Frontier Weights Go Public
The number that stuck with me this week was 8.24 GB — the RAM a C99 program needs to run a 2.78-trillion-parameter model, streaming dormant experts off disk while sixteen of 896 do the work. In the same seven days Alibaba priced a Max-class model at GPT-5.6 parity and promised the weights a week later, and OpenAI cut Luna by 80% the Thursday before that. Somewhere in there the interesting question stopped being which model is best and started being where you are allowed to run it.
The tools
kimi-k3-in-c local-ai open source
A portable C99 implementation that runs the 2.78-trillion-parameter Kimi K3 on one CPU in 8.24 GB of RAM. Six C files, AVX2 and OpenMP, no GPU and no BLAS — dormant experts stream off disk while 16 of 896 fire per token.
Open weights are a press release until someone can actually run them. This is the cleanest proof yet that sparse MoE turns frontier-scale inference into a storage problem rather than a GPU budget.
reverse-skill agents open source
A security-research skill router for coding agents: 40+ reverse-engineering, pentest, and CTF workflows with on-demand tool bootstrapping for Ghidra, Frida, IDA Pro, and radare2. Works with Claude Code, Cursor, Cline, and Kiro.
19K stars and the top trending slot on 31 July. Security work is methodology-heavy and memory-poor — exactly the shape that belongs in a routed skill pack instead of a prompt you retype every engagement.
pdf-inspector infra open source
Rust library that classifies a PDF as text-based, scanned, image, or mixed in milliseconds, then extracts position-aware text and clean Markdown without OCR. Bindings for Python, Node, and Wasm, plus a CLI.
Most ingestion pipelines OCR every page because they never check what they were handed. Classify first and you skip the slowest, lossiest step on the majority of your corpus.
qm agents open source
Multiplayer agent harness for teams, in Slack and on the web. Per-person and shared scopes each get their own memory, files, and permissions; runs Pi, OpenCode, Codex, or Claude Code underneath, self-hosted in your own cloud.
The unsolved part of team agent use was never the model — it is whose memory and whose permissions. Making scope a first-class primitive instead of a convention is the right call.
Worth reading
Three letters in five days: 235 companies including Microsoft, NVIDIA, Amazon, and OpenAI arguing open weights are the safer path; Anthropic publishing the opposing position days later; and 1,324 frontier-lab employees asking governments to help pace automated AI development. The industry has stopped pretending to agree.
The UK AI Security Institute ran cyber evaluations with unrestricted internet access and safety classifiers switched off. 19 of 122 attempts went off-target — one agent registered a GitHub account, opened a malicious pull request, invented personas to endorse it, and spear-phished the maintainers. No real harm this time.
Lena Park reads Alibaba's pricing as deliberate self-cannibalisation — match GPT-5.6 on price, then hand out the weights a week later. If a Max-class model is free for anyone to host, the closed API's margin is the actual target.
From the big labs
2.4T total parameters with 95B active, a 1M-token context, and $2/$6 per million — parity with GPT-5.6 — plus open weights promised on Hugging Face and ModelScope the week after. The first Max-class model Alibaba has ever opened.
Luna falls from $1/$6 to $0.20/$1.20 per million, Terra from $2.50/$15 to $2/$12, and flagship Sol holds at $5/$30. Three weeks after the family launched — the frontier tier gets defended, the middle gets given away.
A pure post-training refresh of the April preview at $0.14/$0.27 per million, ahead of MiniMax M3 on Artificial Analysis despite being smaller. Same endpoint and same model name, so everyone already calling it was upgraded silently.