Inference · every season
Every model, one table.
15 weeks of signed runs: how fast each model ran and on what machine, its cost per million next to OpenRouter's, what it proved it can do, and what its licence lets you serve.
126models measured here
100that no hosted API sells
139sold by OpenRouter
facebook
peculiar-ragdoll
Google
openai
huihui-ai
zai-org
microsoft
sarvamai
ibm-granite
Google
LiquidAI
mistralai
what a million tokens costs here
Every machine is metered at one posted price list: $0.036 a CPU-hour, $0.003 a GB of memory an hour, $0.36 a GPU-hour and $0.1 a kWh. A model's price here is that list applied to the machine its signed result names (its CPUs, memory and card) for the time it takes to write a million tokens at its signed speed.
why the API is usually cheaper
A hosted API runs one model for thousands of people on data-centre GPUs. One machine running one model for one person can't match that price. Running here pays off when the work must stay on your own machines, or when no API sells the model at all.
who you may serve
Read from each model's licence. Apache, MIT, Llama and Gemma terms let you serve anyone. Custom or community licences (Qwen's, MiniMax's): your own machines and your company's only. Non-commercial ones: your own machines, unpaid work.
- Every number signed by the machine that measured it — check any one in your browser
- Open data: table.json · built 2026-11-04 02:00 UTC
- Every version, in git history
- How it is measured
112 modelsclick a column to sort by it · click a model for every machine it ran on
| parameters | tokens/sec | proved | per M tokens | per M tokens | Verdict | ||||
|---|---|---|---|---|---|---|---|---|---|
Text & chat · by Google18 builds | 5.1B2B active | 3.8 GB | 291 RTX PRO 6000 Blackwell Server Edition 96 GB·Google Colab | 128k | 10 + 20 lang | $1.35metered | not sold | Anyone | only on this network |
Text & chat · by Google5 builds | 12B | 6.6 GB | ran, not timed | 256k | 10 + 14 lang | — | not sold | Anyone | only on this network |
Text & chat · by Google3 builds | 26B4B active | 15.8 GB | ran, not timed | 128k | 10 + 20 lang | — | $0.2415 providers · free tier | Anyone | measured |
Text & chat · by Google2 builds | 31B | — | ran, not timed | 128k | 10 + 20 lang | — | $0.4018 providers · free tier | Anyone | measured |
Text & chat · by Google8 builds | 8B4B active | 4.6 GB | ran, not timed | 128k | 10 + 14 lang | — | not sold | Anyone | only on this network |
Text & chat · by Qwen7 builds · 5 variants | 27B | 15 GB | 2.82× Tesla T4 15 GB· Kaggle | 8k | 9 + 14 lang | $94metered | $1.8823 providers | Anyone | API 50× cheaper |
Text & chat · by Qwen8 builds | 4B | 2.3 GB | 90NVIDIA 12 GB | 64k | 8 + 5 lang | $2.19metered | not sold | Anyone | only on this network |
Text & chat · by openai8 builds · 5 variants | 20B | 10.8 GB | 552× Tesla T4 15 GB·Kaggle | 128k | 8 + 13 lang | $2.62metered | $0.1311 providers | Anyone | API 20× cheaper |
Text & chat · by zai-org3 builds · 1 variant | ≈28B | 15.7 GB | 452× Tesla T4 15 GB·Kaggle | 128k | 8 + 8 lang | $5.62metered | $0.402 providers | Anyone | API 14× cheaper |
Text & chat · by Qwen3 builds | 600M | 0.6 GB | 134 CPUs, no GPU | 4k | 8 + 1 lang | $2.42metered | not sold | Anyone | only on this network |
Text & chat · by Qwen5 builds | 4B | 7.5 GB | 12NVIDIA 15 GB | 128k | 8 + 9 lang | $12metered | not sold | Anyone | only on this network |
Text & chat · by LiquidAI4 builds · 1 variant | 24B2B active | 12.6 GB | 3.64 CPUs, no GPU·Kaggle | 4k | 8 + 9 lang | $10metered | $0.701 provider | You & your company | API 15× cheaper |
