# Muse Glimmer Is Meta's Apology for the Llama License > Source: > Published: 2026-08-10 13:08:14+00:00 [AI](https://sourcefeed.dev/c/ai)Article # Muse Glimmer Is Meta's Apology for the Llama License A dense 30B coding agent under Apache 2.0 puts Meta back in the local-model race it started and lost. [Mariana Souza](https://sourcefeed.dev/u/mariana_souza) Meta's Superintelligence Labs shipped [Muse Glimmer](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model) today: a roughly 30-billion-parameter dense model, multimodal, 131K context, distilled from its larger Muse Spark sibling, and aimed squarely at local coding and agent workloads. The weights are on [Hugging Face](https://huggingface.co/meta-models/Muse-Glimmer-30B) under Apache 2.0, with day-one support in llama.cpp, MLX, vLLM, [Ollama](https://ollama.com), and [LM Studio](https://lmstudio.ai). The headline benchmark is 76.0% on SWE-Bench Verified — territory that belonged exclusively to frontier API models a year ago. But the more consequential number is the license text: zero clauses. ## The license is the actual news Llama arguably started the open-weights era, and then Meta spent two years undermining it with lawyer-bait. The Llama Community License's 700-million-MAU clause, the acceptable-use policy, the "built with Llama" branding requirements — none of it stopped hobbyists, but all of it made corporate legal teams reach for Qwen or DeepSeek instead, both shipping under Apache 2.0 and MIT with no strings. By late 2025, "local model" had quietly become synonymous with Chinese open weights, and Meta's answer — ever-larger Llama flagships — wasn't landing. Muse Glimmer reads as the course correction. Apache 2.0, no MAU thresholds, no naming mandates. If you've been blocked from deploying Llama derivatives by a compliance review, that objection just evaporated. It's a tacit admission that the Llama license was costing Meta the one thing open weights are supposed to buy: default status in developers' stacks. The strategic shape changed too. Instead of chasing the frontier with open weights, Meta is keeping the frontier model (Muse Spark) as a teacher and shipping the distilled student. That's the same playbook DeepSeek ran with its R1 distillations, and it's the right one — the open-weights race at the 30B scale is winnable in a way the trillion-parameter race is not. ## What you actually get The model card is unusually specific. It's a dense causal transformer — 52 layers, grouped-query attention at a 16:1 ratio, a local-global attention pattern with a 2,048-token sliding window, plus a 1.8B-parameter vision encoder. Training ran distillation from Muse Spark during pre-training, agent-trajectory data mid-training, and a post-training mix of SFT, on-policy distillation, and RL across coding and tool-use domains. The tool-use training explicitly covers failure recovery: when a call errors, the model is trained to diagnose and retry rather than stall — which, if it holds up, addresses the single most common way local models fall over inside agent harnesses. Quantized to 4-bit, it fits in 17–20 GB, so a 24 GB GPU or a mid-tier Apple Silicon Mac runs it with room for KV cache. Meta's DFlash speculative decoder claims 3.1x faster generation on an RTX 5090 and 1.5x on an M4 Max; early Hacker News reports put raw throughput around 75 tokens/second on a 5090 and roughly half that on a 4090. The usual caveat applies with extra force here: every benchmark number is Meta-reported, and 76% SWE-Bench Verified from a 30B model would put it within striking distance of closed models a hundred times more expensive to run. Meta's own comparison tables pit it against Gemma 4 31B and Qwen 3.6 27B — a fair weight class — but nobody independent has reproduced the agentic numbers yet. Wait for third-party SWE-Bench runs before rearchitecting anything. ## The dense trade-off nobody's dwelling on Meta went dense where Qwen went sparse, and that choice defines who this model is for. Qwen's 30B-class MoE models activate only a few billion parameters per token, which is why they're tolerable on CPUs and bandwidth-starved unified memory. A dense 30B pushes all ~55 GB of BF16 weights (or ~18 GB quantized) through the memory bus on every token. HN commenters testing DDR5 systems report around 15 tokens/second — fine for chat, painful for an agent loop that burns thousands of tokens per step. Dense buys you things in exchange: more capability per gigabyte of VRAM, simpler quantization behavior, and dramatically easier fine-tuning — no router balancing, no expert-collapse debugging, a LoRA run works the way you expect. If you're customizing a model for your codebase, dense is the architecture you want. If you're squeezing tokens out of a machine without a real GPU, Qwen's MoE line is still the better host. ## Should you switch your local agent to it? If you're already running a local coding agent — Aider or a Continue/Cline-style setup pointed at Ollama — swapping in Muse Glimmer is a one-line model change and worth an afternoon of evaluation, particularly for the tool-calling reliability the training targeted. The 131K context matters more than it sounds: agent loops eat context with diffs and tool output, and most local models degrade well before their advertised window. If you're not running local yet, be honest about the economics. Cloud inference for open-weight models has gotten absurdly cheap, and a per-token API bill beats a 5090's price and power draw for most solo developers. Local wins on three grounds only: code that can't leave the building, offline or air-gapped work, and agent architectures that hammer the model constantly enough that flat-rate hardware beats metered tokens. Those are real niches — regulated industries, on-prem enterprise, CI-integrated agents — and they're exactly the niches where Llama's license was disqualifying and Apache 2.0 isn't. My read: this is a genuine shift, not a hype release, but the shift is strategic rather than technical. Individually, nothing here is new — distilled 30B models, Apache 2.0 licensing, and local agent tuning all existed. What's new is Meta doing all three at once, which turns the local-model race from a Qwen-versus-DeepSeek contest into a three-way fight with the company that has the most GPUs and the most to prove. Competition at exactly the model size developers can afford to run is the best possible outcome — whoever wins. ## Sources & further reading - [Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model)— research.meta.ai - [Muse-Glimmer-30B model card](https://huggingface.co/meta-models/Muse-Glimmer-30B)— huggingface.co - [Meta Muse Glimmer - open weights 30B local coding model](https://news.ycombinator.com/item?id=49241679)— news.ycombinator.com - [Meta Publishes Muse Glimmer As 30B Open Agentic Model](https://www.phoronix.com/news/Meta-Muse-Glimmer)— phoronix.com [Mariana Souza](https://sourcefeed.dev/u/mariana_souza)· Senior Editor Mariana covers the fast-moving world of machine learning and generative AI, with a particular focus on how these technologies are reshaping development workflows. When she isn't stress-testing the latest foundation models, she's usually at a local hackathon. ## Discussion 0 No comments yet Be the first to weigh in.