$ ls ./models
The open models worth running
Quality on four axes — reasoning, coding, agentic, long-context — every number cited to its source. Filter by your rig to see what fits, how fast it decodes, and where each model stands on the frontier.
~/mindpool/models --explore
rig
| model | Reasoning | Coding | Agentic | Long-context | fits my rig |
|---|---|---|---|---|---|
Gemma 4 E2B 2B · 3 GB · Apache-2.0 · source ↗ | 60 | 44 | 24.5 | 19.1 | ~120 tok/s |
Gemma 4 E4B 4B · 5 GB · Apache-2.0 · source ↗ | 69.4 | 52 | 42.2 | 25.4 | ~60 tok/s |
Gemma 4 12B 12B · 7 GB · Apache-2.0 · source ↗ | 77.2 | 72 | 69 | 43.4 | ~20 tok/s |
Gemma 4 26B-A4B 26B (4B active) · 15 GB · Apache-2.0 · source ↗ | 82.6 | 77.1 | 68.2 | 44.1 | ~60 tok/s |
Gemma 4 31B 31B · 18 GB · Apache-2.0 · source ↗ | 85.2 | 80 | 76.9 | 66.4 | — too big |
Qwen3 8B 8B · 6 GB · Apache-2.0 · source ↗ | 87.5 | 57.5 | 68.1 | — | ~30 tok/s |
Qwen3 14B 14B · 9 GB · Apache-2.0 · source ↗ | 88.6 | — | — | — | ~17 tok/s |
Qwen3 30B-A3B 30B (3B active) · 18 GB · Apache-2.0 · source ↗ | 89.5 | — | — | — | — too big |
Llama 8B 8B · 6 GB · Llama Community · source ↗ | 48.3 | 72.6 | 76.1 | — | ~30 tok/s |
Llama 70B 70B · 42 GB · Llama Community · source ↗ | 68.9 | 88.4 | — | — | — too big |
Llama 405B 405B · 230 GB · Llama Community · source ↗ | 73.3 | — | — | — | — too big |
DeepSeek-V2-Lite 16B (2.4B active) · 10 GB · DeepSeek · source ↗ | — | 57.3 | — | — | ~100 tok/s |
OLMo 2 13B 13B · 8 GB · Apache-2.0 · source ↗ | 68.5 | — | — | — | ~18 tok/s |
tok/s ≈ bandwidth ÷ (active-params × bytes/param) — analytic decode ceiling; real-world is lower (attention, KV, overhead) and ignores prefill.
Scores are curated from primary sources (each number links to its source ↗); axes a model doesn't publish show "—".
Reasoning mixes MMLU-Redux %, MMLU-Pro %, MMLU-Pro CoT %, MMLU-Pro CoT 5-shot %, MMLU 5-shot % — different benchmarks, not directly rankable. Hover a score for its exact metric.
models as of 2026-06 · benchmarks as of 2026-06 — curated, not exhaustive; re-verify at source.