Meta Muse Glimmer: A 30B Open-Weight Agentic Model for Local AI

Meta open-weighted its second Muse model in five months on August 10, 2026. Where April's Muse Spark stayed closed and July's Muse Spark 1.1 shipped only as a hosted API, Muse Glimmer — a 30-billion-parameter agentic model — landed on Hugging Face with weights under Apache 2.0, small enough to run on a single consumer GPU.

The pitch, direct from Meta's announcement: local agents, function calling, local coding, and LLM-as-a-judge evaluation, "with or without an internet connection." The model is distilled from Muse Spark, the much larger closed teacher model behind Meta's consumer AI stack, and it arrives one week after Muse Code, the terminal coding agent built on Muse Spark 1.2.

The release lands the same day Zuckerberg published a long essay on meta.com, The Future Is for Everyone, which explicitly says Meta "will resume releasing some open source models soon" — Muse Glimmer is that pledge, shipped as a checkpoint rather than a promise.

TL;DR — what people are asking

What is Muse Glimmer, and why release it now?

Muse Glimmer is a dense, 30-billion-parameter model built specifically for always-on local agent workflows — the opposite design goal from Muse Spark's cloud-scale reasoning. Meta frames the motivation as a gap in the market: "most deployments still depend on cloud infrastructure and network access," and running models locally means AI works "anywhere, anytime, with or without an internet connection."

That's the same throughline Zuckerberg argued in his July WSJ op-ed, The AI Future Is for Everyone — personal AI that runs on hardware people own, not just AI mediated through a lab's servers. Muse Glimmer is that thesis expressed as a checkpoint.

How it was trained

Meta describes a three-phase pipeline distilling agentic reasoning down from a much larger teacher:

1. Pre-Training — trained on Muse Spark's own outputs via logit distillation, using a similar data mix to the teacher model.
2. Mid-Training — longer-context, agent-heavy data with richer reasoning traces, mixed with organic data.
3. Post-Training — supervised fine-tuning combined with on-policy distillation and reinforcement learning across general, reasoning, coding, and agentic domains.

Meta says Muse Glimmer was evaluated under its Advanced AI Scaling Framework and cleared for open-weight release "across all relevant categories" — the same governance process explainx.ai flagged when covering Muse Spark's original launch.

What it's built to do

Per Meta's post, Muse Glimmer targets seven agentic capabilities: end-to-end task completion (measured on DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench), reliable tool use with precise function-call schemas, multi-step reasoning over long horizons, failure recovery (diagnosing bad tool results and retrying instead of halting), multimodal input via a dedicated perception encoder, compatibility with agentic orchestration patterns including OpenClaw, controllable reasoning effort, and multilingual coverage across 100+ languages. Readers building agent stacks around these primitives will recognize the same design goals covered in explainx.ai's agent harness guide and MCP overview.

Benchmarks: Muse Glimmer-30B vs Gemma4-31B vs Qwen3.6-27B

This table reproduces Meta's own published comparison chart — Muse Glimmer-30B in "High Reasoning" mode against Gemma4-31B and Qwen3.6-27B, both in "Thinking Mode." As with every vendor chart, Meta chose the comparison set and the configuration; the same caution explainx.ai applied to Muse Code's launch benchmarks applies here. Test on your own workload before routing production traffic.

General agentic

Agentic coding

Multimodal

† Muse Glimmer's 78.8 edges Qwen3.6-27B's 78.4 on Charxiv Reasoning — the closest three-way result on the whole chart.

Safety

General capabilities and reasoning

The pattern several Hacker News commenters called out: Muse Glimmer wins most agentic and tool-calling categories decisively — MCP Atlas by over 20 points, safety scores by a wide margin — but the general-reasoning and pure-coding gap to Qwen3.6-27B (which explainx.ai covered in its own local development guide) is much closer, and Qwen3.6-27B specifically wins Terminal-Bench 2.1 by nine points. One commenter summarized it as "barely edges out Qwen3.6-27B, except for tool-calling/MCP skills"; another said hands-on use "felt better than the benchmark scores seem to indicate." Both can be true — benchmark suites and day-to-day agent reliability don't always move together.

Running it locally: quantization and DFlash

At full precision, a 30B model needs over 55GB of memory — more than any consumer GPU on the market. Meta's answer is aggressive quantization plus speculative decoding, the same two levers explainx.ai's laptop-runnable model guide uses to rank local models.

Quantization. Weights compress to roughly 4-bit, shrinking the language model to under 20GB — leaving headroom for the KV cache, the perception encoder, and the DFlash drafter inside a 24GB or 32GB memory budget. Meta claims "minimal to no degradation on agentic tasks" from the compression.

DFlash speculative decoding. Muse Glimmer ships with a small companion "drafter" model, DFlash, that proposes a block of several tokens at once; the main 30B model then verifies the whole block in a single parallel forward pass instead of generating token-by-token. Meta's measured decode-speed gains:

Meta says output quality is identical with or without the drafter enabled — DFlash only proposes candidates, the base model still verifies every token.