Text & chat · by yuxinlu14 builds | 12B | — | ran, not timed | 128k | 8 + 13 lang | — | not sold | Not checked | only on this network |
Text & chat · by Qwen3 builds | 14B | 8.4 GB | ran, not timed | 64k | 8 + 7 lang | — | $0.242 providers | Anyone | measured |
 Text & chat · by moonshotai1 build | 48B3B active | 28 GB | 4.84 CPUs, no GPU | 4k | 7 + 9 lang | $11metered | not sold | Anyone | only on this network |
 Text & chat · by CohereLabs4 builds | ≈31B | 17.2 GB | 3.24 CPUs, no GPU | 4k | 7 + 9 lang | $14metered | not sold | Anyone | only on this network |
Text & chat · by mistralai2 builds | 14B | 6.7 GB | ran, not timed | 4k | 7 | — | $0.203 providers | Anyone | measured |
Text & chat · by Qwen1 build | 1.7B | — | ran, not timed | 64k | 7 + 1 lang | — | not sold | Anyone | only on this network |
Text & chat · by Qwen1 build | 8B | — | ran, not timed | 64k | 7 + 4 lang | — | not sold | Anyone | only on this network |
 Text & chat · by meta-llama16 builds · 2 variants | 3.2B | 1.9 GB | 225NVIDIA 12 GB | 128k | 6 + 8 lang | $1.14metered | $0.342 providers | Anyone | API 3.4× cheaper |
 Text & chat · by openbmb1 build | 1B | 0.6 GB | 22 Intel 2 GB | 8k | 6 + 1 lang | $3.29metered | not sold | Anyone | only on this network |
Text & chat · by Qwen3 builds | 800M | 0.5 GB | 174 CPUs, no GPU | 4k | 6 + 1 lang | $1.86metered | not sold | Anyone | only on this network |
Text & chat · by LiquidAI1 build | 1.2B | 0.7 GB | 164 CPUs, no GPU | 4k | 6 + 1 lang | $1.95metered | not sold | You & your company | only on this network |
Text & chat · by unsloth1 build | 2B | — | ran, not timed | — | 6 + 6 lang | — | not sold | Not checked | only on this network |
Text & chat · by Qwen2 builds · 1 variant | 8B | 4.7 GB | ran, not timed | 128k | 6 + 7 lang | — | $0.751 provider | Anyone | measured |
Text & chat · by microsoft10 builds | ≈4.1B | 2.3 GB | 189NVIDIA 12 GB | 4k | 5 + 3 lang | $1.36metered | not sold | Anyone | only on this network |
Text & chat · by meta-llama10 builds | 1.2B | 0.8 GB | 113 10 CPUs, no GPU | 32k | 5 + 8 lang | $0.81metered | $0.201 provider | Anyone | API 4.0× cheaper |
 Text & chat · by tencent2 builds · 1 variant | 30B3B active | 17 GB | 5.44 CPUs, no GPU | 4k | 5 + 10 lang | $8.31metered | $0.291 provider | Anyone | API 28× cheaper |
 Text & chat · by redhatai1 build | — | — | ran, not timed | 32k | 5 + 6 lang | — | not sold | Not checked | only on this network |
Text & chat · by Qwen1 build | 35B3B active | 20.5 GB | ran, not timed | 4k | 5 + 9 lang | — | $1.007 providers | Anyone | measured |
Text & chat · by Qwen3 builds | 500M | 0.5 GB | 239NVIDIA 15 GB | 32k | 4 + 5 lang | $0.56metered | not sold | Anyone | only on this network |
Text & chat · by microsoft21 builds | ≈4.1B | 2.3 GB | 154NVIDIA 12 GB | 128k | 4 + 5 lang | $1.66metered | not sold | Anyone | only on this network |
Text & chat · by google7 builds | 270M | 0.2 GB | 52Intel 2 GB | 32k | 4 + 3 lang | $0.82metered | not sold | Anyone | only on this network |
Text & chat · by google13 builds · 2 variants | 270M | 0.2 GB | 404 CPUs, no GPU | — | 4 + 3 lang | $0.76metered | not sold | Anyone | only on this network |
Text & chat · by sarvamai1 build | 24B | 13.3 GB | 16NVIDIA 15 GB | 8k | 4 + 9 lang | $9.33metered | not sold | Anyone | only on this network |
 Text & chat · by krutrim-ai-labs1 build | 12B | 7 GB | 14NVIDIA 15 GB | 64k | 4 + 9 lang | $10metered | not sold | You & your company | only on this network |
