Mistral AI pushed a public preview of Mistral Large 4, nicknamed “le Chonk,” live on October 6, 2026, through Mistral Studio. The French lab’s own model listing puts the mixture-of-experts (MoE) system at 675 billion total parameters with 41 billion active at inference time. That is the number that matters for anyone trying to run or budget around this model. But it is not the number most people are reading in headlines right now, and the gap between the two figures says as much about the state of AI marketing in 2026 as it does about the model itself.
Secondary reports put Mistral Large 4 at 1.05 trillion total parameters and 49 billion active, a materially bigger model than what Mistral’s own listing shows. Neither Mistral’s announcement nor its model card, as captured in the official release, confirms those larger figures. That discrepancy, not the launch itself, is the story worth unpacking: what it means for buyers trying to size GPU clusters, what it signals about the credibility of parameter-count claims across the open-weight AI market, and where le Chonk actually lands against GLM-5.2, DeepSeek, and Reflection AI’s Beam.
What Mistral Actually Confirmed on October 6
Strip away the unverified numbers and the confirmed facts are narrow but concrete. Mistral AI’s announcement, which gave the model its internal codename ML4 and its public nickname “le Chonk,” opened API access through Mistral Studio on October 6, 2026. The architecture is a multimodal mixture-of-experts design, a choice Mistral has leaned on since earlier Mixtral releases, where a router sends each token to a small subset of specialized expert networks instead of running the full parameter count on every pass.
The model listing that Mistral published alongside the preview states 675 billion total parameters and 41 billion active parameters. That is a far more modest footprint than the 1 trillion-plus figures floating around in secondary coverage, and it is the number developers should plan around until Mistral says otherwise. Mistral also committed to releasing the open weights by the end of October 2026, though it did not lock in an exact date in the official announcement. A specific date of October 27 has shown up in some secondary reporting, but that detail is not confirmed by Mistral’s own material.
No pricing was disclosed for the Mistral Studio preview API at launch. No named Mistral executive or researcher gave an on-record quote in the official announcement. Readers should treat any price or quote circulating outside Mistral’s own channels as unverified until the company publishes pricing tiers, something it has done within days of past model launches.
The 675B vs. 1.05T Parameter Gap, Explained
Parameter counts have become a marketing battleground in the open-weight AI race, and Mistral Large 4 is now a case study in why those numbers need a source attached. Total parameter count and active parameter count measure different things in a MoE model. Total parameters describe every weight baked into the model, including experts that sit idle on any given forward pass. Active parameters describe what actually fires when the model processes a single token. A model with a huge total count but a small active count can run cheaper than its headline number suggests, which is exactly the appeal of the MoE architecture.
When outlets report 1.05 trillion total and 49 billion active for Mistral Large 4, they are describing a model roughly 56% larger on paper than what Mistral’s own listing shows. That’s not a rounding error. It’s the difference between a model that needs a modest multi-GPU node for inference and one that needs a meaningfully larger cluster to host, even accounting for MoE sparsity. Until Mistral clarifies which number is correct, enterprise teams evaluating le Chonk for self-hosting should default to the conservative, officially listed figures: 675B total, 41B active.
This is not a new problem. Model cards have drifted from press coverage before, usually because an early draft spec, a leaked internal planning document, or a third-party benchmark aggregator gets picked up before the vendor locks a final number. What makes this instance notable is the size of the gap and the fact that it is happening at a moment when enterprise buyers are already wary of inflated AI claims after a year of benchmark disputes across the industry.
Why Open Weights by End of October Matters More Than the Spec Sheet
The detail that should carry more weight than any parameter count is Mistral’s commitment to open-source the weights by the end of October 2026. That is a narrow window, roughly three to four weeks from the preview launch, and it puts Mistral in a different competitive lane than labs that keep flagship models closed indefinitely.
Mistral has built its identity around open-weight releases going back to the original Mistral 7B and the Mixtral series, both of which still circulate as downloadable checkpoints on Hugging Face’s model hub, and continuing that pattern with a model in the hundreds-of-billions-of-parameters class is a bigger commitment than it looks. Open-weight releases at this scale carry real infrastructure and reputational cost: once weights are public, competitors can fine-tune on them, redistribute them, and benchmark them without Mistral’s cooperation. A lab only does this when it believes the openness itself is a competitive asset, not just a marketing line.
