Muse Glimmer 30B

Glimmer is a dense decoder-only transformer with a separate vision encoder. The 29.6B total includes a roughly 1.8B-parameter ViT-G/14 Perception Encoder, which Meta releases as a frozen component feeding a projector. That is why native_multimodal is false here: images reach the language model through an attached encoder, not early fusion. The language model has 52 layers, hidden size 6,656, SwiGLU feed-forward width 19,968 and gated attention. Attention repeats a pattern of three local sliding-window layers (window 2,048) and one global layer. RoPE (theta 500,000) applies only on local layers. Grouped-query attention uses 32 query heads and 2 KV heads. Only a quarter of the layers keep a full-length KV cache, so Meta measures the whole 131,072-token context at 2.1 GiB of KV cache and compute buffers on the 4-bit build. The vocabulary is 202,048 tokens (200,000 BPE plus 2,048 special), the same size as Llama 4's, and each image can use up to 4,096 visual tokens. Training token count, compute and data mix are undisclosed beyond "a similar data mix as the teacher".

Agentic work is the design target, and Meta's numbers are strongest there (cap_agentic 8.2): MCP Atlas 75.5, DeepSearchQA 74.6, Gaia2 43.3 and OSWorld-Verified 65.9, all vendor-reported. Function calling (7.5) is reliable on single calls per Meta's docs. The model supports one tool call per turn and no parallel calls, which costs round-trips in fan-out agents. Coding (7.8) rests on SWE-bench Verified 76.0, SWE-Bench Pro 51.2 and Terminal-Bench 2.1 51.7, solid for 30B but well behind Qwen3.8-27B on Terminal-Bench. Reasoning (7.8) pairs GPQA Diamond 83.5 with a modest HLE 22.0, and math (8.3) reflects AIME 2026 94.7. Instruction following (8.0) comes from IFBench 77.0. Vision (7.5) is competitive at this size: MMMU-Pro 74, CharXiv Reasoning 78.8, ScreenSpot Pro 75.4. Document parsing (6.5) is weaker: OmniDocBench v1.5 75.8 on Meta's modified scorer, against 91.1 that Qwen reports for Qwen3.8-27B. Long context (6.5) combines strong in-window reasoning (AA-LCR 80.0, BEAM-128K 65.1) with a 131K window, half or less of what its peers offer. Safety calibration (7.0) is covered below. Multilingual and creative writing are unscored because Meta publishes no evaluation for them; it states only that training data covered more than 100 languages. There is no live data access (0).

Every row except LMArena is vendor-reported. Meta ran its model at high reasoning strength, which is also the default. Meta's methodology report says GPQA Diamond, HLE, AA-LCR, MMMU-Pro and Terminal-Bench 2.1 are Artificial Analysis results; Artificial Analysis's own page could not be read, so they are recorded here as Meta-reported. The Qwen3.8-27B column is Qwen's own table, which lists Glimmer with Meta's figures. The two companies report Qwen3.6-27B differently on shared benchmarks (OSWorld-Verified 75.6 from Meta against 63.9 from Qwen; OmniDocBench 77.8 against 89.4), so cross-vendor gaps under about 10 points are within harness noise. Meta's other comparator is Google's Gemma 4 31B (released 2026-03-31), which trails Glimmer on most of Meta's agentic rows but ranks higher on LMArena (1451 ±8). The one independent figure puts Glimmer 104th of 402 text models on 3,666 votes, below Qwen3.8-27B and Gemma 4 31B and far above Llama 4 Scout (1321).

No independent throughput or latency measurement has been published, so the speed fields are null. Meta's own figures use the 17 GB 4-bit build with batch size 1: an RTX 5090 runs 74.9 tokens/s plain and 233.4 tokens/s with the DFlash drafter (3.1x) under llama.cpp. On ExecuTorch, an Apple M4 Max runs 23.7 tokens/s plain and 37.8 with DFlash, and an M5 Max 26.6 and 50.2. Meta's Together AI page relays Together's own measurement on two H100s under long-prompt agent traffic: 307 ms median time to first token and 105 output tokens/s per request. The practical caveat is that reasoning is always on, and Meta says the model routinely writes multi-thousand-token chains of thought. Time to a final answer is longer than the token rate suggests, and the low reasoning strength is the lever for interactive use.

Meta sells no tokens for this model, so every price field is null. The cost is your hardware or a third-party host's meter.

Four Meta repositories on Hugging Face cover the release. Muse-Glimmer-30B holds the BF16 safetensors (about 60 GB, Transformers 5.15 or later). Muse-Glimmer-30B-GGUF holds two 4-bit text builds: K-Quant-17GB (Q4_K_M, 16.8 GB download, about 1.0% average degradation, targets 24 GB cards) and K-Quant-Dynamic (Q4_K_XL, 19.7 GB, about 0.2%, targets 32 GB), plus the vision projector and drafter. The other two are the DFlash drafter checkpoint and prebuilt ExecuTorch exports for phones, tablets and Macs. Runtimes: llama.cpp from release b10353, SGLang, ExecuTorch, Ollama, LM Studio and Unsloth. vLLM support was still an unmerged pull request when its guide was read on 2026-09-21, so it runs from a dedicated Docker image rather than a pip install. Hosted access comes from Together AI (serverless and dedicated, served in FP8 on MLP layers), Fireworks AI and OpenRouter. There is no first-party Meta endpoint, and Meta's launch partner list names no hyperscaler cloud platform. For customization, Meta documents LoRA, QLoRA and full fine-tuning, reinforcement learning, and PyTorch TorchTitan. AMD, Arm, Dell, Intel and NVIDIA are named hardware partners.

Meta evaluated Glimmer under its Advanced AI Scaling Framework. The model falls below the framework's "Frontier AI" threshold because it is less capable than Muse Spark. Meta's preparedness team rates chem/bio risk moderate or lower and infers the same for cyber and loss of control. On chem/bio benchmarks it sits in line with Gemma 4 31B and Qwen3.6-27B. Train-time mitigations include safety SFT, safety RL, and privacy training on data minimization and local-first execution. Meta's own safety table is mixed. In Siren AgentDojo prompt-injection tests the attack success rate is 28.4%, better than Qwen3.6-27B's 40.3% but worse than Gemma 4 31B's 25.6%. In CIMemories, Glimmer leaks sensitive attributes in inappropriate contexts 26.4% of the time, more than double Gemma's 12.1%. No moderation layer ships with the weights, and Meta advises deploying the model inside a system with guardrails and human confirmation for irreversible actions. Self-hosted, no data leaves your infrastructure. A Meta Usage Policy sits alongside Apache 2.0, listing prohibited uses and excluding users under 18. Compliance certifications depend on your host, not the model.

Launch partners named by Meta: AMD, Arm, Dell, Fireworks AI, Hugging Face, Intel, llama.cpp, LM Studio, NVIDIA, Ollama, OpenRouter, SGLang and RadixArk, Together AI, Unsloth, and vLLM and Inferact. OpenClaw is the reference agent harness, and Meta also cites Hermes Agent compatibility. A public cookbook covers quickstart, tool-use fundamentals, a full agent recipe, inference servers, hardware platforms and hosted APIs, and each recipe states its precision, server and observed VRAM. On 2026-09-21 Hugging Face showed about 292,000 downloads over the previous 30 days for the main repository (1,926 likes) and about 416,000 for the GGUF repository. That is a growing footprint, though an order of magnitude below Qwen3.8-27B's roughly 7.2 million over the same period.