The one thing
01Meta ships Muse Glimmer: 30B, Apache 2.0, built to run local agents on one GPU.
Muse Glimmer is a dense ~29.6B multimodal model from Meta Superintelligence Labs, open-weighted under Apache 2.0 and on Hugging Face today.1 It is trained around the agent loop — plan, call tools, check results, recover from failure — not chat. Quantized weights target under 20 GB so the model, KV cache, perception encoder and a DFlash speculative drafter fit a 24–32 GB envelope; Meta claims agentic reliability holds at 4-bit on that class of machine.12
Day-zero runtimes are already moving: transformers v5.15.0 adds the model; Ollama 0.32.7/0.32.8 adds MLX then NVIDIA/AMD paths; Hugging Face documents llama.cpp, vLLM and SGLang support landing with the weights.345 Integrations for Ollama, LM Studio and Unsloth are rolling this week rather than all live at publish time.1
Deeper — what is open, what is not
- Apache 2.0 covers full-precision BF16, 4-bit quants, the DFlash drafter and the ~1.8B perception encoder. Training data and training code are not released.2
- Context is stated at 131k+ tokens; knowledge cutoff 4 Jan 2026; text+image in, text out; 100+ languages; controllable effort via system prompt.2
- This is Meta's first fully open weights release since Muse Spark went proprietary. The licence is the news as much as the size.6
Shipped
05Claude Sonnet 5 · $2/$10 made permanent
Anthropic's Sonnet 5 launch post was edited 10 Aug: introductory API pricing of $2 per million input tokens and $10 per million output stays. The $3/$15 step that was due 1 Sept is cancelled.7 Budget the locked rate. The platform pricing docs still show the old two-row table with an August 31 cutoff — treat the dated edit on the news page as the primary until docs catch up.8
OpenAI Agents Python 0.20.0
Implicit default model is now gpt-5.6-luna. Explicit model strings, run-level overrides and OPENAI_DEFAULT_MODEL still win. Local MCP supports Python SDK v1 and v2; apps with custom MCP HTTP auth or client factories must use the HTTP types of the installed MCP major, or pin mcp<2.9
vLLM 0.27.0
Kimi K3 full stack in one release, Qwen3.5 dense/MoE, PyTorch 2.13 (breaking env), deeper FlashAttention 4 on SM100, and a pile of DeepSeek-V4 serving wins. 561 commits, 242 contributors.10
Vercel Sandbox · Managed Images
Sandbox SDK v3 defaults to vercel/sandbox/universal:latest on Ubuntu 26.04 with Node 24, Python 3.14, and preinstalled opencode, claude-code, codex and pi. Old runtime stays as a deprecated alias; Amazon Linux stays only via runtime. Source images live in the public vercel/sandbox repo.11
Cline v4.1.8
Yolo Mode toggle removed — it was cosmetic; auto-approve is now the single control, and prior Yolo setups migrate to approve-all. Vertex accepts any model ID by hand, including Fable 5, and stops applying Anthropic list prices to Vertex Fable bills.12
Promised, not shipped
Muse Glimmer partner runtimes (Ollama beyond first paths, LM Studio, Unsloth) and optimized llama.cpp / MLX / ExecuTorch — rolling this week · Muse Spark 1.2 open weights — Zuckerberg said "soon", no date · Claude Code auto mode default — 14 Aug 2026 (issue 005)
The conversation
01A 30B Apache model aimed at always-on local agents reopens the question the cloud defaults closed: what workload still needs a metered frontier call once tool-use quality is good enough on one GPU.1
- Meta research blog · primaryVendor claim, dated
Positions Glimmer against Gemma4-31B and Qwen3.6-27B on agentic benches (DeepSearch QA, MCP-Atlas, τ-Bench, SWE-Bench), and ships DFlash block drafting so generation stays interactive on M4/M5 Max and RTX 5090 class boxes.1
- @aussieguy1234 · Hacker NewsReader, unverified
Notes SWE-bench-verified in the Opus-of-not-long-ago band and asks the useful question: cloud is faster and more reliable for most software — do you need that margin for this task?13
- @hougaard · Hacker NewsField report, unverified
Ran the full build (~120 GB RAM), fed defective code, watched it loop and dig in. One session is not a bench; it is a reminder that agentic claims die on recovery, not on the first tool call.13
Open weights under Apache matter more than the parameter count: you can fine-tune, ship, and keep the stack when a lab changes its mind. That is the difference from Muse Spark.
The decision is not "local vs cloud". It is which loops are private or cost-sensitive enough to pin a 24 GB box, and which still buy frontier recovery. Measure your own scaffold before you rewrite the default.
GitHub releases: 28 watchlist repos polled, sweep complete (public API). Vendor feeds: 12 sources, sweep complete, 16 in-window. HN above the noise floor with thread comments. Fresh Reddit pulse (6 subreddits, 8 posts above floor). X accounts digest last collected 05:30 UTC 10 Aug (~24h old) and X keyword digest same age — background only, stated. No 11 Aug X digest in-repo.
Skip this
06OpenAI Daybreak / GPT-5.6-Cyber partner programme. Authorized cyber research for approved partners. Not a tool you install this week.
OpenAI letter to Governor Abbott on Texas AI infrastructure. Policy. Wrong beat.
GitHub custom thread subscriptions deprecated. Notification UX. Converts custom → subscribed. Not a toolchain change.
AI SDK monorepo patch storm (ai@6.0.248 and siblings). aspectRatio:'adaptive' on generateVideo and dependency bumps. Collapsed.
Needle2 14MB agentic LLM for phones. Show HN, interesting edge bet, no independent measurement we opened.
OpenAI status: increased error rates, 12.7h minor. Resolved minor. Weather, not a story without customer impact numbers.
Everything we saw
7878 candidates scanned · 14 used in this issue — the rest, with the reason each one was left out
| Item | Source | Signal | Call |
|---|---|---|---|
| Muse Glimmer open weights | research.meta.ai | 1083p · 592c HN | led the issue |
| Muse Glimmer model card | huggingface.co | primary card | one thing |
| Sonnet 5 pricing made permanent | anthropic.com | edit 10 Aug | shipped |
| OpenAI Agents Python v0.20.0 | github releases | 2.9h | shipped |
| vLLM v0.27.0 | github releases | 8.8h | shipped |
| Vercel Sandbox Managed Images | vercel changelog | 12h | shipped |
| Cline v4.1.8 | github releases | 2.1h | shipped |
| transformers v5.15.0 | github releases | 19.6h | support path for Glimmer |
| Ollama v0.32.8 | github releases | 6.3h | Glimmer runtime |
| Vercel: network boundary essay | vercel.com | 6h | context for sandbox, not a separate ship |
| Claude Code v2.1.227 | github releases | 7.1h | flag/Bash fixes |
| LiteLLM v1.96.0 | github releases | 7.7h | proxy fixes + cosign note |
| Zuckerberg FT on open models | HN 453p | paywalled colour | background to Glimmer |
| Reddit pulse — Sonnet pricing, Cursor token waste, Muse Spark 1.2 rumour | reddit pulse | 8 posts, 6 subs | pricing confirmed via Anthropic primary |
| X digests 10 Aug 05:30 | digests | ~24h old | background only |
End of feed. That is everything from the window worth your time.
Next issue tomorrow, 06:00 UTC — and if nothing ships, it will say so in two hundred words.