That calculus matters because the open-weight tier of the market has gotten crowded fast. Reflection AI’s Beam, Zhipu’s GLM-5.2, and DeepSeek’s V4.1-Flash have all pushed open or semi-open releases in recent months, each making efficiency claims that strain easy comparison for the same reason Mistral’s own numbers are confusing right now: total parameters, active parameters, and training compute are reported inconsistently across labs, and few are publishing third-party-audited benchmark runs alongside the model card.
Mistral Large 4 vs. the Open-Weight Field
Mistral Large 4 does not land in a vacuum. It lands into a market where every major open-weight release of the past two months has made a parameter-efficiency claim as its headline. Reflection AI priced Beam at 501 billion total parameters with 23 billion active, and leaned on that gap as its core pitch: a model that performs far above what its active-parameter footprint would suggest. GLM-5.2 and DeepSeek’s V4.1-Flash have each pursued a similar story, cutting inference cost rather than chasing raw scale.
Mistral’s position, at least using its own confirmed numbers, is a 675B/41B model, which puts it in a similar active-parameter class to Beam and comfortably above DeepSeek’s lighter models. If the 1.05T/49B figures that secondary sources are circulating turn out to be accurate, Mistral Large 4 would be the largest total-parameter model in this specific MoE-efficiency cohort, which changes how it should be judged: a bigger total parameter pool paired with a disproportionately small active count is usually a stronger argument for the MoE approach’s efficiency math, not a weaker one, provided the benchmarks hold up.
| Model | Total Parameters | Active Parameters | Open Weights | Source Status |
|---|---|---|---|---|
| Mistral Large 4 (“le Chonk”) — official | 675B | 41B | Due by end of Oct. 2026 | Confirmed by Mistral’s model listing |
| Mistral Large 4 — secondary reports | 1.05T | 49B | N/A | Unconfirmed, not in official announcement |
| Reflection AI Beam | 501B | 23B | Yes | Previously reported by Mistral AI and Reflection AI’s own materials |
| DeepSeek V4.1-Flash | Not disclosed in this fact set | Not disclosed in this fact set | Yes | Previously reported pricing cut of 70% on output tokens |
| GLM-5.3 | Not disclosed in this fact set | Not disclosed in this fact set | Partial | Previously reported safety-bypass testing results |
The Enterprise Angle: What “41B Active” Means for Deployment Cost
For an engineering team deciding whether to evaluate Mistral Large 4 against incumbents already in production, the active parameter count is the number that actually drives GPU memory and latency math, not the total. A 41-billion active parameter model sits in a range that typically requires a multi-GPU inference setup for low-latency serving at reasonable batch sizes, though exact memory needs depend heavily on quantization choices Mistral has not yet published in its developer documentation for this model.
That is a meaningfully different planning exercise than if the 49-billion active figure from secondary reports turns out to be the real one. An 8-billion-parameter swing in active weights, on top of an already large base, is the kind of difference that pushes a deployment from “fits on a single high-memory node” to “needs a dedicated multi-node inference cluster.” Enterprise buyers evaluating le Chonk against the usual shortlist of hosted and self-hosted models should wait for Mistral to publish final, audited specs before locking in infrastructure commitments, rather than sizing hardware against numbers that haven’t cleared Mistral’s own documentation.
Mistral Studio’s role here is also worth flagging. By routing the preview exclusively through its own hosted platform rather than an immediate open release, Mistral keeps full control over rate limits, pricing experiments, and usage telemetry during the weeks before weights go public. That is a pattern other labs have used for staged rollouts: hosted preview first, open weights later, so the vendor can tune serving infrastructure and pricing before the model becomes self-hostable.
Historical Context: From Mixtral to le Chonk
Mistral’s bet on mixture-of-experts architecture didn’t start with Large 4. The lab built its early reputation on Mixtral 8x7B, a model that punched well above its active-parameter weight class and helped popularize MoE as a credible alternative to dense scaling for labs without hyperscaler-level compute budgets. Each subsequent Mistral Large release has grown the total parameter envelope while trying to keep active parameters, and therefore inference cost, from scaling at the same rate.