Availability

Weights are live on Hugging Face today. Meta says integrations for llama.cpp, MLX, and ExecuTorch are landing "in the coming days," alongside Ollama, LM Studio, Unsloth, vLLM, SGLang, Together AI, Fireworks AI, and OpenRouter. PyTorch TorchTitan support is available for further customization, and Meta named AMD, Arm, Dell, Intel, and NVIDIA as hardware launch partners. For a primer on the runtime most local users will reach for first, see explainx.ai's what is llama.cpp guide.

What Hacker News found running it on real hardware

The Hacker News thread on the launch ("Meta Muse Glimmer – open weights 30B local coding model") hit 555 points and 293 comments within hours. Setting aside the corporate-ethics debate that tends to attach to any Meta release, the technical substance is worth reading closely:

• Apache 2.0 is the real headline for developers, not just a footnote. One commenter noted: "If even Meta can use a permissive license for its model weights, so can any other company" — a pointed contrast with Llama's more restrictive terms.
• The tool-calling chat template uses <atem:function_calls> / <atem:invoke> / <atem:parameter> tags. Multiple commenters independently guessed "atem" is "meta" spelled backwards, likely chosen to avoid tokenizer collisions with literal HTML <meta> tags in scraped training data or agent output.
• llama.cpp support landed within hours of the announcement — the merged PR is at github.com/ggml-org/llama.cpp/pull/26841.
• Real hardware numbers, not vendor slides. On an AMD 7900XT (20GB VRAM, ~800GB/s bandwidth), Unsloth's Q4_K_XL GGUF quant ran in ~19GB VRAM with 4 parallel 113K-context slots, hitting ~700 tok/s prompt processing and ~36 tok/s generation — rising to ~60 tok/s once DFlash speculative decoding was switched on. The smaller Q3_K_XL quant dropped to 15.6GB VRAM with the full 131K context at similar speed.
• Dense-vs-MoE tradeoffs matter on unified memory. Both Gemma4 and Qwen3.6 ship mixture-of-experts variants that are less memory-bandwidth-bound; Muse Glimmer is dense, so on Mac and NVIDIA DGX Spark-class unified-memory hardware it's more bandwidth-limited. One back-of-envelope estimate put dense-mode decode around 14-15 tok/s on a DGX Spark at 4-bit before DFlash, rising to 33-60 tok/s once speculative decoding kicked in.
• "Open weight" is not "open source," and several commenters pushed back on Meta's framing for exactly that reason — no training data, no training code, no independent reproducibility. Worth stating plainly rather than glossing over.
• The low end of the hardware range is real but slow — one hobbyist reported running Muse Glimmer on an old 32GB Mac Mini via Ollama; it worked, just slowly.
• Not available in Hong Kong at launch, per one commenter's report — unconfirmed whether mainland China faces the same restriction.

Is Muse Spark 1.2 next for open weights?

The most consequential line in the entire announcement cycle wasn't about Muse Glimmer's benchmarks — it was Zuckerberg's aside that Meta will "soon" open-weight Muse Spark 1.2, the model currently powering Muse Code. Alexandr Wang, Meta's Chief AI Officer, repeated the same commitment in his own thread. Neither gave a date.

That's a real shift from Muse Code's August 5-6 launch post, which described Muse Spark 1.2 as closed "regardless" of what happened with the harness around it. Five days later, that framing changed. If you're evaluating Muse Code for production, it's worth tracking this — an open-weighted Muse Spark 1.2 would mean self-hosting the exact model behind Meta's coding agent, not just calling a hosted API.

The bigger picture: Meta's open-weight strategy is compounding

Muse Glimmer isn't an isolated release — it's the fourth Muse-family announcement since April, and the second in five days after Muse Code. Meta is now running two release tracks in parallel: closed, hosted frontier models (Muse Spark 1.1 via the Meta Model API) and open-weight distillations sized for local hardware (Muse Glimmer, and soon Muse Spark 1.2). That mirrors the split explainx.ai has tracked across the wider industry — see the top 10 open-weight models for laptops and loop engineering for local agents — where the interesting competitive question isn't "closed vs open" in the abstract, but which capability tier a lab is willing to release locally-runnable.

Related on explainx.ai

• Muse Spark and personal superintelligence (April 2026)
• Muse Spark 1.1 and the Meta Model API (July 2026)
• Muse Code: terminal coding agent on Muse Spark 1.2 (August 2026)
• Zuckerberg: The AI Future Is for Everyone (WSJ op-ed)
• Zuckerberg's "The Future Is for Everyone" essay and open-source pledge (August 10, 2026)
• Top 10 open-weight models you can run on a laptop
• Qwen 3.6 27B local dev guide
• Gemma 4 31B on Cerebras
• What is an agent harness? Complete guide
• What is loop engineering?
• What is llama.cpp?

Official source: Meta Superintelligence Labs, Introducing Muse Glimmer (August 10, 2026).

Details reflect Meta's August 10, 2026 announcement, X posts from Mark Zuckerberg and Alexandr Wang the same day, and the Hacker News discussion thread as of publication. Benchmarks are Meta's own self-published comparison chart — verify against your own workload before switching production tooling. Pricing, regional availability, and integration timelines may change; check Hugging Face and Meta's developer docs for current status.