Muse Glimmer is Meta’s pivot back to open source — a 30-billion-parameter dense multimodal model released on August 10, 2026 under Apache 2.0, the most permissive license Meta has ever applied to a foundation model. It’s designed for one thing: running autonomous agents on hardware you own.
Five days after launching Muse Spark 1.2 as a closed API product, Meta open-sourced Glimmer and committed to releasing Spark 1.2 weights as well. The message: distribution matters more than control.
Architecture
| Spec | Value |
|---|---|
| Parameters | 30B dense (29.6B including vision encoder) |
| Type | Dense multimodal transformer |
| Context window | 131K+ tokens |
| Input modalities | Text, image (vision encoder included) |
| Output modalities | Text |
| Precision | BF16 (quantizations available) |
| License | Apache 2.0 |
DFlash speculative decoding
Muse Glimmer ships with a quantized DFlash drafter — a small companion model that accelerates generation through speculative decoding:
| Hardware | Speedup |
|---|---|
| NVIDIA RTX 5090 | 3.1× |
| Apple M5 Max | 1.8× |
| Apple M4 Max | Supported (speed TBD) |
This makes the 30B model fast enough for fluid conversation and real-time agent interaction on consumer hardware.
Local deployment
The whole point of Glimmer is running locally. Meta benchmarks it at 24GB VRAM after quantization — a single RTX 4090 or Apple Silicon Mac handles it.
Day-0 support across the ecosystem:
| Runtime | Status |
|---|---|
| Hugging Face transformers | Day 0 |
| llama.cpp | Day 0 |
| vLLM | Day 0 |
| MLX (Apple Silicon) | Day 0 |
| ExecuTorch (mobile) | Coming days |
| Ollama / LM Studio / Unsloth | Coming days |
| Together AI / Fireworks / OpenRouter | Available |
Benchmarks
Meta compares Glimmer against Gemma4-31B and Qwen3.6-27B — the relevant size class. No AA Intelligence Index score or independent evaluation has been published at launch. Meta claims “strong performance on key agentic use cases” but the specific numbers are from self-selected benchmarks in the launch post.
How it compares
- vs Muse Spark 1.2: Spark is the frontier model (AA Index 54, TB 2.1 82.9%). Glimmer is the 30B local model. Different jobs — Spark for cloud inference, Glimmer for edge agents.
- vs Gemma4-31B / Qwen3.6-27B: Direct size-class competitors. Meta claims Glimmer performs strongly against both on agentic benchmarks. All three target similar hardware profiles.
- vs GLM-5.2 / Kimi K2.7 Code: GLM-5.2 (MIT, 753B MoE) and Kimi K2.7 (Modified MIT, 1T MoE) are much larger open models requiring server hardware. Glimmer trades capability for true consumer-device deployment.
Why it matters
This is Meta’s first Apache 2.0 model release — ever. After the Llama era of custom licenses with commercial caps, and the Muse era of proprietary API-only models, Apache 2.0 signals a strategic shift. Mark Zuckerberg accompanied the launch with a 14-page letter arguing for broadly available AI systems, and Meta committed to opening Spark 1.2 weights “in the coming weeks.”
The practical implication: if you want to run an agent locally for privacy, cost, or latency reasons, Glimmer is the most capable Apache 2.0 model in the 30B class — and it ships with the runtime ecosystem to actually do it.
Limitations
- 30B is not frontier. Don’t expect Opus 5 or Fable 5 reasoning quality. This is a local-class model.
- No independent benchmarks. Self-selected comparisons at launch.
- 131K context. Significantly shorter than the 1M context windows on frontier models.
- No official hosted pricing from Meta. For API access, you depend on partners.
For the full AI model landscape, see the 2026 LLM guide.
Best for
- Always-on local coding agents on laptops and workstations
- Privacy-sensitive agentic workflows that can't send data to the cloud
- Function calling and tool use on consumer hardware
- LLM-as-a-judge evaluation locally
- Fine-tuning base for domain-specific agents
Pros & cons
Strengths
- Apache 2.0 — Meta's first fully permissive open-source model, no revenue caps or usage restrictions
- Runs on 24GB VRAM (single consumer GPU) after quantization — true local deployment
- DFlash speculative decoding drafter ships with weights — up to 3.1× faster generation on RTX 5090
- Strong for 30B class on agentic benchmarks — competes with Gemma4-31B and Qwen3.6-27B
Limitations
- 30B class — not frontier-competitive with 100B+ models on hard reasoning or deep coding
- No independent benchmark scores published at launch — Meta's comparisons are self-selected
- 131K context is small vs 1M context frontier models
- No official hosted API from Meta — depends on partner availability
Frequently asked questions
When was Muse Glimmer released?
August 10, 2026, with weights immediately available on Hugging Face under Apache 2.0.
What hardware does Muse Glimmer need?
Full BF16 is ~30B parameters. With quantization (4-bit), it fits in 24GB of VRAM — an RTX 4090, M4 Max, or similar. The DFlash drafter enables fluid generation speed at 3.1× on RTX 5090 and 1.8× on M5 Max.
How does Muse Glimmer relate to Muse Spark?
Different roles. Muse Spark 1.2 is Meta's frontier model (API-only at launch, open weights coming). Muse Glimmer is a smaller 30B model designed specifically for local deployment. Glimmer is distilled from Muse — optimized for edge, not cloud.
Why Apache 2.0?
This is Meta's first Apache 2.0 foundation model release. Previous Muse models were proprietary, and Llama used Meta's custom license with commercial restrictions. Apache 2.0 means no revenue caps, no usage restrictions, and full commercial freedom.