The nickname “le Chonk” fits a company that has leaned into playful, French-inflected branding for a lineup that otherwise competes on dense technical ground. But the joke sits oddly next to this week’s confusion over the model’s actual size. Mistral built goodwill by being one of the few labs willing to put real weights into the world rather than gate everything behind an API. A parameter-count dispute right out of the gate, even an unintentional one caused by secondary reporting outrunning the official model card, is the kind of friction that chips at that goodwill if it isn’t cleared up quickly.
It also lands at a moment when Mistral has been raising its profile on multiple fronts, including a reported valuation climb and new AI safety tooling released alongside a $24 billion valuation milestone. A high-profile model launch muddied by conflicting specs is a worse look during a period when the company is trying to project operational maturity to investors and enterprise customers alike.
Competitive Pressure: Why the Timing Isn’t an Accident
Mistral Large 4’s preview did not launch into a quiet news cycle. Reflection AI’s Beam has spent the past several weeks making the exact efficiency argument Mistral now needs to answer, and Google, OpenAI, and Anthropic have each shipped or previewed new flagship models in the same window, each pushing a different combination of pricing, context length, and benchmark wins as the headline. Google’s Gemini 4 Argon arrived with split benchmark results and new per-token pricing, while Anthropic’s Claude Opus 5.5 claimed a benchmark win over GPT-5.6 Sol at a third of the cost.
Mistral launching a preview with no pricing and no benchmark disclosures, into a week this crowded, reads as a company trying to claim mindshare before it has finished the harder work of validating and documenting the model. That is not unusual for a fast-moving AI market, but it does mean the open-weight release at the end of October carries extra weight: that is when independent researchers will be able to actually verify total and active parameter counts themselves, rather than relying on either Mistral’s listing or secondary reporting.
Zhipu’s GLM-5.3 has its own credibility questions after safety-bypass testing results circulated this year, and DeepSeek’s aggressive pricing cuts on V4.1-Flash have already reset expectations for what “efficient” open-weight inference should cost. Mistral entering this field with an unresolved spec dispute means its first real test won’t be a benchmark leaderboard. It will be whether the open-weight release in late October matches what the company told the market on October 6.
Market Impact: Valuation, Investor Signal, and the Credibility Cost
Model launches do double duty for AI labs right now: they are product news and investor signal at once. Mistral has been raising its profile through its valuation climb, and a flagship model preview is exactly the kind of event that feeds into how investors and enterprise partners price the company’s next funding round or contract renewal. A clean launch with confirmed, consistent specs reinforces that story. A launch where the headline parameter count and the official model card disagree by 56% does the opposite, even if the underlying model performs well once independent testing starts.
The practical market impact falls into three buckets. First, enterprise procurement teams now have a concrete reason to wait for the open-weight release before committing budget, which slows Mistral’s ability to convert preview buzz into signed deals this month. Second, competitors get a news cycle to point at the confusion as evidence that parameter-count marketing across the industry needs more scrutiny, a framing that benefits labs like DeepSeek that have leaned on transparent, documented pricing cuts instead of scale claims. Third, Mistral itself has a short window, the same three-to-four-week stretch before open weights ship, to clarify its own numbers before the ambiguity hardens into a lasting credibility mark against the release.
What to Watch Before the Open-Weight Release
Three things are worth tracking between now and whenever Mistral actually publishes the open weights. First, whether Mistral issues a clarification or correction on the parameter count, either confirming 675B/41B as final or revising it upward toward the 1.05T/49B figures circulating in secondary coverage. Second, whether Mistral Studio pricing drops before the open-weight release, which would give the clearest signal yet of how the company is positioning le Chonk against hosted rivals. Third, whether independent benchmark runs appear once API access widens beyond the initial preview, since no official benchmark numbers were part of the October 6 announcement.
None of those have resolved yet, and that is exactly why treating this launch as settled news would be premature. The responsible read, for now, is that Mistral shipped a real, working preview of a large MoE model, committed to open-sourcing it within the month, and left enough specification ambiguity that the next three to four weeks matter more than the launch-day headlines.
