Model

Comma v0.1-2T

OPEN

EleutherAI · Comma family · released Jun 6, 2025

7B model trained for 2T tokens (1.93T main phase + 75.5B cool-down) entirely on the openly-licensed/public-domain 'Common Pile v0.1' dataset (8TB), demonstrating that legally clean training data can rival models trained on unlicensed web text; a 1T-token sibling checkpoint (comma-v0.1-1t) was released alongside it.

ReasoningCodingVisionFunction callingTool useAgentic

-71.4

Elo · rank #410

Parameters

Active params

7B (dense)

Context

4K tokens

Architecture

Dense decoder-only transformer, Llama 3 architecture, 7B, trained with 4096-token sequences

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

ARC-ChallengeReasoning45.8%#134

best: Llama 3.1 405B · 96.9%

ARC-EasyReasoning71.8%#44

best: Phi-3-medium (14B) · 97.7%

HellaSwagReasoning64.4%#136

best: Claude 3 Opus · 95.4%

HumanEvalCoding44.2%#133

best: Claude Opus 4.5 · 99.4%

MBPPCoding41.5%#81

best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%

MMLUKnowledge49.8%#232

best: OpenAI o3 · 92.9%

OpenBookQAReasoning46.2%#34

best: Claude 1 · 90.8%

PIQAReasoning72.5%#73

best: GPT-4o mini · 93.1%

Social IQaReasoning52.3%#19

best: Apple DCLM-Baseline 7B · 82.9%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Comma family

Elo progression across releases

Comma v0.1-2TJun 2025-71.4

Resources

HFcommon-pile/comma-v0.1-2t DOCShttps://blog.eleuther.ai/common-pile/

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