What happened
Mistral’s changelog lists Mistral Large 4, model identifier mistral-large-4, as available in Public Preview on October 5. The entry describes it as an open-weight, general-purpose multimodal model built on a Mixture-of-Experts architecture. It gives the context window as 1M tokens. The entry does not describe benchmark results, modalities in detail, or a specific intended workflow beyond the general-purpose characterization.
The changelog says open weights are coming soon. That is a stated future plan, not a record that the weights were already downloadable on October 5. The listing therefore separates two routes builders may care about: preview access through Mistral’s service now, and possible access to weights later. The entry does not specify a weight release date, licence terms, hardware requirements, or whether the eventual weights will match the preview model exactly.
Mistral also lists launch pricing as 50% off for two weeks. It does not give the underlying price, the exact calendar end date of the offer, or a price table in this entry. The discount is consequently useful as an indication that preview use has a promotional window, but insufficient for a cost estimate on its own. Teams comparing providers would need to consult the current model pricing and confirm what the discount applies to before projecting sustained usage.
The public-preview label matters: the changelog does not call Large 4 generally available. It also does not list rate limits, service-level commitments, regional availability, or other preview restrictions. Builders can treat the listing as an opening for evaluation, not as confirmation that the model is ready for every production workload. Mistral links the changelog entry to a model card, but the changelog itself is the published record for the release date, access status, context-window claim, and launch promotion.
That distinction matters operationally. A team can begin a contained comparison against an existing model while preserving a fallback path if preview behavior, access rules, or economics change. The release notice does not promise version stability, migration guarantees, or a timetable for moving from preview to general availability. It is therefore not enough to assume that a prototype built during the discounted period will have identical commercial terms or operational characteristics later.
What the archive adds
New Runtime’s October 6 note on the Large 4 preview framed open weights and self-deployment as the practical test of Mistral’s sovereignty pitch. This changelog entry is a restatement of the preview announcement’s core facts, adding the compact record that the launch offer was 50% off for two weeks; it does not confirm weights or a later pricing change. New Runtime’s October 30 coverage of the released weights is a later, distinct development, not evidence that weights were available at preview launch.
Trend signal
This release does not yet fit a pattern New Runtime tracks. An open-weight label and a preview API listing alone do not show how teams will deploy the model, what completed tasks will cost, or whether a particular harness or protocol becomes the durable integration boundary.
Why now
The changelog dates public preview to October 5 and describes a two-week launch discount. That creates a short evaluation window for teams interested in testing the model through Mistral’s platform. The source does not explain why Mistral chose this timing or connect the offer to a market event.
Why it matters
A one-million-token context window can make Large 4 worth evaluating for workloads that need broad context in a single request. But context capacity is not the same as reliable performance, acceptable latency, or lower total cost. The launch discount could make short trials cheaper, yet the changelog gives no baseline rate from which to calculate spend after the offer.
For builders, the practical implication is to separate three decisions: whether the preview can handle a representative task, whether API economics work at expected volume, and whether the promised open weights eventually support a viable self-hosting path. The source confirms only the preview, the model’s stated architecture and context size, and the promotional discount. It does not settle the other two decisions. Teams should avoid designing around future weight availability until release timing, licence and deployment requirements are clear.
A useful evaluation should record task success, input size, output size, response time, and failure handling under the team’s own workload. Those measurements make the promotional period useful even if the eventual list price differs. They also prevent a large advertised context window from becoming the sole selection criterion when retrieval, prompt construction, and verification may determine the actual cost and quality of an agent workflow.
Limits and uncertainty
Mistral’s changelog identifies the model as a public preview and gives its context window and launch discount, but does not quantify quality, latency, throughput, or reliability. It does not disclose the underlying price or an exact date when the two-week offer ends. The entry says open weights are coming soon without a date or licence terms. These are company-published claims and release details; they are not an independent evaluation of the model or confirmation of production readiness.
Reader takeaway
Mistral Large 4 is available to try in public preview, with a stated 1M-token context window and a temporary 50% launch discount. Treat the offer as an evaluation opportunity, not a cost forecast: confirm current rates and preview limits, and wait for the actual weights and their terms before planning self-hosting.
Open question
When will Mistral publish Large 4’s weights, and what licence and deployment requirements will accompany them?