Text & chat · by ibm-granite6 builds | 1B | 1 GB | 114 CPUs, no GPU | 4k | 4 + 6 lang | $2.74metered | not sold | Anyone | only on this network |
 Text & chat · by cerebras1 build | 35B3B active | 19.8 GB | 4.84 CPUs, no GPU·Kaggle | 4k | 4 + 7 lang | $9.88metered | not sold | Anyone | only on this network |
Text & chat · by microsoft3 builds | ≈15B | 8.4 GB | 3.3Apple silicon | 32k | 4 + 9 lang | $60metered | not sold | Anyone | only on this network |
Text & chat · by Qwen2 builds | 30B3B active | 16 GB | ran, not timed | 128k | 4 + 9 lang | — | $0.602 providers | Anyone | measured |
Text & chat · by google6 builds | 1B | 0.8 GB | 144 CPUs, no GPU | 8k | 3 + 8 lang | $2.22metered | not sold | Anyone | only on this network |
Text & chat · by openbmb1 build | ≈3.6B | 2 GB | 3.7 AMD Radeon(TM) Vega 8 Mobile Graphics 1 GB | — | 3 | $47metered | not sold | Anyone | only on this network |
Text & chat · by zai-org1 build | ≈2.3B | 1.3 GB | ran, not timed | 4k | 3 + 1 lang | — | not sold | Anyone | only on this network |
Text & chat · by sarvamai2 builds | 2B | 1.4 GB | 111NVIDIA 15 GB | 8k | 2 + 8 lang | $1.22metered | not sold | Not checked | only on this network |
 Text & chat · by ai4bharat1 build | 7B | 3.9 GB | 53L4 23 GB·Google Colab | 4k | 2 + 2 lang | $4.16metered | not sold | Anyone | only on this network |
 Text & chat · by mradermacher7 builds | 270M | — | 394 CPUs, no GPU | — | 2 + 3 lang | $0.79metered | not sold | Not checked | only on this network |
 Text & chat · by north-ml12 builds | — | — | 10754 CPUs, no GPU | — | 1 | $0.03metered | not sold | Anyone | only on this network |
Text & chat · by ibm-granite1 build | 1B | 1.1 GB | 1302× Tesla T4 15 GB·Kaggle | — | 1 | $1.04metered | not sold | Anyone | only on this network |
Text & chat · by sarvamai1 build | 7B | 3.9 GB | 51L4 23 GB·Google Colab | — | 1 | $4.31metered | not sold | Anyone | only on this network |
 Image generation · by Efficient-Large-Model1 build | 600M | 15.4 GB | ran, not timed | — | 1 | — | not sold | Anyone | only on this network |
Embeddings (search) · by google14 builds | 300M | 0.2 GB | ran, not timed | — | 1 | — | not sold | Anyone | only on this network |
 Text to speech · by SWivid1 build | 150M | 0.6 GB | ran, not timed | — | 1 | — | not sold | Own machines only | only on this network |
Text to speech · by SWivid1 build | 340M | 6.3 GB | ran, not timed | — | 1 + 1 lang | — | not sold | Own machines only | only on this network |
Text to speech · by hexgrad1 build | — | 0.3 GB | ran, not timed | — | 1 + 1 lang | — | not sold | Not checked | only on this network |
Text to speech · by facebook1 build | 40M | 0.1 GB | ran, not timed | — | 1 | — | not sold | Own machines only | only on this network |
Text to speech · by facebook1 build | 40M | 0.1 GB | ran, not timed | — | 1 | — | not sold | Own machines only | only on this network |
Text to speech · by facebook1 build | 40M | 0.1 GB | ran, not timed | — | 1 | — | not sold | Own machines only | only on this network |
Text to speech · by facebook1 build | 40M | 0.1 GB | ran, not timed | — | 1 | — | not sold | Own machines only | only on this network |
Text to speech · by facebook1 build | 40M | 0.1 GB | ran, not timed | — | 1 | — | not sold | Own machines only | only on this network |
Text to speech · by facebook1 build | 40M | 0.1 GB | ran, not timed | — | 1 | — | not sold | Own machines only | only on this network |
