Kimi-K3 weights land on Hugging Face — a 2.8T-parameter flagship anyone can download, under a license that isn't open source
Moonshot AI published the weights for Kimi-K3, its current flagship: 2.8 trillion total parameters with 104B active per token (16 of 896 experts), a 1M-token context window, and native vision/video in the same model. The weights ship as MXFP4 (with MXFP8 activations) straight from quantization-aware training — a 1.56TB download — and Moonshot credits its new Kimi Delta Attention and Attention Residuals designs with a claimed ~2.5× scaling-efficiency gain over K2. Self-reported benchmarks are aggressive (93.5 GPQA Diamond, 67.5 DeepSWE, 91.2 BrowseComp); no independent evaluation exists yet. The license is the other story: Moonshot calls this open-weight, not open source — products over 100M monthly users or $20M monthly revenue must display "Kimi K3", and Model-as-a-Service businesses past $20M in any 12 months need a separate Moonshot agreement, a notch more restrictive than K2's modified-MIT approach. Hosted access is already live, with OpenRouter listing K3 at $3/$15 per million tokens at the full 1M context.
Sources & depth
Anthropic says it has never wanted open weights banned — and lays out what it wants instead
With Washington debating restrictions on open-weight models, Anthropic published its position: it denies advocating a ban ("Anthropic has never advocated for a ban on open-weights models" — Dario Amodei) and calls benign open-weight models public goods. What it proposes instead: tighter chip export controls on China as the primary lever, a crackdown on industrial-scale distillation operations, and mandatory pre-release safety testing for all sufficiently capable models, open or closed alike. The company also commits to banning accounts that use its own models for distillation-based misuse. The statement lands days after Nvidia, Microsoft and Meta warned against overregulating open weights and startup founders urged Washington not to shut off Chinese open models — putting the major US lab most identified with safety arguments on the anti-ban side of the ledger.
Sources & depth
NVIDIA invests in Ilya Sutskever's SSI, which gets an "order of magnitude" more compute
NVIDIA and Safe Superintelligence Inc. announced a long-term strategic partnership: NVIDIA takes an investment stake in SSI, and SSI gets access to NVIDIA's next-generation Vera Rubin platform — enough, the companies say, to increase SSI's compute by an order of magnitude. Dollar amounts and chip counts were not disclosed. Notable on both sides: Sutskever says SSI has "research that is worthy of scaling up," the first concrete signal about the famously quiet lab's trajectory, while NVIDIA gains what the release calls rare access to SSI's closely guarded research and its input on future compute platforms. It extends NVIDIA's pattern of anchoring frontier labs to its roadmap through equity plus priority hardware access.
Sources & depth
40+ companies form the Open Secure AI Alliance to cast open AI as a cybersecurity asset
NVIDIA, Microsoft, IBM, Cisco, Cloudflare, Hugging Face, Palo Alto Networks and 30-plus other companies launched the Open Secure AI Alliance, arguing that defenders need open frontier AI systems and that policymakers should treat them as "defensive assets, not liabilities." The launch comes with concrete contributions: NVIDIA's open-source NOOA agent-governance framework, Microsoft's MDASH multi-model bug-discovery scanner, IBM/Red Hat's Lightwell signed-patch tooling for open-source supply chains, HPE's SPIFFE/SPIRE zero-trust identity for AI agents, and Hugging Face's Safetensors. SpaceX AI is open-sourcing its Grok Build coding agent and says model weights will follow. Coming the same week as Anthropic's open-weights statement, it's a coordinated industry push to move the open-model policy debate onto security grounds.
Sources & depth
Opus 5 quadruples the previous best on SlopCodeBench's long-horizon coding test — and still can't run unattended
A writeup benchmarks Claude Opus 5 on SlopCodeBench, UW Madison's long-horizon coding benchmark that reveals requirements incrementally across checkpoints instead of all upfront — closer to real software maintenance than one-shot tasks. Opus 5 hit a 24% strict pass rate on a 17-checkpoint subset (4 of 17 checkpoints defect-free), against 6% for both Opus 4.8 and Sonnet 5, with the original paper reporting Opus 4.6 at 17% and GPT-5.4 at 11%. The author's conclusion cuts both ways: a real generational jump, but "for real-shaped software engineering work, building one issue at a time, today's models can't be relied on to run lights-off without steering." Caveats worth carrying: only 3 problems tested, Opus 5 wrote ~3× more code than its predecessors (over half of it tests), and all models showed code-quality degradation as checkpoints accumulated.
Sources & depth
Six weeks after Bun's "11-day, $165K" AI rewrite in Rust — still no release
A skeptical status check on the most-cited AI-productivity story of the summer: Jarred Sumner's claim that the Bun JavaScript runtime was rewritten from Zig to Rust in 11 days for $165,000 in Claude API costs. Six weeks after the rewrite reportedly merged, no release has shipped — no release tag since mid-May — and the automated Claude bot has 2,475 open pull requests. The author argues the $165K figure excludes CI/CD infrastructure and direct Anthropic engineering involvement, and estimates the true ongoing cost may approach $800K, noting similar flagship AI projects (Anthropic's C compiler, Cursor's FastRender) have also gone quiet. A useful counterweight to headline AI-coding numbers: the expensive part isn't generating the code, it's landing it.
Sources & depth