Five Predictions for the Weeks Ahead
- Mistral will likely publish a clarifying post or updated model card within one to two weeks, given how far apart the official and secondary parameter figures are.
- Pricing for Mistral Studio access to Large 4 will probably land before the open-weight release, following the pattern of recent flagship launches from other labs.
- Independent benchmark comparisons against Beam, GLM-5.3, and DeepSeek V4.1-Flash will surface within days of broader API access, not after the open-weight drop.
- If the open weights ship on schedule by the end of October, expect a wave of community fine-tunes and quantized variants within the first week, mirroring how quickly the ecosystem moved on past Mistral open releases.
- Expect competitors to use the parameter-count confusion in their own marketing, framing it as a reason to trust labs that publish audited, consistent specs from day one.
| Date | Milestone | Status |
|---|---|---|
| October 6, 2026 | Mistral Large 4 preview launches on Mistral Studio | Confirmed |
| October 6, 2026 | Official model listing shows 675B total / 41B active params | Confirmed |
| Unspecified, “secondary reports” | 1.05T total / 49B active params claimed | Unconfirmed |
| By end of October 2026 | Open weights release committed by Mistral | Confirmed commitment, exact date pending |
| October 27, 2026 (reported) | Specific open-weight release date | Unconfirmed, secondary reporting only |
Key Takeaways for Developers and Buyers
Strip the noise and three things hold up. Mistral shipped a real preview of a large multimodal MoE model on October 6, not a teaser or a waitlist. The company’s own listing says 675B total and 41B active parameters, and that is the figure worth planning around, not the larger 1.05T/49B numbers circulating in secondary coverage. And the open-weight commitment by the end of October is the detail that will actually let outside researchers settle the parameter-count question, since self-hosted testing is the only way to verify a model’s real footprint independent of any vendor’s marketing copy.
For teams building roadmaps around open-weight models, the practical move is to treat this week’s headlines as provisional. Budget infrastructure conversations around the confirmed 41B active parameter figure, watch for a Mistral clarification, and hold off on hard commitments until either the open weights ship or Mistral locks its own numbers down in writing.
Frequently Asked Questions
How many parameters does Mistral Large 4 actually have?
Mistral’s own model listing states 675 billion total parameters with 41 billion active at inference time. Secondary sources report a larger 1.05 trillion total and 49 billion active, but those figures are not confirmed by Mistral’s official announcement. Until Mistral clarifies, the 675B/41B figures are the ones backed by the company’s own documentation.
When will Mistral Large 4’s weights be open-sourced?
Mistral’s announcement commits to releasing the weights by the end of October 2026 without naming an exact date. A specific date of October 27 has appeared in secondary reporting but is not confirmed by Mistral’s own material.
Is “le Chonk” an official name or a nickname?
It’s a nickname. Mistral’s announcement describes the model as “unofficially ML4, very officially: le Chonk,” with Mistral Large 4 serving as the formal product name.
How much does access to Mistral Large 4 cost?
No pricing was confirmed in Mistral’s official announcement. Any price figures circulating elsewhere should be treated as unverified until Mistral publishes its own Mistral Studio pricing tiers.
How does Mistral Large 4 compare to Reflection AI’s Beam?
Using Mistral’s officially confirmed figures, Large 4’s 675B total / 41B active parameter footprint is larger in both total and active count than Beam’s previously reported 501B total / 23B active configuration. Direct performance benchmarks between the two have not been published as of this writing.
Is Mistral Large 4 multimodal?
Yes. Mistral’s announcement describes Large 4 as a multimodal mixture-of-experts model, though specific details on supported input types beyond text were not detailed in the official announcement.
Why do the parameter counts for Mistral Large 4 differ across news sites?
Discrepancies like this typically trace back to draft specs, internal planning documents, or benchmark aggregator pages that circulate before a vendor finalizes its own model card. In this case, Mistral’s official model listing and widely circulated secondary figures disagree by a wide margin, and the gap has not yet been resolved publicly.
Can I self-host Mistral Large 4 once the weights ship?
Mistral has a track record of releasing open weights for self-hosting, consistent with its commitment to open-source Large 4 by the end of October 2026. Exact hardware requirements will depend on the final confirmed active parameter count and whatever quantized variants the community or Mistral itself produces after release.