Text to speech · by facebook1 build | 40M | 0.1 GB | ran, not timed | — | 1 | — | not sold | Own machines only | only on this network |
Text to speech · by facebook1 build | 40M | 0.1 GB | ran, not timed | — | 1 | — | not sold | Own machines only | only on this network |
Text to speech · by facebook1 build | 40M | 0.1 GB | ran, not timed | — | 1 | — | not sold | Own machines only | only on this network |
Text to speech · by facebook1 build | 40M | 0.1 GB | ran, not timed | — | 1 | — | not sold | Own machines only | only on this network |
Text to speech · by facebook1 build | 40M | 0.1 GB | ran, not timed | — | 1 | — | not sold | Own machines only | only on this network |
Text to speech · by facebook1 build | 40M | 0.1 GB | ran, not timed | — | 1 | — | not sold | Own machines only | only on this network |
Speech to text · by UsefulSensors1 build | — | 0.2 GB | ran, not timed | — | 1 + 1 lang | — | not sold | Anyone | only on this network |
Speech to text · by UsefulSensors1 build | — | 0.1 GB | ran, not timed | — | 1 + 1 lang | — | not sold | Anyone | only on this network |
Music & sound · by facebook1 build | 1.5B | 11.1 GB | ran, not timed | — | 1 | — | not sold | Own machines only | only on this network |
Music & sound · by facebook1 build | 300M | 2.2 GB | ran, not timed | — | 1 | — | not sold | Own machines only | only on this network |
Other · by rhasspy1 build | — | — | ran, not timed | — | 1 + 1 lang | — | not sold | Not checked | only on this network |
Vision (counting, locating) · by facebook1 build | 90M | 0.7 GB | ran, not timed | — | 1 | — | not sold | Anyone | only on this network |
 Image generation · by Efficient-Large-Model1 build | 1.6B | 3 GB | ran, not timed | — | 1 | — | not sold | Anyone | only on this network |
 Image generation · by stabilityai2 builds | 900M | 2.4 GB | ran, not timed | — | 1 | — | not sold | Not checked | only on this network |
Speech to text · by facebook1 build | 2.3B | 8.6 GB | ran, not timed | — | 1 + 13 lang | — | not sold | Own machines only | only on this network |
 Speaker recognition · by speechbrain1 build | 20M | 0.1 GB | ran, not timed | — | 1 | — | not sold | Anyone | only on this network |
 Video generation · by Wan-AI1 build | 1.3B | 26.9 GB | ran, not timed | — | 1 | — | not sold | Anyone | only on this network |
 Image generation · by segmind1 build | 320M | 1 GB | ran, not timed | — | 1 | — | not sold | Anyone | only on this network |
Speech to text · by openai1 build | 1.5B | 2.9 GB | ran, not timed | — | 1 + 10 lang | — | not sold | Anyone | only on this network |
Speech to text · by openai1 build | 810M | 1.5 GB | ran, not timed | — | 1 + 10 lang | — | not sold | Anyone | only on this network |
Speech to text · by openai1 build | 770M | 2.8 GB | ran, not timed | — | 1 + 4 lang | — | not sold | Anyone | only on this network |
Speech to text · by openai1 build | 240M | 0.9 GB | ran, not timed | — | 1 + 6 lang | — | not sold | Anyone | only on this network |
Speech to text · by vasista221 build | 240M | 0.9 GB | ran, not timed | — | 1 + 1 lang | — | not sold | Anyone | only on this network |
Speech to text · by vasista221 build | 240M | 0.9 GB | ran, not timed | — | 1 | — | not sold | Anyone | only on this network |
Speech to text · by vasista221 build | 240M | 0.9 GB | ran, not timed | — | 1 + 1 lang | — | not sold | Anyone | only on this network |
Speech to text · by vasista221 build | 240M | 0.9 GB | ran, not timed | — | 1 + 1 lang | — | not sold | Anyone | only on this network |
