Meta Superintelligence Labs released Muse Glimmer on Monday: a 30-billion-parameter multimodal model, open weights, Apache 2.0, published on Hugging Face. Mark Zuckerberg said the weights for Muse Spark 1.2 — Meta’s latest foundation model — will follow.
The parameter count is not the interesting number. The memory footprint is. At full precision a 30B model needs more than 55 GB. Meta compresses the weights to roughly 4-bit, which brings the language model under 20 GB and leaves headroom inside a 24 GB or 32 GB envelope for the KV cache, the vision encoder and the speculative-decoding drafter. Two quantised builds ship: K-Quant-Dynamic targets 32 GB at 0.2% average accuracy degradation, K-Quant-17GB targets 24 GB at 1.0%, measured across 15 common benchmarks.
Then there is speed, which is what usually kills local models inside an agent loop. Meta pairs the model with DFlash, a block-diffusion drafter that proposes 16 tokens in a single forward pass while the main model verifies the block in parallel. On an RTX 5090 at batch size 1, throughput goes from 74.9 to 233.4 tokens per second — a 3.1x speedup. Apple silicon gains less, roughly 1.5x to 1.8x on M4 Max and M5 Max. A model you can run is not the same thing as a model you can run in a loop.
Where it wins, and where it doesn’t
Meta benchmarks Muse Glimmer against Gemma4-31B and Qwen3.6-27B in thinking mode. It leads on MCP Atlas at 75.5 against 54.2 and 62.5 — a wide margin on tool orchestration. It also leads DeepSearch QA at 74.6, Gaia2 at 43.3 and SWE-Bench Pro at 51.2, with AIME 2026 at 94.7 on the reasoning side.
The losses are the more useful half of the scorecard. Qwen3.6-27B stays ahead on OSWorld-Verified, 75.6 to 65.9. It leads TerminalBench 2.1 at 60.7 against 51.7, and edges SWE-Bench Verified 77.2 to 76.0. The pattern is consistent: Muse Glimmer is stronger at planning and calling tools, weaker at driving a computer and working a terminal. Those last two are not edge cases. For a lot of teams, that is the job.
Our take: Yesterday Microsoft put a token meter on its own engineers. Today Meta handed out a model that removes the meter for a whole class of always-on work — the scheduled, repetitive, low-glamour agent tasks where a per-token bill compounds fastest. That is the real trade this release forces: rent frontier capability by the token, or own a fixed cost that runs offline and never leaves your building. But read the losing benchmarks before you move anything. And note the safety line Meta published: on the Siren AgentDojo suite the attack success rate is 28.4 with utility at 94.2 — close to three in ten prompt-injection attempts landing. Meta’s own guidance is to wrap it in system-level guardrails rather than expose it as a bare endpoint, which is exactly the discipline most firms skipped when they handed six in ten agents allow-all access.
Why Meta is giving it away
Constellation Research’s Larry Dignan reads the strategy plainly: Meta already monetises through advertising, so it can “box in foundation labs because it doesn’t depend on selling foundation models.” Every open-weight release costs Meta little and costs the labs selling inference a great deal. Zuckerberg’s framing is national rather than commercial — that the US and its allies should lead the open-source AI ecosystem. Both things can be true.
Distribution is the tell. Optimised support for llama.cpp, MLX and ExecuTorch is due within days, with planned access through Ollama, LM Studio, vLLM, SGLang, Together AI, Fireworks AI and OpenRouter, plus device-level work with AMD, Arm, Dell, Intel and NVIDIA. This is not a research drop.
What to watch
- Muse Spark 1.2 weights. Zuckerberg promised them. A genuinely open frontier-class model from a US lab changes the pricing conversation far more than a 30B does.
- Whether the computer-use gap closes. If OSWorld and TerminalBench stay with Qwen, Meta owns the orchestration layer and someone else owns the hands.
- On-prem procurement. Healthcare, legal, financial services and public sector are the buyers who care about data residency and offline operation more than about the last two benchmark points.
- Whether OpenAI and Anthropic respond with weights. The pricing squeeze only works if the labs selling tokens feel it.
- Injection defences. A 28.4 attack success rate is a governance problem before it is a model problem — and it needs a named owner on every agent you run.