Mistral Releases 1T-Parameter Large 4 Weights, With a Commercial Catch
Mistral released the weights of its 1T-parameter Large 4 under a modified MIT license, claiming coding and cyber gains, but the license bars model-as-a-service use without a commercial deal.
Mistral released the weights of Mistral Large 4 (Le Chonk), a 1-trillion-parameter multimodal model, after a public preview. The company says three weeks of additional reinforcement learning lifted coding and defensive cybersecurity performance while cutting inference costs.
Mistral reports the model leads open-weight coding benchmarks: 19.7% on Cognition's FrontierCode 2.0, 41% on Terminal-Bench 4.0, and 48% on the Artificial Analysis Coding Agent Index, tied with GLM-5.3 while using 47% fewer tokens per task. It scores 51% on the Artificial Analysis Cyber Index and refuses 96.6% of malicious cyber requests, per Mistral.
The weights carry a modified MIT license requiring a separate commercial license for model-as-a-service offerings.
Key takeaways
- Mistral offers three reasoning modes: Low for simple tasks, Medium for everyday work, and High as the default for hard use cases, letting developers trade performance against inference cost.
- In Surge Code's blind pairwise evaluations, Mistral Large 4 was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5 among the models evaluated, according to Mistral.
- Mistral reports Finch score 73.3% for finance workflows, second to Kimi K3 among open models; BrowseComp 84.5%; and 97.1% attack resistance on the B3 Agent Security Benchmark, up from 93.3% in preview.
- The model is available through Mistral Studio, the Mistral API, OpenRouter, Hugging Face, vLLM, SGLang and NVIDIA NIM, with managed deployment via Microsoft Foundry. API pricing is $0.70 per million input tokens, $0.07 cached input and $2.10 output.
- Mistral says it monitored live cybersecurity traffic, observed refusals against black-box attacks on live targets but saw potential for real-attack uplift, and is modifying guardrails. A vetted variant for penetration testing and red-teaming is limited to authorized systems.
What people think
Analysis details
32 comments analyzed. Analyzed .
The sample selects active discussion branches and includes available parent comments. It may omit parts of the full thread. The analysis included 32 of 59 usable stored comments because of the input limit. Selected themes do not measure community opinion or how common a view is.
Open weights or open source?
Mistral's modified MIT license requires a commercial license for Model-as-a-Service, prompting debate over whether calling it open source is misleading. The OSI definition demands training data and code, which no major lab provides. Some argue open weights never meant open source, while others say weights available is more accurate. The dispute reflects broader concerns about rights and fairness in AI licensing.
Commercial license for service providers
The license requires businesses operating a Model-as-a-Service to negotiate a separate commercial agreement with Mistral. This mirrors HashiCorp's model, where general use is free but competing cloud services need a license. Some see this as fair and accurate, while others note it restricts commercial applications and makes the open label questionable.
Price cuts and hidden cache costs
Mistral cut API prices to $0.7 per million input tokens and $2.1 per million output tokens, nearly half the preview rates. However, users highlight that cache rates are critical; one reports a >97% input cache rate, making providers with exorbitant or hidden cache costs unattractive. Even at the new price, some find it too expensive and slow compared to alternatives like Opus 5.5.
Benchmark numbers and competitiveness
Critics call the benchmark numbers dreadful and note the cost is higher than Sol, arguing there are smarter models for less money. Others suggest the model wasn't benchmaxxed, which could be a reason to try it. Self-reported evaluations indicate it is not competitive with top-tier models, though the full release may warrant reassessment for some use cases.
Training data barriers to open source
Truly open source would require releasing training data and code, but the data is likely petabytes and its release would violate original licenses. Without training data, models cannot be reproduced. The OSI definition requires training code and data composition, which only Nvidia and Allen AI meet among major vendors. Distributed training faces technical hurdles like fast interconnect requirements.
Capital concentration and worker leverage
Training frontier models is prohibitively expensive, concentrating capital at the top and reducing leverage in the job market. Some worry open weights may not survive once the AI cycle cools. Suggestions for distributed training or SETI@HOME-style efforts face technical and funding challenges. Nvidia's incentive to subsidize open models is questioned because consumer GPU sales are small compared to data centers.