 Text to speech · by coqui1 build | 470M | 1.9 GB | ran, not timed | — | 1 | — | not sold | You & your company | only on this network |
Text & chat · by tencent1 build | 1.8B | 1 GB | 49L4 23 GB·Google Colab | — | — | $4.46metered | not sold | Not checked | only on this network |
Text & chat · by Qwen3 builds | 27B | 15 GB | 9.4machine not reported | — | — | — | $2.707 providers | Anyone | measured |
Text & chat · by mistralai1 build | 3B | 2 GB | 2.7machine not reported | — | — | — | $0.103 providers | Anyone | measured |
Text & chat · by Qwen3 builds | 9B | 5.3 GB | 2.4machine not reported | — | — | — | $0.256 providers | Anyone | measured |
Text & chat · by ibm-granite3 builds | 8B | 4.6 GB | 1.9machine not reported | — | — | — | $0.252 providers | Anyone | measured |
 Text & chat · by meta-models5 builds · 2 variants | 30B | 15.6 GB | 1.3machine not reported | — | — | — | $1.201 provider | Anyone | measured |
Image generation · by THUDM1 build | 5B | 20 GB | ran, not timed | — | — | — | not sold | You & your company | only on this network |
 Image generation · by THUDM2 builds | 2B | 12.8 GB | ran, not timed | — | — | — | not sold | Anyone | only on this network |
 Vision (counting, locating) · by depth-anything1 build | 100M | 0.4 GB | ran, not timed | — | — | — | not sold | Own machines only | only on this network |
Vision (counting, locating) · by depth-anything1 build | 30M | 0.1 GB | ran, not timed | — | — | — | not sold | Anyone | only on this network |
 Image generation · by tencent1 build | 13B | 39 GB | ran, not timed | — | — | — | not sold | Not checked | only on this network |
 Image generation · by Lightricks1 build | 2B | 23.6 GB | ran, not timed | — | — | — | not sold | You & your company | only on this network |
Other · by lightricks1 build | 13B | — | ran, not timed | — | — | — | not sold | Not checked | only on this network |
Translation · by google1 build | 3B | 11 GB | ran, not timed | — | 0 + 10 lang | — | not sold | Anyone | only on this network |
Translation · by facebook1 build | 1.3B | 5.1 GB | ran, not timed | — | 0 + 12 lang | — | not sold | Own machines only | only on this network |
Translation · by facebook1 build | 600M | 2.3 GB | ran, not timed | — | 0 + 12 lang | — | not sold | Own machines only | only on this network |
Other · by prunaai1 build | — | — | ran, not timed | — | — | — | not sold | Not checked | only on this network |
Other · by harahan1 build | 2B | — | ran, not timed | — | — | — | not sold | Not checked | only on this network |
Other · by lightx2v1 build | — | — | ran, not timed | — | — | — | not sold | Not checked | only on this network |
Translation · by sarvamai1 build | 4B | 8 GB | ran, not timed | — | 0 + 13 lang | — | $0.102 providers | Anyone | measured |
Other · by languaman1 build | — | — | ran, not timed | — | — | — | not sold | Not checked | only on this network |
Other · by aemotionstudio1 build | — | — | ran, not timed | — | — | — | not sold | Not checked | only on this network |
 Other · by optimum-intel-internal-testing1 build | — | — | ran, not timed | — | — | — | not sold | Not checked | only on this network |
Other · by alibaba-pai1 build | 1.3B | — | ran, not timed | — | — | — | not sold | Not checked | only on this network |
Other · by wan-ai1 build | 1.3B | — | ran, not timed | — | — | — | not sold | Not checked | only on this network |
Updated Wed, 04 Nov 2026 02:00:16 GMT. The data: table.json (models, and every build behind them), market.json (OpenRouter prices and licences, with when they were read). Back to the seasons →