Meta returns to open weights with a 30B Apache-2.0 agent model

Thirty billion dense parameters, Apache 2.0, and roughly 18GB at 4-bit quantization: Muse Glimmer is Meta’s return to open weights, and the field reorganized around it inside a few hours. The model is multimodal with a dedicated perception encoder, carries 128K context, speaks 100+ languages, and exposes a controllable reasoning-effort dial. What separates it from a normal weights drop is how it was made: Meta logit-distilled Glimmer from Muse Spark and trained it on agentic traces from the beginning, rather than shipping a base model and bolting tool use on afterward. A DFlash speculative-decoding drafter ships alongside the weights, which is the detail that matters if you intend to keep an agent loop running locally all day, since the drafter exists to hold interactive latency down on a single consumer card.

The license is the part practitioners should register. Simon Willison called it out immediately: every Llama release before this carried a non-OSI license with usage carve-outs, and Apache 2.0 removes the lawyer step from local deployment. On capability, Artificial Analysis puts Glimmer at 35 on its Intelligence Index, behind Qwen3.6-27B at 38 and roughly level with Kimi K2.5 at 36, while scoring 44 on openness. It runs about 60GB in BF16 and drops under 20GB for the language model at 4-bit, leaving headroom on a 24GB card for KV cache and the perception encoder. The reported weak spot is hallucination and knowledge calibration, so treat it as a tool-using local agent rather than a knowledge oracle. Alexandr Wang says open weights for Muse Spark 1.2 arrive soon, which would be the more interesting release of the two.

The second story of the day came from a model nobody can use. Anthropic ran an unreleased research version of Claude at the Riemann Hypothesis and reported that it raised the lower bound on the fraction of zeta zeros satisfying the hypothesis from 41.6% to 67.2%. The conjecture stands. What moved is a real, longstanding bound in analytic number theory, and the method is worth more attention than the headline: Jarred Sumner noted the run consumed 31M output tokens across repeated retries and wide exploration. That is search with a verifier, not a single brilliant completion, and it is the same shape as the agent loops most of us run against test suites all day.

OpenAI spent its day on the other end of the capability question. GPT-5.6-Cyber is a purpose-trained cybersecurity model gated behind Daybreak Red, restricted to approved defenders with extra monitoring for higher-risk work, and OpenAI says it has already been used to find previously unknown vulnerabilities in open-source software including Chrome’s V8 engine. Daybreak Blue, the broader tier, stays on GPT-5.6 Sol. The gating tells you how OpenAI reads the current threat picture, and it lands two weeks into a run of agent-driven supply-chain incidents that made the argument for them.

Pricing moved underneath all of that. Anthropic made Claude Sonnet 5’s introductory rate permanent at $2 per million input tokens and $10 per million output, a decision that reads as a response to an open field getting good enough to matter, and one of several threads in yesterday’s AINews issue worth reading in full. The harness benchmarks in that issue are the sharper signal: Composio ran DeepSeek V4 Flash through four different harnesses over 30 agentic tasks and found the cheapest harness also produced the best results. Same weights, same tasks, different scaffolding, materially different outcomes. Alongside that, a paper summary making the rounds claims typed Python stubs executed in-code beat native JSON tool calling in 11 of 14 models tested, with the GPT-5.6 family up 10.6% over the JSON baseline on BFCL v4.

Tool-interface design got a harder empirical result too. A 54,000-trial study across 24 models (9 Claude, 6 Gemini, 9 GPT) against a production legal-information MCP server asked whether clients actually read reference data embedded in server instructions. With the competing search tool removed, 23 of 24 models read the embedded table reliably at 98%+ hit ratio. With a search tool merely present in the toolset, 9 models dropped below 15%. The capability was never missing; the models simply preferred to go searching. Combining three instruction-level interventions restored 86%+ for 20 of 24 models, but individual interventions backfired for specific model families, which is why the authors argue for a host-level mechanism that puts server instructions ahead of tool selection instead of leaving every server author to prompt-engineer around it.

Cost discipline is becoming its own engineering discipline. JetBrains described how development-related AI spending grew roughly tenfold in six months and what it built in response: not an approved-tools whitelist, but a shared access and accounting layer that preserves tool choice while making consumption visible. The Stack Overflow podcast landed on the same nerve from the other side, with Coder’s Rob Whiteley arguing that token consumption and PR merge counts have become Goodhart-style proxies that stop measuring anything once teams optimize for them.

Two shipping notes for people building on agent platforms. Amp added global plugins and skills, Amp-hosted and available across every surface Amp runs on, so a personal formatter hook follows you instead of living in one repo’s config. And GitHub published a walkthrough of the Copilot SDK for Java, which lets enterprise Java apps drive agents without a Langchain4j or Spring AI dependency in between.

Watch whether Muse Spark 1.2’s weights actually land, and on what license. Glimmer at 30B is a credible local agent; Spark is the model that would put Meta back in the frontier open-weights conversation rather than the consumer-hardware one.