Olmo 3 7B Instruct SFT
Ai2 · released 2025-12-01 · Apache-2.0 license
Ai2's Olmo 3 7B Instruct SFT is the supervised-finetuned checkpoint of a fully-open 7B: vendor IFEval 81.7, GPQA 30.0, LiveCodeBench v3 20.0, AIME 6.7. Later DPO/RLVR variants score notably higher.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 7.3B total |
| Architecture | olmo3 |
| Context window | 66K tokens |
| Knowledge cutoff | 2024-12-01 |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 30% |
|---|---|
| SWE-bench Verified | — |
| AIME | 6.7% |
| MMLU-Pro | — |
| BFCL v3 (tool use) | 48.9% |
| Composite score | 3.76 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 5.7 GB | 4.2 GB | — | 66K |
Strengths & weaknesses
Strengths: Fully open (weights + data + training code); Decent instruction following (IFEval 81.7) and strong safety; Apache-2.0
Weaknesses: Weak coding (LiveCodeBench v3 20.0) and math (AIME 6.7); This is the SFT-only checkpoint; DPO/RLVR variants score higher