Models
Size · speed · intelligence.
The field on one axis. Cyan marks what Conifer runs locally.
Intelligence
Artificial Analysis Intelligence Index · higher is better · Artificial Analysis · 60 models
- 59.9Claude Fable 5: 59.9 index
- 55.7Claude Opus 4.8: 55.7 index
- 54.8GPT-5.5: 54.8 index
- 53.5Claude Opus 4.7: 53.5 index
- 44.3DeepSeek V4 Pro: 44.3 index
- 44.2Kimi K2.6: 44.2 index
- 40.8Claude Opus 4.5: 40.8 index
- 40GPT-5.4 mini: 40 index
- 39.6Gemini 3 Pro: 39.6 index
- 38.2GPT-5.4 nano: 38.2 index
- 36.9GPT-5.1: 36.9 index
- 36.4Claude Sonnet 4.5: 36.4 index
- 34.7GPT-5: 34.7 index
- 33.3Grok 4: 33.3 index
- 32.7Kimi K2 Thinking: 32.7 index
- 32DeepSeek V3.2: 32 index
- 31.7Qwen3 Max: 31.7 index
- 30.6Grok 4.1 Fast: 30.6 index
- 28.7GLM-4.6: 28.7 index
- 27.4Grok 4 Fast: 27.4 index
- 25.8Gemini 2.5 Pro: 25.8 index
- 23.8gpt-oss 120B: 23.8 index
- 21.8Nova 2.0 Pro: 21.8 index
- 20.9Nova 2.0 Omni: 20.9 index
- 20.1DeepSeek R1 0528: 20.1 index
- 20.1Gemini 2.5 Flash: 20.1 index
- 20.1Qwen3.5 4B: 20.1 index, runs locally on Conifer
- 19.6Qwen3 235B: 19.6 index
- 19.2Devstral 2: 19.2 index
- 18.2Nova 2.0 Lite: 18.2 index
- 17.9Qwen3 VL 32B: 17.9 index
- 17.7MiniMax M1: 17.7 index
- 17o1-preview: 17 index
- 16.5GLM-4.5 Air: 16.5 index
- 15.9Mistral Large 3: 15.9 index
- 15.4DeepSeek V3 0324: 15.4 index
- 14.9gpt-oss 20B: 14.9 index
- 14.8GPT-4.1 mini: 14.8 index
- 14.4Qwen3 30B A3B: 14.4 index, runs locally on Conifer
- 14.3Llama 4 Maverick: 14.3 index
- 14.2Nemotron 3 Nano: 14.2 index
- 12.5Mistral Medium 3: 12.5 index
- 11.5Qwen3 32B: 11.5 index, runs locally on Conifer
- 11R1 Distill 32B: 11 index, runs locally on Conifer
- 10.4Qwen3 14B: 10.4 index, runs locally on Conifer
- 9.8R1 Distill 14B: 9.8 index, runs locally on Conifer
- 9.4Llama 3.3 70B: 9.4 index
- 8.4Qwen3 4B: 8.4 index, runs locally on Conifer
- 8.3Qwen3 8B: 8.3 index, runs locally on Conifer
- 7.6Llama 3.1 8B: 7.6 index, runs locally on Conifer
- 7.5Qwen2.5 32B: 7.5 index, runs locally on Conifer
- 7.4Gemma 3 27B: 7.4 index
- 7.1Qwen2.5 Coder 32B: 7.1 index, runs locally on Conifer
- 6.4R1 Distill Llama 8B: 6.4 index, runs locally on Conifer
- 5.5Gemma 3 12B: 5.5 index, runs locally on Conifer
- 4.9Phi-4: 4.9 index, runs locally on Conifer
- 4.5Qwen2.5 Coder 7B: 4.5 index, runs locally on Conifer
- 4.2Llama 3.2 3B: 4.2 index, runs locally on Conifer
- 2.6Qwen3 1.7B: 2.6 index, runs locally on Conifer
- 1.1Gemma 3 4B: 1.1 index, runs locally on Conifer
how to read this
The whole Artificial Analysis field on one axis, and lit in cyan: the open models Conifer runs locally.
