is now an ernd partner: 10% back on inference 
In this guide
Summary
- Independent providers served 315 trillion open-weight tokens on OpenRouter from October 6 to November 4, 2026. That’s up 55% from 203 trillion in the 30 days before.
- Together served the most, with 13.0% of that total. The top five providers together served 48%.
- DeepSeek V4.1 Flash was the most-served open-weight model. Its two versions made up 39.5% of all open-weight tokens on OpenRouter over the 30 days.
Disclosure: ernd partners with Wafer, one of the providers ranked here. Its numbers come from the same sums as everyone else’s.
Cite this: ernd, “State of open-weight inference, November 2026,” ernd.ai/reports/open-weight-inference-november-2026, using OpenRouter public data for October 6 to November 4, 2026.
Your AI bill, with ernd
What could your AI bill earn?
Enter an amount from $0 to $100,000,000.
back a year at 4%
For US companies. No annual fee. Illustration based on AI vendor spend only. Rewards are points redeemable as cash back.
Who served the most tokens?
Together led the 30 days. The last week looked different. Wafer was first with 13.9%, then DeepInfra and Together. Wafer’s 30-day rank was fourth, about 0.8 trillion tokens behind DeepInfra.
| Rank | Provider | Last 30 days | Share | Last 7 days | Share | Rank, last 7 days |
|---|---|---|---|---|---|---|
| 1 | Together | 40.8T | 13.0% | 9.5T | 10.6% | 3 |
| 2 | inference.net | 30.0T | 9.5% | 8.1T | 9.1% | 4 |
| 3 | DeepInfra | 28.3T | 9.0% | 10.6T | 11.8% | 2 |
| 4 | Wafer | 27.6T | 8.7% | 12.4T | 13.9% | 1 |
| 5 | Relace | 25.9T | 8.2% | 6.2T | 6.9% | 5 |
| 6 | DigitalOcean | 18.9T | 6.0% | 5.1T | 5.7% | 6 |
| 7 | CoreWeave | 12.7T | 4.0% | 3.6T | 4.0% | 8 |
| 8 | Morph | 11.7T | 3.7% | 4.5T | 5.0% | 7 |
| 9 | Fireworks | 11.5T | 3.7% | 3.1T | 3.5% | 9 |
| 10 | Novita | 9.9T | 3.1% | 1.7T | 1.9% | 14 |
| 11 | OpenInference | 9.6T | 3.0% | 3.1T | 3.4% | 10 |
| 12 | Parasail | 8.2T | 2.6% | 1.7T | 2.0% | 12 |
T is trillion tokens. The 30-day window is October 6 to November 4, 2026 and the 7-day window is October 29 to November 4, both in UTC. Shares are of all open-weight tokens served by independent providers: 315.0 trillion over 30 days and 89.5 trillion over 7 days.
Labs serving their own models handled another 247.4 trillion open-weight tokens in the 30 days. So independents carried 56% of open-weight traffic on OpenRouter, up from 55% the month before.
How fast did providers grow?
The independent total grew 55%. Half of the top 12 grew faster than that, and two shrank.
| Provider | Sep 6 to Oct 5 | Oct 6 to Nov 4 | Change |
|---|---|---|---|
| Together | 28.5T | 40.8T | +43% |
| inference.net | 4.0T | 30.0T | +642% |
| DeepInfra | 18.0T | 28.3T | +57% |
| Wafer | 9.7T | 27.6T | +183% |
| Relace | 27.9T | 25.9T | -7% |
| DigitalOcean | 1.9T | 18.9T | +893% |
| CoreWeave | 9.4T | 12.7T | +35% |
| Morph | 1.8T | 11.7T | +565% |
| Fireworks | 5.2T | 11.5T | +121% |
| Novita | 16.7T | 9.9T | -41% |
| OpenInference | 7.0T | 9.6T | +38% |
| Parasail | 8.0T | 8.2T | +3% |
OpenRouter’s data for inference.net starts on September 12, so its earlier month is short by six days.
DigitalOcean went from under 2 trillion to almost 19. Relace and Novita were second and fourth in September and lost ground in October.
Which open-weight models were served most?
