Model

Muse Glimmer 8B

OPEN

Meta · Muse Glimmer family · released Sep 15, 2026

Meta's September 15, 2026 open-weights release — the 8.2B consumer-GPU tier of the Muse Glimmer line that debuted at 30B in August. The point of the release is footprint: 5.8GB at Q4_K_M (16.5GB at BF16) with ~95 tok/s on an RTX 4090, i.e. a genuinely local multimodal model. The capability story is proportionate — most scores land at 70-80% of the 30B sibling, with agentic rows (MCP Atlas, OSWorld-Verified, Terminal-Bench) falling off hardest, as expected at this size. Apache 2.0, weights on HuggingFace, no first-party API.

ReasoningCodingVisionFunction callingTool useAgentic

1566.8

Elo · rank #133

Parameters

8.2B

Active params

8.2B (dense)

Context

131K tokens

Architecture

8.2B dense causal transformer with a distilled ViT-S/16 perception encoder (~0.4B); 131K-token context, the Glimmer line's consumer-GPU tier

License

Apache 2.0

Languages

—

API price (in/out)

No hosted API

Modalities

text · vision

Benchmark results

Bar shows position within the tracked field; marker = field best

AIMEMath82.1%#73

best: GPT-5.2 · 100.0%

CharXivVision63.5%#38

best: Qwen3.8-Flash-Next · 90.6%

GDPval-AAAgents812#53

best: Claude Fable 5 · 1932

GPQA DiamondReasoning68.7%#149

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning12.4%#97

best: Claude Opus 5 · 64.7%

IFBenchReasoning64.8%#38

best: Inkling-Medium · 85.3%

MCP AtlasAgents61.2%#10

best: Kimi K3.1 · 85.6%

MMMU-ProVision58.3%#54

best: Claude Opus 4.7 · 85.5%

OSWorld-VerifiedAgents52.7%#30

best: Claude Fable 5 · 85.0%

ScreenSpot-ProVision61.8%#7

best: GPT-6 Astra · 92.7%

SWE-bench ProCoding38.4%#67

best: Claude Fable 5.1 · 81.2%

SWE-bench VerifiedCoding61.2%#87

best: Claude Opus 5 · 96.0%

Terminal-Bench 2.0Coding36.9%#81

best: Gemini 3.8 Flash · 89.4%

Run it locally

VRAM @ Q4

5.8 GB

VRAM @ FP16

16.5 GB

Fits on (Q4)

RTX 3060 12GBRTX 4070 Ti 16GBRTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

~95 tok/s on RTX 4090 (Q4, llama.cpp)

Quantizations

GGUF Q4 · Q4_K_M · BF16

Fine-tune it

Permissive

QLoRA7.2 GB1× RTX 3060 12GB

LoRA19.1 GB1× RTX 3090 24GB

Full fine-tune133.2 GB1× H200 141GB

QLoRA SFT on ~10k samples ≈ $4.80 (1× RTX 3060 12GB)

Plan a fine-tune →

Muse Glimmer family

Elo progression across releases

Muse Glimmer 30BAug 20262032.9 Muse Glimmer 8BSep 20261566.8

Resources

HFmeta-llama/Muse-Glimmer-8B DOCShttps://developer.meta.com/ai/models/muse-glimmer

API price weights · each benchmark row carries its own source badge (see methodology)