On the Intelligence composite the cloud giants tower; flip to GPQA, AIME or coding to watch the laptop-sized models punch far above their weight.
methodology & sources
Published figures, cross-checked against Artificial Analysis and each model's own technical report (snapshot Jul 2026— refreshed automatically from the live AA leaderboard).
Intelligence is the live Artificial Analysis Intelligence Index (a multi-eval composite on AA's current scale); capability scores are reasoning / thinking mode, no external tools, and benchmark versions vary by lab.
“Runs locally” marks the open weights in Conifer's catalogue that fit a personal machine; larger open models (Kimi, DeepSeek V3/V4, gpt-oss 120B, Qwen3 235B) sit in the field.
How far behind is one consumer GPU? Six to twelve months.
- AA Intelligence Index6.3 mo
- MMLU-Pro7.3 mo
- GPQA-Diamond7.4 mo
- LM Arena Elo12.4 mo
source & method
Epoch AI, “Frontier AI performance becomes accessible on consumer hardware within a year” (Somala & Emberson, Aug 2025, CC-BY): the best open model that fits a single consumer GPU matches frontier scores from 6–12 months earlier — and it is gaining (+125 vs +80 Elo per year on LM Arena).
Their bar for “fits”: full weights, 4-bit quantized, in one card's VRAM — models to ~28B on an RTX 4090, ~40B on an RTX 5090. Epoch's caveat: small open models are likelier to be benchmark-tuned, so real-world lag may run somewhat longer.
| model | size | tok/s | intelligence |
|---|---|---|---|
| Llama | |||
| Llama 3.2 1B | 1.2B | 299 | 49.3 |
| Llama 3.2 3B | 3.2B | 118 | 63.4 |
| Llama 3.1 8B | 8.0B | 49 | 69.4 |
| Hermes 3 · Llama 3.1 8B | 8.0B | ~49 | 64.8 |
| Qwen 3 | |||
| Qwen 3 0.6B | 0.6B | ~290 | 52.8 |
| Qwen 3 1.7B | 1.7B | ~165 | 62.6 |
| Qwen 3 4B | 4.0B | 90 | 73.0 |
| Qwen 3 4B Instruct 2507 | 4.0B | ~90 | · |
| Qwen 3 8B | 8.2B | 46 | 76.9 |
| Qwen 3 14B | 14.8B | ~25 | 81.1 |
| Qwen 3 32B | 32.8B | ~12 | 83.6 |
| Qwen 3 30B A3B 2507 | 30.5B · 3B active | · | 87.1 |
| Qwen 3 Coder 30B A3B | 30.5B · 3B active | · | · |
| Qwen 2.5 | |||
| Qwen 2.5 0.5B | 0.5B | 288 | 47.5 |
| Qwen 2.5 1.5B | 1.5B | 166 | 60.9 |
| Qwen 2.5 3B | 3.1B | 89 | 65.6 |
| Qwen 2.5 7B | 7.6B | 53 | 74.2 |
| Qwen 2.5 14B | 14.7B | ~27 | 79.7 |
| Qwen 2.5 Coder 1.5B | 1.5B | ~166 | 53.6 |
| Qwen 2.5 Coder 7B | 7.6B | ~53 | 68.0 |
| Gemma | |||
| Gemma 2 2B | 2.6B | 138 | 51.3 |
| Gemma 2 9B | 9.2B | ~41 | 71.3 |
| Gemma 3 4B | 4.3B | 66 | 59.6 |
| Gemma 4 12B | 11.9B | 15 | 74.5 |
| DeepSeek | |||
| R1 Distill Qwen 1.5B | 1.8B | ~166 | · |
| R1 Distill Qwen 7B | 7.6B | ~53 | · |
| R1 Distill Llama 8B | 8.0B | ~49 | · |
| R1 Distill Qwen 14B | 14.8B | ~27 | · |
| R1 Distill Qwen 32B | 32.8B | ~12 | · |
| DeepSeek V2 Lite | 15.7B | · | 55.7 |
| Phi | |||
| Phi 3.5 Mini | 3.8B | ~78 | 69.0 |
measurement notes
tok/s: decode, Apple M3 Max, Q4_K_M, 512-token prompt; ~ projected from a measured sibling. intelligence: MMLU 5-shot, as published. · marks not measured / not published.