Muse Glimmer
- Provider
- Meta
- Status
- available
- Context
- 131,072 tok
- SWE-bench
- 76%
- Knowledge
- 2026-01
Muse Glimmer is Meta’s 30-billion-parameter open-weights model, released on 10 August 2026 under an Apache 2.0 licence — and it is the most consequential thing Meta has shipped since it walked away from open weights in April. It is distilled from the proprietary Muse Spark 1.2, built specifically for local agentic work, and it fits on a single consumer GPU: the 4-bit build is about 17GB, down from roughly 60GB at full precision.
The headline is the licence, not the benchmark. Apache 2.0 is an OSI-approved licence with no revenue cap and no monthly-active-user ceiling — which makes Glimmer more permissively licensed than any Llama model ever was, and a direct answer to years of criticism that Meta’s “open source” branding didn’t survive contact with the Llama Community Licence’s 700-million-user clause. Mark Zuckerberg framed the release in an accompanying essay: “Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it” (Engadget).
Verdict — 12 August 2026. Muse Glimmer is the best local agentic model in its size class, and the most permissively licensed model Meta has ever published. On Meta’s own numbers it beats Qwen 3.6 27B and Gemma 4 31B decisively on tool use and agent orchestration (MCP Atlas 75.5 vs 62.5 and 54.2), while trailing Qwen on computer use and terminal work. Two caveats keep this from being a bigger story than it is. First, it is 30B — deliberately mid-size, and not a frontier model or a rival to Kimi K3 or DeepSeek V4 at the top of our open-source ranking. Second, every published figure is Meta-reported; nothing independent has landed yet. If you want one sentence: this is the model to reach for when you want an agent that runs on your own hardware, and the licence is the reason to pick it over the alternatives.
Quick specs
| Provider | Meta (Meta Superintelligence Labs) |
| Released | 10 August 2026 |
| Status | Available — open weights on Hugging Face |
| Licence | Apache 2.0 (no revenue or MAU cap) |
| Parameters | ~29.6B total — dense, ≈27.8B language model + ≈1.8B vision encoder |
| Architecture | Dense causal transformer (not MoE) + ViT-style perception encoder |
| Context window | 131,072 tokens |
| Modalities | Text and images in; text out |
| Knowledge cutoff | 4 January 2026 |
| Distilled from | Muse Spark 1.2 |
| Runs on | 24GB GPU (17GB 4-bit build), 32GB GPU, or an M4/M5 Max Mac |
| Price | Free to self-host; no first-party Meta API |
| MCP Atlas | 75.5 (Qwen 3.6 27B: 62.5, Gemma 4 31B: 54.2) |
| SWE-bench Pro | 51.2% (Qwen 50.2, Gemma 36.9) |
| Best for | Local agents, tool use, private/offline work, commercial self-hosting |
| Limitations | Mid-size not frontier; trails Qwen on computer use; vendor-reported benchmarks only |
Why this release matters
Meta spent 2023–2025 as the standard-bearer for open-weight AI, then reversed course: Muse Spark, the April 2026 flagship from Meta Superintelligence Labs, was proprietary, and Muse Spark 1.1 in July became the first Meta model developers had to pay for. The Meta provider page covers that pivot in full. Glimmer is the first real step back.
Three things make it more than a gesture:
The licence is genuinely open. Llama shipped under a bespoke community licence that restricted commercial use above 700 million monthly active users and drew sustained criticism for being marketed as open source when it did not meet the Open Source Definition. Apache 2.0 has none of that: commercial use, modification, redistribution and sublicensing, no caps (VentureBeat).
It is built for agents, not chat. Meta trained Glimmer specifically for multi-step tool use, failure recovery and long-running local workflows — schedule management, file organisation, coding — rather than for conversational benchmarks. That is the axis it wins on.
More is promised. Meta’s superintelligence chief has committed to releasing open weights for Muse Spark 1.2 — the full flagship, not a distillation — “soon” (The Register). If that ships, it is a far bigger event than Glimmer itself. Until it does, treat it as an intention.
