Meta Releases Muse Glimmer, a 30B Open-Weights Model Under Apache 2.0 License
Meta has released Muse Glimmer, a 30 billion parameter open-weights model licensed under Apache 2.0. The model targets agentic task completion, tool use, and multi-step reasoning, with benchmark results claimed on DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench.
Meta has released Muse Glimmer, a new 30 billion parameter open-weights language model, marking a return to open model releases after a period of relative quiet from the company on that front. The model is licensed under Apache 2.0, a notable departure from the more restrictive custom licenses Meta attached to previous Llama releases.
What's new
According to Meta, Muse Glimmer is built specifically for agentic workloads rather than general chat use. The company highlights three capability areas:
- End-to-end agentic task completion — Meta claims strong success rates on full-task benchmarks including DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench. These benchmarks evaluate an agent's ability to operate within scaffolds, write and debug code, and complete multi-turn tasks from start to finish.
- Reliable tool use — the model is designed to invoke function calls with precise schemas across extended workflows.
- Multi-step reasoning — Meta says the model sustains coherent plans over long-horizon, complex tasks.
Meta has not published specific numeric benchmark scores in the announcement; the claims above are Meta's own characterization of the model's performance on the named benchmarks. No independent verification of these results is currently available.
No context window size has been disclosed. As an open-weights release, there is no per-token API pricing — the model is intended to be downloaded and run locally or self-hosted.
Simon Willison, who covered the release on his weblog, tested a quantized version of the model via LM Studio — an 18.16 GB file — and noted it fits comfortably on machines with 32 GB of RAM or more while leaving headroom for other applications. Willison's own machine has 128 GB of RAM.
Training and release details
Meta has not disclosed a training data cutoff date, exact parameter architecture details, or a formal model card with quantitative benchmark scores as of this writing. The release appears to be part of a broader trend among labs releasing mid-sized open models (in the 20B–40B range) optimized for agentic and tool-use scenarios rather than raw chat quality.
What this means
The Apache 2.0 license is the most consequential detail here. Meta's prior Llama license terms included commercial-use restrictions and attribution requirements that discouraged some enterprise adoption. A clean Apache 2.0 license removes that friction and puts Muse Glimmer on more permissive footing than most of Meta's earlier open releases.
The 30B size class is a deliberate choice: it's large enough for meaningful reasoning and tool-use capability but small enough to run on consumer and prosumer hardware — a single machine with 32–64 GB of RAM can serve it locally. That positions Muse Glimmer as a practical alternative for developers building local agents who don't want to depend on hosted APIs for cost, privacy, or latency reasons.
The absence of published numeric benchmark scores and a context window figure makes it difficult to assess how Muse Glimmer stacks up against similarly sized open models like Qwen, Mistral, or DeepSeek's smaller releases. Until Meta or independent evaluators publish concrete numbers, the agentic-performance claims should be treated as unverified.
Related Articles
Alibaba Open-Sources Qwen3.8-2.4T-A95B, Its First Qwen-Max-Class Model With Public Weights
Alibaba's Qwen team released Qwen3.8-2.4T-A95B on August 12, 2026, the open-weight version of Qwen3.8-Max and the first Qwen-Max-class model made publicly available. The 2.4 trillion-parameter mixture-of-experts model activates only 95 billion parameters per token and supports context windows up to 1 million tokens.
DeepSeek Launches V4.1 Flash: Low-Cost MoE Model Claims to Beat V4 Pro
DeepSeek has released V4.1 Flash, a sparse mixture-of-experts model priced at $0.30 per 1M input tokens and $1.20 per 1M output tokens with a 1 million token context window. DeepSeek claims the model exceeds the larger V4 Pro on performance, speed, and task completion time.
DeepSeek Releases V4.1-Flash: 552B MoE Model Cuts KV Cache to 890 Bytes Per Token
DeepSeek has released V4.1-Flash, a 552B-parameter multimodal Mixture-of-Experts model supporting 1M-token context and activating only 8B parameters during prefill. The model uses a new Causal Encoder-Decoder architecture and Compressed Sparse Attention 2 to cut global KV cache to 890 bytes per token, roughly a quarter of its predecessor.
IBM's Granite PatchTST-FM-r2 Tops Permissively Licensed Time Series Forecasting Models on GIFT-Eval
IBM has released Granite Time Series PatchTST-FM-r2, a 385M-parameter forecasting model that ranks #2 overall among replicable zero-shot models on the GIFT-Eval benchmark and first among those with permissive, commercial-friendly licensing. The model uses conformer-based architecture and supports probabilistic forecasts across up to 8,192-step contexts.
Comments
Loading...