These totals cover every provider, labs included. Share is of the 562.5 trillion open-weight tokens on OpenRouter over the 30 days.
| Rank | Model | Last 30 days | Share | Providers | Served by its own lab |
|---|---|---|---|---|---|
| 1 | DeepSeek V4.1 Flash 1017 | 113.7T | 20.2% | 39 | 31.6% |
| 2 | DeepSeek V4.1 Flash | 108.5T | 19.3% | 37 | 13.3% |
| 3 | Z.ai GLM 5.3 Flash | 58.5T | 10.4% | 39 | 5.7% |
| 4 | Tencent Hy4 preview | 35.9T | 6.4% | 7 | 96.3% |
| 5 | StepFun Step 5 Preview | 33.2T | 5.9% | 3 | 99.9% |
| 6 | Xiaomi MiMo-V2.6-Flash | 29.8T | 5.3% | 10 | 89.6% |
| 7 | MiniMax M3.1 Flash | 23.2T | 4.1% | 12 | 95.3% |
| 8 | NVIDIA Nemotron 3 Ultra | 21.2T | 3.8% | 3 | 99.6% |
| 9 | Z.ai GLM 5.3 | 18.8T | 3.4% | 37 | 9.1% |
| 10 | DeepSeek V4 Flash 0731 | 18.1T | 3.2% | 29 | 0.0% |
Names as listed in OpenRouter’s catalog. Providers counts every provider with tokens on that model in the window.
The models fall into two groups. On the DeepSeek models, independents serve most of the traffic. Tencent, StepFun, Xiaomi, MiniMax and NVIDIA serve nearly all of their own.
DeepSeek V4.1 Flash 1017 first shows up in the data on October 17, and it still finished first. In the last week it took 38.5% of open-weight tokens by itself.
Who had the best price and latency?
For the models with the most independent traffic, these are the cheapest listing and the fastest providers on OpenRouter.
| Model | Cheapest listing, input / output per 1M tokens | Fastest first token | Highest throughput |
|---|---|---|---|
| DeepSeek V4.1 Flash 1017 | Sail Research, $0.10 / $0.15 (fp4) | Together, 0.25 s | Together, 248 tok/s |
| DeepSeek V4.1 Flash | Morph, $0.02 / $0.40 (fp8) | Together, 0.23 s | Fireworks, 241 tok/s |
| Z.ai GLM 5.3 Flash | Relace, $0.03 / $0.09 (fp4) | Modal, 0.40 s | Fireworks, 128 tok/s |
| MiniMax M3.1 Flash | GMICloud, $0.11 / $0.45 (fp8) | inference.net, 0.52 s | DigitalOcean, 190 tok/s |
| Z.ai GLM 5.3 | Wafer, $0.07 / $1.40 (fp4) | Modal, 0.32 s | Baseten, 173 tok/s |
| DeepSeek V4 Flash 0731 | DeepInfra, $0.06 / $0.18 (fp8) | Baseten, 0.29 s | Cohere, 158 tok/s |
Prices are OpenRouter list prices on November 5, 2026, before caching discounts. Cheapest means the lowest blend of three parts input to one part output, on standard-tier endpoints. Speeds are OpenRouter’s daily figures averaged over October 29 to November 4, for endpoints with at least five days of data.
Cheap listings often run at lower precision, so the quantization is shown. Check output quality on your own prompts before you switch.
How this report was made
OpenRouter doesn’t publish provider rankings. Each provider’s page on OpenRouter has a public chart of daily tokens by model, and every number here is a sum of those charts. We pulled all 121 providers on November 6, 2026 at about 02:00 UTC.
A model counts as open-weight when OpenRouter’s catalog links it to public weights on Hugging Face. Independent means a provider serving a model it didn’t make, so a lab’s tokens for its own models are left out of the provider totals.
Each provider’s chart names its top models and groups the rest as “Others.” Those tokens can’t be classified, so they’re left out. They were 22.1 trillion across all providers over the 30 days. Adding them back moves DeepInfra above inference.net and leaves the rest of the top four as it is.
These are OpenRouter tokens only. A provider’s direct customers don’t show up here, so this measures share of OpenRouter traffic, not share of the market.
This report is updated monthly.
The ernd card
Earn on what’s left.
Once you’ve worked through your AI costs, make the remaining bill work for you. Earn 4% back at AI vendors like OpenAI and Anthropic.
Request your cardThe ernd card is a credit card for US companies, with partners across the AI stack built in.
Common questions.
Which inference provider serves the most open-weight tokens on OpenRouter?
Over the 30 days from October 6 to November 4, 2026, Together served the most, with 40.8 trillion tokens and 13.0% of the independent total. In the last week of that window, October 29 to November 4, Wafer served the most, with 12.4 trillion tokens and 13.9%.
Does OpenRouter publish provider rankings?
No. OpenRouter publishes daily token counts per model on each provider's page. The rankings here are sums of those counts.
What counts as an independent provider?
Any provider serving an open-weight model it didn't make. Tokens a lab serves for its own models, such as DeepSeek serving DeepSeek models, are left out of the provider totals.
Sources & methodology
Checked . Providers change prices and terms; the linked official page is the final word. This guide is published by ernd.
ernd is run by Provenance Research Inc.