Architecture
Glimmer is a dense model, which is the interesting design choice. Almost every recent open release at this scale — gpt-oss, the larger Qwen and GLM models — is mixture-of-experts, which buys cheap FLOPs at the cost of holding every expert in memory. Meta went the other way: all ~29.6B parameters activate on every token, giving predictable latency and no routing overhead, which matters for long agent loops where consistency beats peak throughput.
| Component | Detail |
|---|---|
| Language model | 52 layers, hidden dimension 6,656, SwiGLU FFN (19,968 intermediate) |
| Attention | Gated grouped-query attention, 32 query heads / 2 KV heads (16:1) |
| Attention pattern | Local, Local, Local, Global — 2,048-token sliding window on the local layers |
| Positional encoding | RoPE with NoPE on the text decoder; 2D RoPE for vision |
| Vision encoder | ≈1.8B ViT-style perception encoder, up to 4,096 visual tokens per image |
| Vocabulary | 202,048 tokens (200K BPE + 2,048 special) |
| Languages | 100+ |
The 16:1 GQA ratio is the detail that makes the local story work. It shrinks the KV cache so far that the full 131K context fits on a 24GB card without quantising the cache — the usual bottleneck when people try to run long-context models at home.
Hardware: what you actually need
Meta shipped three weight builds, and the middle one is the sweet spot.
| Build | Size | Target hardware | Quality loss vs BF16 |
|---|---|---|---|
| Full precision (BF16) | ~60GB | 64GB card, or 1×80GB H100 | — |
| K-Quant-Dynamic | ~30GB | 32GB GPU (RTX 5090) | ~0.2% |
| K-Quant-17GB | ~17GB | 24GB GPU (RTX 4090 / 3090) | ~1.0% |
Apple Silicon works too — an M4 Max or M5 Max with enough unified memory runs the quantised builds directly. For fine-tuning, LoRA fits on a single 80GB H100; a full fine-tune needs 8×80GB with FSDP or ZeRO-3.
Speculative decoding is the performance story. Meta ships a separate DFlash drafter (Muse-Glimmer-30B-assistant), a block-diffusion model that proposes 16 tokens per pass. The reported speedups are large: 74.9 → 233.4 tokens/second on an RTX 5090 (3.1x), 26.6 → 50.2 on an M5 Max, and 23.7 → 37.8 on an M4 Max. If you are running Glimmer locally, running it without the drafter leaves most of its speed on the table.
Day-0 support landed in Transformers, llama.cpp (including the WebUI), vLLM, Ollama and LM Studio, with GGUF weights published alongside the safetensors.
Benchmark performance
Every number below is Meta-reported from the model card. Independent evaluations had not landed at the time of writing, and our methodology treats vendor figures as a ceiling.
| Benchmark | Muse Glimmer | Qwen 3.6 27B | Gemma 4 31B |
|---|---|---|---|
| MCP Atlas (agentic tool use) | 75.5 | 62.5 | 54.2 |
| DeepSearch QA | 74.6 | 71.1 | 61.7 |
| AA-LCR (long-context reasoning) | 80.0 | 73.3 | — |
| τ³-Banking (agentic) | 23.5 | 16.7 | — |
| SWE-bench Pro | 51.2% | 50.2% | 36.9% |
| AIME 2026 (maths) | 94.7 | 94.1 | 89.2 |
| Charxiv Reasoning | 78.8 | — | — |
| SWE-bench Verified | 76.0% | 77.2% | 66.6% |
| GPQA Diamond | 83.5 | 84.2 | 85.7 |
| OSWorld-Verified (computer use) | 65.9 | 75.6 | — |
| Terminal-Bench 2.1 | — | 60.7 | — |
The shape is clear and worth stating plainly: Glimmer wins where an agent has to plan, call tools and recover from failure, and loses where it has to drive a GUI or a terminal. MCP Atlas is the flagship result — 13 points clear of Qwen and 21 clear of Gemma — and it is the benchmark that most closely matches what Meta says the model is for. By third-party tallies it takes the best result on 12 of 24 published rows, beating Gemma on 19 of 24 and Qwen on 14.
What it is not: a frontier model. Against the top of our best open-source LLMs board — DeepSeek V4 at 80.6% SWE-bench Verified, Kimi K3 at the top of the independent intelligence index — Glimmer is a tier below on raw capability. That is by design; those models need server-class hardware and Glimmer needs a gaming GPU.
How Muse Glimmer compares
- vs Muse Spark 1.2 — Glimmer is the distilled, open, local sibling. Spark 1.2 is the closed flagship with a 1M-token context, metered at $1.25/$4.25 per MTok; Glimmer is free, runs offline, and gives up substantial capability for it. Different jobs, not competing tiers.
- vs Qwen 3.6 27B — the closest match on size, licence (both Apache 2.0) and hardware target. Glimmer leads on agentic orchestration and tool use; Qwen leads on computer use, terminal work and SWE-bench Verified. If your workload is MCP servers and tool chains, take Glimmer; if it’s driving a desktop or a shell, take Qwen.
