Open Source Did the Shipping This Week
Flagships were quiet Aug 4-10. Open source wasn't — a 30B local agentic model under Apache 2.0, a 2.6B on-device runner, and a wave of agent memory tooling.
The flagship labs took the week off. Open source didn't.
For Aug 4-10, if you were watching OpenAI/Anthropic/Google for a new base model, there was nothing verifiable. If you were watching Hugging Face and arXiv, it was one of the busiest weeks in months. Labs harden and tier, open source ships the plumbing.
###Muse Glimmer: 30B that fits on one GPU
Meta dropped Muse Glimmer on Aug 10 — 30B, multimodal, Apache 2.0. That license is the story. Not a Llama-style commercial carve-out, just Apache 2.0. BF16, GGUF quants, ExecuTorch, plus a DFlash speculative drafter.
A 30B model normally wants 55GB at full precision. Meta compresses to ~4-bit (~20GB) and claims 3.1x decode speedup via block-level speculation so it can live inside an agent loop on a single 24GB card or a Mac. As-claimed, but Simon Willison already had it running via LM Studio:
They're targeting DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench — full agentic task completion, not chat vibes. If you run agents with MCP servers, this is the one to benchmark. Apache 2.0 means you can actually ship it.
###LFM2.5-2.6B: forget it's even running
Liquid AI's LFM2.5-2.6B landed Aug 4 — 2.69B params, 131K context, text-only. 22 double-gated convolutions + 8 GQA blocks, 30 layers, ~34T tokens, GGUF/MLX/ONNX day one.
It's not trying to be Glimmer. Liquid says don't use it for coding or knowledge-heavy work. It's for cheap background agents — invoice parsing, form extraction, robotics — where you want 128K on-device without paying for 30B. Glimmer shrinks the frontier to fit; LFM makes it so small you forget it's running.
###Shieldstral: safety as a prompt
Mistral's Shieldstral 1.0 3B (Aug 7, Apache 2.0) flips moderation on its head. No fixed harm taxonomy — you supply policy as a plain question at inference and get a calibrated yes/no score:
score = model.classify(text, policy="Does this contain self-harm instructions?")
Mistral claims 84.9% avg F1, matching a 20B safeguard model 7x its size. If it holds, it's the efficiency story of the week. The catch is auditability — a score without a taxonomy is harder to explain to a regulator. But per-customer policies without retraining is exactly the right shape.
###The substrate wave
Four infra pieces in three days, all about time and memory:
- Shepherd (Aug 8, MIT,
pip install shepherd-ai) — Git-like typed traces for agent runs, fork process+filesystem 5x faster than Docker, >95% cache reuse on replay. - TencentDB Agent Memory v2.0 (Aug 7, MIT) — team memory hub turning chats/docs/code into Chat Memory, Skill, LLM-Wiki, and Code-Graph. Self-hosted via Docker.
MCP won the tool boundary. This week was the next one: what happened (Shepherd) and what we remember (Agent Memory). Neither solves auth for you — budget for that.
###Papers worth reading
- CreativeInstruct (arXiv:2608.07460, Aug 7) — teaching LLMs to balance quality, creativity, and diversity without RLHF flattening everything. Code on GitHub.
- CoinRAG (arXiv:2608.07458) — KV-cache reuse for long-context RAG via contextualized nuggets. Practical if you're actually running long contexts.
- Evidence-RL (arXiv:2608.08021, Aug 8) — counterfactual evidence disentanglement inside GRPO for visual reasoning. No inference overhead.
###What I'd do Monday
Run Glimmer on your agentic loop with your own MCP servers and measure task completion, not token speed. Put LFM2.5 on a spare device for the boring extraction you keep paying Sol to do. And try Shieldstral with your actual policy wording — it's the first classifier that treats policy as context.
Open source had the better week. It usually does when the labs are in harden mode.