Meta Muse Glimmer: 30B Apache 2.0 open-weight model lands for on-device agentic workflows

3 sources

Meta Superintelligence Labs released Muse Glimmer on 2026-08-10: a 30B-parameter model released under Apache 2.0, weights on Hugging Face, integrations planned with llama.cpp, MLX, ExecuTorch, Ollama, LM Studio, Unsloth, vLLM, SGLang, Together AI, Fireworks AI, OpenRouter; partner work with AMD, Arm, Dell, Intel, NVIDIA. Built for end-to-end agentic task completion, reliable tool use, multi-step reasoning, and failure recovery; quantized to ~4-bit the LM weights shrink to under 20 GB, designed to fit within 24/32 GB device envelopes alongside KV cache, a perception encoder, and a DFlash drafter; the LM Studio build runs at 18.16 GB on machines with 32 GB+ RAM. Vision-capable via a dedicated perception encoder. Trained and evaluated on DeepSearch QA, MCP-Atlas, τ-Bench, SWE-Bench — specific scores were not extracted in the captured summary, so describe capabilities qualitatively.

Why now

The story is the editorial frame for 'Meta bets that on-device open-weight agentic models are the next surface' — useful because creators who previously needed frontier API access can now consider on-device workflows for routine agentic tasks, and the Apache 2.0 license removes the friction that closed-weight agentic models impose.

Why it is worth publishing

Demo potential: a side-by-side of the same agentic task (e.g., a 5-step research-and-summarize workflow) run via Muse Glimmer locally on a Mac vs the cloud frontier API, measuring latency, cost, and reliability.

Evidence basis

Official Meta Research blog + Simon Willison's independent commentary (2026-08-10) + The Decoder editorial framing

“Meta just shipped a 30B open-weight model under Apache 2.0 that fits in a 32 GB Mac — and it's built specifically for the agentic workflows every solo creator has been paying frontier-API prices for.”

Angle

Frame the story as 'Apache 2.0 + 30B + on-device = the first credible open-weight agentic surface for solo creators and small teams'. The license is the differentiator from closed-weight agentic models, not the parameter count.

Format

Long-form explainer

Demo idea

Record a 10-minute side-by-side: 1 min 'what shipped today', 2 min 'what Apache 2.0 + on-device changes', 3 min running the same 5-step research-and-summarize agentic task on Muse Glimmer locally vs a frontier cloud API, 2 min on hardware setup, 2 min 'when to use which'.

Platform notes

Do not quote specific DeepSearch QA / MCP-Atlas / τ-Bench / SWE-Bench scores without re-extracting them from the Meta blog — only the qualitative framing was captured. The Simon Willison post corroborates the 30B count, Apache 2.0 license, and the 18.16 GB LM Studio figure.

Usable claims

Meta Superintelligence Labs shipped Muse Glimmer on 2026-08-10: a 30B-parameter model released under Apache 2.0, with weights on Hugging Face; built for end-to-end agentic task completion and designed to run locally on a Mac or PC with a single consumer GPU.
Quantized to ~4-bit, Muse Glimmer's LM weights shrink to under 20 GB, designed to fit within 24 GB or 32 GB device envelopes alongside KV cache, a perception encoder, and a DFlash drafter; the LM Studio build runs at 18.16 GB on machines with 32 GB+ RAM.

Evidence pipeline

From the news

Breakdown

Muse Glimmer is the first credible Apache 2.0 + on-device open-weight agentic surface for solo creators — the editorial framing ('30B + Apache 2.0 + fits in a 32 GB Mac') is useful but turns into a benchmark overclaim if you quote specific DeepSearch QA / MCP-Atlas / τ-Bench / SWE-Bench scores without re-extracting them from the Meta blog. This explainer uses the on-device + Apache 2.0 combination as the lens to compare Muse Glimmer to frontier closed-weight agentic APIs, while keeping capability claims qualitative (trained for end-to-end agentic task completion, reliable tool use, multi-step reasoning, failure recovery).

Sources

Risks

Refer to the benchmarks by name and describe the model qualitatively ('trained for end-to-end agentic task completion, reliable tool use, multi-step reasoning'). Open the Meta blog post directly to extract specific scores before stating them on the record.

Demo ideas

Hardware test: same 5-step agentic task on MacBook M4-Max, M5-Max, RTX-5090 — capture latency and reliability, not specific throughput numbers.
License comparison: Apache 2.0 (Muse Glimmer) vs closed-weight frontier (GPT-5.6, Mythos) — what each license lets you ship without negotiation.
Migration guide: which existing on-device open-weight workflows (today: Llama 3.x, Qwen3-Omni, Mistral Magistral) can swap to Muse Glimmer without retraining.