Model
Model explorer

Snowflake Arctic

OPEN
Snowflake · Arctic family · released Apr 24, 2024

Apache-2.0 480B-total/17B-active dense+MoE hybrid (128 experts) tuned for enterprise SQL/code/instruction-following, trained for under $2M.

ReasoningCodingVisionFunction callingTool useAgentic
485.8
Elo · rank #308
Parameters
480B
Active params
17B (MoE)
Context
4K tokens
Architecture
Dense-MoE Hybrid (10B dense transformer + 128x3.66B residual MoE experts, top-2 gating)
License
Apache-2.0
Languages
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
GSM8KMath74.2%#93
best: Llama 3.1 405B · 96.8%
HumanEvalCoding64.3%#93
best: Claude Opus 4.5 · 99.4%
IFEvalReasoning52.4%#131
best: Gemma 4 26B A4B · 98.5%
MMLUKnowledge67.3%#167
best: OpenAI o3 · 92.9%
Run it locally
VRAM @ Q4
291 GB
VRAM @ FP16
960 GB
Fits on (Q4)
M3 Ultra 512GB
480B total parameters (Q4 ~291GB) require a multi-GPU server; not runnable on a single consumer GPU, and mainline llama.cpp needs a community fork for Arctic's architecture.
Quantizations
GGUF Q4
Arctic family
Elo progression across releases
API price weights · each benchmark row carries its own source badge (see methodology)