Mistral publishes weights, halves prices, and makes resellers sign a deal; seven Chinese models outrank it
Le Chonk's open weights arrive with a resale license, a memory wall and a scoreboard that still ranks seven Chinese models ahead. Sovereign buyers get AI no provider can switch off; Mistral gets a funnel into its own paid serving.
Richard Tang · 3 min read
On October 29 Mistral$14B — Mistral, private, latest valuation $14B, the Paris-based AI lab, published the weights of Mistral Large 4, the trillion-parameter model it calls Le Chonk, and cut its API price roughly in half the same day 12. Read the license alongside the generosity. Large 3 shipped under Apache 2.0; Large 4 ships under a modified MIT license that is free to run, fine-tune and commercialize on one condition: any business offering the model to third parties as a service must sign a separate commercial agreement with Mistral 3.
Self-hosting has its own floor. The compressed 2-bit build still needs 512GB of unified memory, a Mac Studio M5 Ultra that costs $22,299 in the US 5. So the weights are free to evaluate, and the money changes hands when a customer scales: rent from Mistral, or pay the toll to resell.
The v5 index puts the closed frontier 32 points above Mistral Large 4
Data
| Value | |
|---|---|
| Claude Opus 5.5 | 58 |
| GPT-6 Astra | 57 |
| Kimi K3.1 Pro | 46 |
| DeepSeek V4.1 Flash | 32 |
| GLM-5.3 | 32 |
| Mistral Large 4 | 26 |
| Muse Spark Open | 25 |
The scoreboard is the risk
What the toll buys sits behind rivals. On the Cyber Index from Artificial Analysis, an independent benchmarking firm, which measures whether a model can find, reproduce and patch code vulnerabilities, Mistral trails both DeepSeek and Xiaomi's+3.12% — Xiaomi, up 3.12 percent today MiMo 3.
On the redesigned v5 Intelligence Index it scores 26, eighth among open models, behind seven Chinese models and just ahead of Meta, against a closed frontier led by Anthropic's$350B — Anthropic, private, latest valuation $350B Claude Opus 5.5 and OpenAI's$500B — OpenAI, private, latest valuation $500B GPT-6 Astra 24. Our earlier piece this week, "Mistral now speaks OpenAI's agent format; its flagship scores 25 against OpenAI's 57", caught the preview at 25; the released model scores 26, one point more, and its position has not changed since that first test 2.
The v5 overhaul widened the gap between the best closed and the best open model from 3 points to 12, but that gap counts Kimi K3.1 Pro's 46, and Kimi is scheduled to release its weights on November 4 4. Against the open models anyone can download today, DeepSeek V4.1 Flash and GLM-5.3 at 32, the released field sits 26 points below the leader 4. The open leaderboard resets this week.
Mistral's cyber defense claim trails both Chinese open rivals
Data
| Value | |
|---|---|
| DeepSeek V4.1 Flash | 57 |
| Xiaomi MiMo-V2.6-Pro | 56 |
| Mistral Large 4 | 51 |
Who the money touches
The buyers arrived anyway. Orange Cyberdefense is integrating a cyber-focused variant into selected services 8. La Banque Postale, a French citizen bank, calls technological sovereignty a prerequisite for customer trust 1. Microsoft−1.57% — Microsoft, down 1.57 percent today Foundry is the first cloud to serve the model 1. Listed exposure runs wider: Nvidia−5.01% — Nvidia, down 5.01 percent today, whose 3,800 Grace Blackwell GPUs trained the model 6; Apple−0.95% — Apple, down 0.95 percent today, which sells the machine the memory wall points at 5; Orange; and Xiaomi, which beats Mistral on the cyber score 3. Samsung led Mistral's September round 9.
The toll is worth what the road is worth
Arthur Mensch's pitch is the capital: "Other labs have raised hundreds of billions to develop equivalent capabilities; in three years, Mistral raised roughly $6 billion", with an IPO within a few years 67. The €3 billion September round at a €21 billion valuation is what that discipline has bought so far 9.
The giveaway is the marketing, the license toll and the halved API are the margin, and the scoreboard is the risk. Sovereign demand is real: institutions want AI no provider can switch off, and the arrival of Orange Cyberdefense, La Banque Postale and Microsoft alongside a model that trails rivals suggests demand outweighs capability in this market. Mistral is doing the work in between, tailoring the weights to a bank's detection logic and a defender's threat environment through its Forge and Studio products 1, and that work is where sovereign value gets created.
But a toll prices the road, and this week the road ranks eighth among open weights, behind labs that answer to a different government. Watch November 4: if Kimi's scheduled weights close the gap Mistral claims to lead, the toll buys less.
Get Model by email, free
The AI labs, their models, and the business behind them. In our daily email, every weekday.
