That makes the release more than a parameter-count headline, while leaving the most consequential question—what users can actually run and under what terms—unresolved.
Implication: The split between API preview and later weights means access, reproducibility, and deployment options may change only if the promised release arrives with usable terms.
Thesis
The preview signals a staged return to frontier competition, but scale and one reported score do not establish parity.
Observed facts
The source says the API preview is available and describes the model as having 1 trillion parameters, 49 billion active parameters.
1The source says training used Mistral’s own cluster of 3,800 NVIDIA Grace Blackwell GPUs.
1The source reports two API reasoning settings, “none” and “high,” and an Artificial Analysis score of 38 versus 9 for Large 3.
1The source says Mistral promises open weights at the end of the month; this is a future commitment, not a completed release.
1
Mechanism
Mistral is offering API access before the promised open-weights release, while presenting a model with 1 trillion total parameters and 49 billion active parameters.
Why now
A concrete preview is available through the API, while the company has set an “end of this month” expectation for open weights.
Implications
The split between API preview and later weights means access, reproducibility, and deployment options may change only if the promised release arrives with usable terms.
Evidence boundary
The supplied source is a commentator’s account of the release, and its single benchmark score cannot establish broad capability, cost, or frontier standing.
Watch conditions
Check whether Mistral releases the weights on schedule, with what license, and how independent evaluations compare across tasks.