- vs gpt-oss 20b — the most direct philosophical comparison: both Apache 2.0, both with adjustable reasoning effort, both aimed at self-hosting. gpt-oss-20b is smaller (14GB) and text-only; Glimmer is multimodal, longer-context and stronger on agentic work.
- vs Gemma 4 31B — Glimmer wins on 19 of 24 published rows, though Gemma holds GPQA Diamond and remains the better pick at genuinely small sizes.
- vs Llama 4 — Glimmer supersedes Llama as Meta’s open-weights flagship, with a much better licence. Llama 4 Scout keeps the context-length record (up to 10M) and a broader fine-tuning ecosystem for now.
For the hardware-first view of where Glimmer fits alongside these, see best local LLMs.
Known limitations
Vendor-reported benchmarks only. No Artificial Analysis entry, no arena Elo, no standardised third-party SWE-bench run at the time of writing. Treat the table above as Meta’s best case.
Deliberately mid-size. At 30B, Glimmer is not competing with Kimi K3, Qwen3.8-Max or the frontier — a point critics have made pointedly, framing it as positioning for enthusiasts and small-to-medium businesses rather than a bid for open-model leadership (The Register).
Weaker at computer use and terminal work. OSWorld-Verified and Terminal-Bench 2.1 both go to Qwen 3.6 27B. If your agent lives in a shell, this is the wrong model.
Open weights, not open source in full. Meta published weights, quantisations, a drafter model and documentation — not the training data or the full training pipeline. The licence is genuinely open; the artefact is still a weights release.
The safety trade-off is real. Open weights can be fine-tuned to strip guardrails, run with no server-side moderation and no ability for Meta to update or revoke them — and a locally-run model with tool access is a materially different security proposition from a local chatbot. That is the standing objection to every open-weights release, and it applies here in proportion to the fact that Glimmer is specifically built to use tools.
Text output only. Images in, text out — no audio, no image or video generation. Meta’s generative media work lives in the separate Muse Image and Muse Video models.
FAQ
What is Muse Glimmer?
Muse Glimmer is Meta’s 30-billion-parameter open-weights model, released on 10 August 2026 under an Apache 2.0 licence. It is distilled from the proprietary Muse Spark 1.2 and built for local agentic work — multi-step tool use, planning and failure recovery — on a single consumer GPU.
Is Muse Glimmer open source?
The licence is: Apache 2.0 is OSI-approved and permits commercial use, modification and redistribution with no revenue or user cap — which makes it more permissive than the Llama Community Licence it supersedes. But like almost every model in this class, Meta released the weights, not the training data or full pipeline, so “open weights” is the more accurate term.
What hardware do I need to run Muse Glimmer?
A 24GB GPU (RTX 4090 or 3090) runs the 17GB 4-bit build with about 1% quality loss; a 32GB GPU (RTX 5090) runs the ~30GB dynamic-quantised build with about 0.2% loss. An M4 or M5 Max Mac with sufficient unified memory works too. Full BF16 precision needs about 60GB — a single 80GB H100.
How fast is Muse Glimmer locally?
On an RTX 5090, roughly 75 tokens/second on its own and up to 233 tokens/second with Meta’s DFlash speculative-decoding drafter — a 3.1x speedup. An M5 Max goes from 26.6 to 50.2 tokens/second, an M4 Max from 23.7 to 37.8.
Is Muse Glimmer better than Qwen 3.6?
At agentic work, yes on Meta’s numbers — MCP Atlas 75.5 vs 62.5, DeepSearch QA 74.6 vs 71.1, SWE-bench Pro 51.2% vs 50.2%. At computer use and terminal work, no: Qwen 3.6 27B leads OSWorld-Verified (75.6 vs 65.9), Terminal-Bench 2.1 and SWE-bench Verified. Both are Apache 2.0 and both fit a 24GB card.
How much does Muse Glimmer cost?
Nothing to download and self-host. There is no first-party Meta API for Glimmer — a deliberate contrast with Muse Spark 1.2, which is metered at $1.25 in / $4.25 out per million tokens ($0.10/$0.20 on the data-contributor tier). Third-party inference hosts offer per-token access at rates that vary.
Does this mean Meta is back to open source?
Partly. Glimmer is Meta’s first Apache 2.0 model and its first significant open release since the pivot to the proprietary Muse Spark line. Meta has also said open weights for Muse Spark 1.2 — the full flagship — are coming “soon”. Until that lands, Meta’s flagship remains closed and metered, and Glimmer is a distillation of it rather than the thing itself.
Last verified 12 August 2026. All benchmark figures are Meta-reported from the Muse Glimmer model card; no independent evaluations had been published at the time of writing. Confirm against the Hugging Face model card before relying on specific numbers.