Muse Glimmer 30B: Agents That Stay on the Bench
Meta distilled Spark for the workshop — a 30B model that runs locally and keeps the job on the bench.
Meta released Muse Glimmer 30B on August 10, 2026 as an open-weight 30B model for local agent workflows — meta-models/Muse-Glimmer-30B on Hugging Face (Apache 2.0). It is a dense 29.6B causal transformer distilled from Muse Spark, with a 1.8B ViT-G/14 perception encoder for interleaved text and images.
Finding
Glimmer is built for autonomous runs on consumer hardware. The full model needs about 64 GB VRAM; with 4-bit K-Quant it compresses to roughly 17 GB for a 24 GB card and about 20 GB for 32 GB, with 0.2 to 1.0 percent average degradation across 15 benchmarks.
Local speed comes from a DFlash drafter — a five-layer speculative decoder that proposes 16-token blocks for the main model to verify in parallel. On a single stream with greedy decoding, measured throughput goes from about 75 to 233 tokens per second on an Nvidia RTX 5090 (3.1x), 24 to 38 on an Apple M4 Max (1.5x), and 27 to 50 on an M5 Max (1.8x).
It keeps the agentic stack in one model: multi-step planning, schema-precise tool calling, long-horizon reasoning with controllable effort (low / medium / high / xhigh via “Reasoning strength”), failure recovery after tool errors, scaffold compatibility across OpenClaw and Hermes Agent, and multimodal input over 131,072 tokens. Max visual tokens per image is 4096.
On benchmarks reported in the model card, the high-reasoning setting compares to Gemma 4 31B and Qwen 3.6 27B in the same size class. Glimmer leads on several workflow-oriented tests: MCP-Atlas 75.5, DeepSearch QA 74.6, Tau-3 Banking 23.5, WildClawBench 47.6, Gaia2 43.3, SWE-Bench Pro 51.2, SciCode 43.6, IFBench 77.0, AIME 2026 94.7, AA-LCR 80.0, and Beam128K 65.1. It trails Qwen on SWE-Bench Verified (76.0 vs 77.2), TerminalBench 2.1 (51.7 vs 60.7), ScreenSpot Pro, OmniDocBench, MMMU Pro, and GDPVal-AA v2. For agentic safety, Siren AgentDojo attack success is 28.4 percent at 94.2 utility (Gemma 25.6 / 90.8, Qwen 40.3 / 92.7).
Sampling recommendation from the card is temperature 1.0, top_p 0.95, top_k 64.
Meaning
The agentic workload studied in “Coding Agents Are Not Chatbots” does not fit a counter that clears between single requests. It fits a workshop bench — tools and manuals stay laid out for the duration of one job so the next step is fast. The Copilot traces showed why that matters: cache reuse rises from 45 percent to 93 percent within a turn, idle drops from 172 seconds across turns to 1.2 seconds within a turn, and input dominates output 68K to 247 tokens.
Glimmer is the model side of that same shift. If the bench must stay set up for minutes, the model that sits on it cannot live only in the cloud. It has to fit the 24 to 32 GB envelope, keep its KV cache and perception encoder resident, and generate fast enough to feel interactive while a tool chain is in progress. Distillation plus 4-bit quantization is the cost paid to fit the bench. DFlash is the mechanism that keeps generation moving without waiting token by token.
That choice has a trade. A 30B local model will not match Spark or Frontier-class models on every general capability, and the tables show it. Where it compensates is where the bench matters most — tasks that require holding context, calling tools in sequence, and continuing after a failure, rather than answering one perfect prompt. The reported 94.7 on AIME 2026 and 75.5 on MCP-Atlas suggest the long-horizon training transferred, but the unit is the turn and the session, not the single request.
What remains open is the same as for any local model in that study’s structure: long-tail retry behavior (9 percent of turns with failures, 4x retries), compaction thresholds when prompts grow, and model-switch cold starts that drop reuse to 8 percent. A local model that can be kept warm avoids the switch entirely, which is itself a performance strategy.
Connection
The move follows the same pattern that made the workshop faster in the first place. FGN in WeatherNext 2 reduced ensemble spread to one globally applied perturbation instead of an iterative diffusion loop. Glimmer reduces agentic latency to one locally verified block diffusion instead of iterative single-token verification. In both cases the cost that was split across many steps is gathered into one coherent update that preserves structure.
Source: Meta Superintelligence Lab, Muse Glimmer 30B model card, August 10, 2026. Hugging Face. Sampling and benchmark figures as reported in the card.