Muse Glimmer — Meta comes back to open weights, and aims them at your GPU

The take

What Meta actually shipped, verified against their own research blog and the model card: ~29.6B parameters dense, 52 layers, including a ~1.8B ViT-G/14 perception encoder for interleaved image input; 131,072+ context; knowledge cutoff January 4, 2026. Released artifacts are full BF16 weights, **two** 4-bit quantizations, a DFlash speculative-decoding drafter head, and the perception encoder — all Apache 2.0. Not just the weights: the drafter and the encoder too. Meta has not released training data or training code, which is the standard caveat on every "open source" model release and worth stating plainly.

The engineering is aimed squarely at the constraint that matters. At full precision a 30B model needs over 55 GB. Meta quantizes to roughly 4-bit, putting the language model under 20 GB, and states the target explicitly: leave enough headroom for the KV cache, the vision encoder and the drafter to run **simultaneously inside a 24 GB or 32 GB envelope**. Their degradation table claims 0.2% at the 32 GB target and 1.0% at the 24 GB target, averaged across 15 benchmarks. The DFlash drafter proposes 16-token blocks in a single forward pass and the main model verifies them in parallel — on Meta's numbers an RTX 5090 goes from 74.9 tok/s to 233.4 tok/s. Vendor-measured, and we have not reproduced it, but the architecture is public and the mechanism is standard speculative decoding rather than a claim you have to take on faith.

The honest part of the model card is what makes us trust the rest of it. Meta does not claim Glimmer beats its size class — it says performance is "approximately in line with other models in its size class", and the chem/bio safety table prints rows where **Gemma 4 31B beats it outright** (MBCT 50.6 vs 41.5, VCT 43.5 vs 37.0, WMDP-Chem 80.5 vs 75.2). A vendor publishing the benchmarks it loses is the signal we weight most. So the reason this is a pick is not a claimed score: it is that the model was trained end-to-end around the agent loop — plan, call a tool, read the result, keep going, recover from a failed call — rather than around chat, and then compressed to fit the card most readers actually own.

Our call: **new planner pick at `agents.top`**, alongside Qwen 3.6-35B-A3B rather than displacing it — the benchmarks are vendor-run and four days old, and our own rule is that independent numbers displace incumbents, not vendor tables. New model entry at `/models/muse-glimmer/`. Two things to keep in view. First, the licence is genuinely the story alongside the model: Apache 2.0 has no equivalent of Llama's 700-million-user clause, so this is more open than anything Meta shipped under the Llama name. Second, and worth saying because the framing invites the opposite conclusion: running an agent locally protects your data from the cloud, it does **not** protect you from prompt injection or from the agent taking an irreversible action. Meta recommends human-in-the-loop confirmation on destructive steps. Follow that. Zuckerberg has also said Muse Spark 1.2 weights are coming; if that lands, an American lab will have opened an actual frontier model, which no US lab has done at that tier.

Where this fits

Models: Muse Glimmer 30B · Qwen 3.6-35B-A3B · gpt-oss-20b · Gemma 4 (31B dense + 26B A4B MoE + 12B multimodal)

Hardware: NVIDIA RTX 5090 · NVIDIA RTX 4090 · NVIDIA RTX 3090 (used, single) · Mac Studio M4 Max 64 GB

Sources

• Introducing Muse Glimmer — Meta AI Research (Aug 10, 2026)
• Muse Glimmer 30B model card — Hugging Face
• Muse Glimmer developer documentation — Meta

Next step