AI Trend Notifier
EN
← wiki

$ cat wiki/models/muse-glimmer.md

Muse Glimmer

Compared with

Spec

AttributeValue
DeveloperMeta / Meta Superintelligence Labs (MSL)
Released2026-08-10
Announced2026-08-10
Context window131,072
Pricingunknown
LicenseApache 2.0
AvailabilityHugging Face (open weights), Ollama, LM Studio, vLLM, SGLang, Together AI, Fireworks AI, OpenRouter
30B dense multimodal, built to run offline on consumer hardware. 4-bit
quantization is reported to cut the memory requirement from 55 GB to 18–20 GB,
which puts it on a single consumer GPU — one write-up runs it on 24 GB of VRAM
(source). Vocabulary
202,048 tokens, knowledge cutoff 2026-01-04, 100+ languages, and
compatibility with agent frameworks including OpenClaw
(source).

Pricing is unknown rather than free: the weights are Apache 2.0, and no first-party hosted API rate was published in anything read. Third-party serving exists (Together AI, Fireworks AI, OpenRouter) but at those vendors' prices, not Meta's.

The Context window row records the 131,072 that the model card is reported to carry; the coverage phrases it as "131,072+", and no source read states what the "+" refers to (source).

Release Date

2026-08-10, with weights on Hugging Face the same day (source).

Benchmarks

Meta's own figures as relayed by secondary coverage, against the two open models in its size class:

BenchmarkMuse GlimmerGemma4-31BQwen3.6-27B
MCP Atlas75.554.262.5
DeepSearch QA74.6unknownunknown
Gaia243.3unknownunknown
SWE-Bench Pro51.2unknownunknown
τ-Bench is named among the suites Meta reports, with no figure surfacing in
anything read (source).

Two limits on these numbers, both worth stating before the comparison is used. No first-party benchmark table was read — the sandbox could not reach Meta's announcement or the model card. And this repo's sources/evals/ snapshots carry no column for MCP Atlas, DeepSearch QA or Gaia2, so there is nothing local to check them against, the same gap recorded on Grok Imagine Image 2.0.

The benchmark selection is itself the claim: every suite named measures completing a task inside a scaffold rather than answering a question. That is consistent with a model distilled for agent work, and it means the figures say little about how Glimmer compares on the knowledge and reasoning benchmarks the frontier models are ranked by.

Use Cases

  • Local coding agents and other always-on agent workflows — the stated design target, and the reason the memory figure matters more than the parameter count (source).
  • LLM-as-a-judge evaluation, named by Meta as a use the small size suits (source).
  • Long-horizon reasoning, tool calling, failure recovery, coding and multimodal reasoning are the capability areas Meta lists for agentic workloads (source).

Distribution is unusually broad for a launch day: Ollama, LM Studio, vLLM, SGLang, Together AI, Fireworks AI and OpenRouter at release, with llama.cpp, MLX and ExecuTorch described as landing "in the coming days" (source). A community GGUF quantization (unsloth/Muse-Glimmer-30B-GGUF) appeared the same day.

Compared To

ModelWeightsLicenseParamsRuns locally
Muse GlimmerreleasedApache 2.030B densesingle consumer GPU (18–20 GB at 4-bit)
Muse Spark 1.2not releasedunknownundisclosedno — API only
LFM2.5-2.6Breleasednot yet read2.69Bphone, under 2.5 GB
InklingreleasedApache 2.0975B MoE / 41B activeno
The interesting comparison is the first row against the second. Glimmer is
distilled from Muse Spark (1.0 / 1.1) — pre-training used **logit distillation
on Muse Spark's outputs**, with mid-training adding longer-context, agent-heavy
data and richer reasoning traces
(source). So Meta is
shipping the teacher as a closed paid API and the student as Apache 2.0,
on the same product line, five days apart. That is a deliberate split rather than
a change of position, and it is the shape Open-Weights Policy Fight has
been tracking.

Conflicting Reports

None. The licence was the one figure worth checking — an early summary did not confirm Apache 2.0, and VentureBeat, Open Source For You and Phoronix each name it explicitly (source).

Open Questions

  • Training compute and data — undisclosed. The distillation lineage is stated; the token budget is not.
  • What the "+" in "131,072+" means. No source read resolves it.
  • Is there a first-party hosted endpoint, and at what price?
  • The comparison set is Gemma4-31B and Qwen3.6-27B, both chosen by Meta. How does Glimmer measure against DeepSeek V4-Flash or Kimi K3, the open models Zuckerberg's own framing names as the competition?
  • Apache 2.0 covers the weights. No source read states whether the evaluation harness that produced the benchmark figures was released with them — the distinction Eval Harness Configuration exists to record.

Referenced by

Sources