fireworks, whose cloud runs open-source models for developers, raised $1.5 billion at a $17.5 billion valuation. It passed $1 billion in annualised revenue, five times last year's, handles 40 trillion tokens per day, and CEO Lin Qiao puts its cost five to 10 times cheaper than equivalent-quality closed models. Over half its revenue once came from Cursor, though Qiao says it has diversified as finance executives push open alternatives.

The European Commission issued two specification decisions on Thursday requiring Google to give third-party AI providers extensive access to Android, including voice activation in place of Gemini and taking actions in apps. A parallel decision makes Google share anonymised search data with rival engines and AI chatbots, bundling users into groups of at least 1,000. Search sharing starts January 2027, Android AI access July 2027, and litigation looks likely.

Microsoft CEO Satya Nadella told Copilot engineers that Anthropic's limits on Fable requests do not make sense, asking when they last used a creation tool so editorially controlled. Anthropic had said Fable 5's safeguards would flag a slightly higher fraction of harmless requests, after an export-control directive briefly cut off access. Nadella's wider line is the one to watch: renting token capital from two labs makes no economic sense.

Linus Torvalds told the kernel mailing list that Linux is not an anti-AI project, and anyone who disagrees can fork it or walk away. The thread concerned Sashiko, an agentic review system its creators claim independently finds 53.6 percent of bugs humans later fix, with false positives they put well within 20 percent. Maintainers eat the false positives, and Torvalds will very loudly ignore anyone arguing otherwise.

Hugging Face disclosed an intrusion driven by an autonomous agent framework that entered via a malicious dataset abusing two code-execution paths in its pipeline. Reconstructing over 17,000 recorded attacker events on frontier APIs failed, because provider guardrails cannot distinguish an incident responder from an attacker, so the team ran the forensics on open-weight GLM 5.2 in-house. Hugging Face's advice is to vet a capable self-hosted model before you need it.

Two engineers spent $100k over four attempts to get Opus to rewrite Postgres into 1.8M lines of idiomatic Rust. Attempt one stalled at 96 percent of the Postgres tests when its planner internals diverged, and attempt three left around ten thousand seams whose type signatures did not match their implementations. The public version passes the full regression suite and runs about 8x slower than Postgres.

Modal rebuilt its sandbox platform so no datastore sits in the creation path, making every worker its own source of truth publishing state into a Redis stream. Creation costs two network hops and one cheap CPU operation, and Modal started 1 million concurrent sandboxes in under a minute, with median start under half a second. It is in Beta, and the long latency tail is still worse than they want.

An M4 Max serving a 35B mixture-of-experts with 3B active saw a chat stream fall to 0.67 tok/s during a 60k-token prefill, against 65 idle. Prefill is compute-bound and decode is bandwidth-bound, and the engine grants roughly one decode step between 2,048-token prefill chunks, so a continuous benchmark leaves every decoder on scraps. A bigger prefix cache and an admission gate capping concurrent large prefills restored 60 to 70 tok/s.

Cassis ran 2,484 text-to-SQL trials on derivation distance, the work left between a model's context and the answer. On hard questions a two-dozen example library took Haiku from 70 to 82 percent and Sonnet from 81 to 93 percent, while three examples, relevant or not, matched the baseline. Coverage is what pays: Haiku plus examples matched Sonnet on metrics alone at roughly a third of the cost.

A FLOPs analysis asks whether China can train Kimi K3 on export-legal H800s and answers yes, comfortably. The model is 2.8T total parameters with 896 experts and 16 active, so only about 70B activate per token, and reading 2.8T as dense compute overshoots the real bill roughly 40 times. The H800's FLOPs were never cut, only its interconnect, so the real constraint is aggregate compute and Huawei chip supply.

Semgrep argues that deploying any model, open weight included, means accepting black box risk, because weights do not predict behaviour. It cites Anthropic's Small Samples research, where the samples needed to plant a persistent backdoor are few and do not grow with model size. No public evidence of a poisoned open-source model exists yet, so the authors want provenance standards and independent auditors, since benchmarks are easy to fine-tune against.

A VP of about 100 engineers spends five figures a month on coding agents and is not sure it is working. Token use does not correlate positively or negatively with productivity, so all four quadrants of the productivity versus token use 2x2 are valid, and measuring engineers on token usage is unequivocally a bad idea. Trying is the only way to know, and trying is expensive.

The author grants nearly every LLM critique, the slop, the copyright, the environmental cost, the collapse of trust in open source contributions, then reports spending almost $10,000 on tokens last month. The argument is that LLMs amplify what you already have, so if you have thoughts they come out sharper, and if you have nothing, nothing comes out very fluently. That difference is invisible from outside, which leaves only trust.

1Password for Claude fills a login or one-time code straight into the page after a biometric approval, scoped to the current task. The company says the credential never enters the model or its memory, and that its new Agentic Mode locks the extension down while an agent drives. It is 1Password's own post, and the integration is Mac only, needing both 1Password apps plus Claude desktop and Claude in Chrome.

ttsc is a drop-in tsc replacement on typescript-go, reading the same tsconfig.json and emitting the same JavaScript, with plugins running in the pass that type-checks the project. Sharing that AST pass is what lets @ttsc/lint replace ESLint and Prettier with rules that surface as compiler errors, which the README puts at up to 800x faster. It is MIT, and ttsx will not run a file until the whole project type-checks.

Ratel keeps an agent's tool and skill catalog out of the system prompt, so each turn searches for what it needs and gets only the matches. Indexes use BM25 over tool schemas and skill metadata, so retrieval is deterministic and needs no vector DB, with semantic ranking opt-in. The core is Apache-2.0 and the SDKs MIT, and the README carries no numbers, pointing at a separate benchmark site for results.