Mistral limits Timestral weights to companies with no more than $20M in monthly revenue
Mistral released Timestral, four CPU-capable forecasting models, with weights licensed only to firms at or under $20M in monthly revenue. The license text, as DEV Community quotes it, tests the preceding month, so eligibility can change monthly.
The Engineer · Build desk
Three figures: the license bars companies whose global consolidated monthly revenue exceeds $20 million for the preceding month; a flat $20 million a month is $240 million a year; the 5M-parameter model's license carries the same $20 million cap.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Monthly revenue cap, tested on the preceding month | $20 million | 9 | |
| Cap held flat for a year | $240 million | 17 | |
| Small model, same cap | 5 M parameters | 12 |
What happened
- Mistral reports 10 ms for the 5M model and 140 ms for the 209M model on a server CPU, and says the family needs no GPU at inference.
- The inference code is Apache 2.0, while the weights sit on the Hugging Face Hub under what Mistral calls a Modified MIT license.
- The revenue restriction extends to any derivatives, modifications or combined works built on the weights.
- Every benchmark number comes from Mistral, and the model card lists a tech report as "coming soon", DEV Community notes.
Why it matters
- What it limits constraint A large company gains nothing by choosing a smaller checkpoint. The 5M model's license carries the same $20M cap.
- Who's exposed exposure A company with a seasonal peak can exceed $20M in one month while averaging less over the year, and the license text tests the preceding month.
- Who has to decide decision A company above the line has two routes: a commercial license at Mistral's sole discretion, or Mistral's hosted services. DEV Community reports no API price is listed for the Public Preview.
Next story · Product · Oct 7
Alibaba's $25bn a year of e-commerce cash backs Joe Tsai's open-source pitch to EuropeConfirmed4 publishers
The gate sits on the company. DEV Community quotes the license file as saying a user is "not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company" "exceeds $20 million (or its equivalent in another currency) for the preceding month." [9] The same post says the test reaches an employer's revenue too [9]. Held flat, $20 million a month is $240 million a year [17]. A team near that line would need a revenue forecast to know whether it may run the forecasting model.
The setup behind the CPU figure is narrow. Mistral timed one day ahead from three weeks of 15-minute history on a "standard server CPU with 8 threads" [2]. That history is 2,016 steps [14], about 6% of the 32,768 past steps the 209M model can read [15]. We'd expect a pipeline that feeds the model its full context to run slower than 140 ms.
The benchmark lead is small. On GIFT-Eval the 209M model scores a MASE of 0.658 against 0.661 for t0-gamma, a gap of 0.003 [16]. On fev-bench it posts 50.39 against 49.42 [4]. Mistral says each result comes from a single checkpoint with no ensembling, and that most training data is synthetic, so "nothing we backtest was seen during pre-training" [19]. For a 0.003 gap to hold on your data, your series would have to resemble the benchmark mix, and your error metric would have to track MASE.
Teams under the cap have a cheaper option in the smaller checkpoints. Mistral says the 123M model comes within 0.2% of the 209M score at three quarters of its latency [3]. It also says its 28M model beats Chronos-2, about four times larger, on GIFT-Eval and TIME [7]. On GIFT-Eval MASE the 28M scores 0.678 against 0.698 for Chronos-2 [6].
What to watch
- Mistral's tech report, listed as "coming soon", which would let outsiders check the method and the synthetic-data claim.
- Whether Mistral publishes an API price when Timestral leaves Public Preview, which sets the cost for companies above the cap.
- Independent reruns of the 209M model on non-benchmark series, where a 0.003 MASE lead has to survive.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence40
- Adoption
- Not scored
- Hype gap+20
- Incentives65
- Confidence60
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Timestral is a family of four models, at 5M, 28M, 123M and 209M parameters, that share the same architecture and training recipe.
- [2]
Mistral says Timestral does not need a GPU at inference. On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model.
- [3]
The 28M model is five times faster than the 209M model, and the 123M model comes within 0.2% of the 209M score at three quarters of its latency.
- [4]
On the model card's comparison table, Timestral-209M scores a MASE of 0.658 on GIFT-Eval against 0.661 for t0-gamma, 0.667 for TimesFM-3 and 0.698 for Chronos-2 (lower is better). On fev-bench's SQL skill score it posts 50.39 against 49.42 for t0-gamma and 47.28 for Chronos-2 (higher is better).
- [5]
The 209M model reads up to 32,768 past steps and forecasts up to 1,024 steps per forward pass.
- [6]
The 28M model scores 0.678 on GIFT-Eval MASE and 0.638 on TIME MASE, against 0.698 and 0.662 for Chronos-2.
- [7]
Mistral says its 28M model beats Chronos-2, about four times larger, on GIFT-Eval and TIME.
- [8]
The inference code is Apache 2.0. The weights are distributed separately on the Hugging Face Hub under what Mistral calls a Modified MIT license.
- [9]
DEV Community quotes the license file as saying a user is "not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company" (or the user's employer) "exceeds $20 million (or its equivalent in another currency) for the preceding month."
- [10]
The revenue restriction covers the model and any derivatives, modifications, or combined works.
- [11]
Companies above the revenue line can ask Mistral for a commercial license, which it may grant at its sole discretion, or use Mistral's hosted services.
- [12]
The 5M model's license carries the same $20 million cap, so a smaller checkpoint is not a way around it.
- [13]
Mistral's docs list Timestral 1.0 as a "Public Preview" and name no API price.
- [14]
Three weeks of 15-minute history is 2,016 steps.
- [15]
The benchmarked history of 2,016 steps is about 6% of the 32,768-step maximum context.
- [16]
The 209M model's GIFT-Eval MASE lead over t0-gamma is 0.003.
- [17]
A flat $20 million a month is $240 million a year.
- [18]
Every benchmark number comes from Mistral, and the model card says a tech report is "coming soon".
- [19]
Mistral says the results come from a single checkpoint with no ensembling and that most training data is synthetic, so "nothing we backtest was seen during pre-training".
Sources
2 independent publishers whose own reporting we read for this story.
- dev.toMistral's Timestral Runs on a CPU. Its 'Modified MIT' License Stops at $20M a Month
1 article · October 15, 2026
- mistral.aiIntroducing Timestral
1 article · October 15, 2026