Mistral Launches 'Lechonk' AI Model

The model's name, 'Lechonk,' emerged from community memes related to its substantial size.

It features a mixture-of-experts architecture with a total of one trillion parameters, showcasing a significant engineering effort. Matthew Berman describes it as "Mistral Large 4 aka Lechonk."

With a total size of one trillion parameters, the model utilizes a subset of active parameters per request to unify instruction, reasoning, and agentic capabilities, as demonstrated by strong performance in the Cyber Gym benchmark. Despite its total size of one trillion parameters, it only employs a subset of active parameters per request, enhancing efficiency.

The model demonstrates strong performance in cybersecurity benchmarks, such as Cyber Gym, and its hybrid nature combines instructing and reasoning functions. Matthew Berman explained, "It is a hybrid instruct and reasoning which is a mixture of experts which means the model total size a trillion parameters is obviously very big and most people can't run that."

Mistral Large 4 is priced at $1.36 per million input tokens and $4.18 per million output tokens, as announced by Matthew Berman. This model operates with 49 billion active parameters, a notable ratio against its total parameter count, indicating a high level of efficiency.

The performance of Mistral Large 4 is reported to be competitive with other global open-source models, highlighting its optimized architecture for cost-effective operations.

US closed-source frontier models, such as those from OpenAI and Anthropic, maintain a significant lead, estimated to be at least six months ahead of current open-source alternatives. These closed-source providers are continually improving speeds and reducing prices, intensifying market competition.

Businesses face a strategic decision between embracing the privacy and ownership benefits of open-source models and leveraging the advanced capabilities and speed offered by closed-source frontier labs.

Mistral asserts that its new model significantly surpasses any other open-weight model developed in the US or Europe, though the speaker notes this is "not a high bar" given the strong competition from Chinese open-source models. The model is currently undergoing 'red-teaming' in cybersecurity environments to enhance its robustness.

The model weights are scheduled for release by the end of the current month, allowing broader access and community engagement.

Frontier Labs, as observed by Matthew Berman, has established a 'Trusted Partner' framework where they provide reduced moderation and guardrails to vetted partners and state authorities. This strategy contrasts with the intense guardrails maintained for public use, selectively removing them for specific trusted entities.

The speaker acknowledges this differentiated access approach as a practical and necessary method for managing the deployment of advanced open-source models, balancing public safety with specialized applications.

On the Deep 1.1 benchmark, Mistral Large 4 achieved a score of 62, positioning it behind Qwen (GLM 5.3) and Kimi K3.

In Terminal Bench rankings, Kimi K3 leads with 22, followed by GLM 5.3 at 40, while Mistral Large 4 is not explicitly ranked here. However, the AA Cyber Index shows Mistral Large 4 tied for first place at 50, sharing the lead with GLM 5.3.

Matthew Berman highlighted the dramatic improvement, stating, "Look at this jump from Mistral medium 3.5 to Mistral large 4 preview. It went from kind of an embarrassingly low score to now they are competing."

Training occurred on 3,800 Nvidia Grace Blackwell GPUs in a European data center, ensuring the base model was conceived, designed, and inferred within the region to bolster sovereign AI infrastructure. This independent infrastructure marks a departure from the dominance of US and Chinese AI development.

The model is a base model, trained from scratch using 3,800 Nvidia Grace Blackwell GPUs located in a European data center, rather than being a fine-tuned version of existing models like Kimi or Qwen.

Mistral Large 4 is currently accessible via API, but it lacks the intuitive, plug-and-play 'chat' application experience common with closed-source models. The speaker encountered technical difficulties, such as excessive 'thinking tokens' that quickly exhausted the context window, underscoring the complexity for general users.

Achieving widespread adoption for open-source models hinges on bridging this usability gap, making them as straightforward to integrate and use as their closed-source counterparts. As Matthew Berman noted, "It is not easy to take an open-source model and just plug it into an agentic harness and it'll just work."

Matthew Berman stated, "Most people won't be doing that."

Consequently, open-source models are currently best suited for companies with the necessary technical expertise and the capacity to operate at scale, where customization and control outweigh ease of initial deployment.

Le Chonk has achieved a top-five global ranking for cybersecurity analysis, outperforming other open-weights models developed outside China. Its performance is evaluated on its ability to identify and rectify security flaws in software.

Matthew Berman emphasized its strong showing, stating, "In fact, not just good for an open-source open weights model, but actually competitive with all models, including closed source."

Anthropic's Claude Opus 5.5 consistently holds the top spots in intelligence indices, notably surpassing competitors like Fable. Matthew Berman remarked on its dominance, stating, "Every single top spot is taken by an anthropic model starting with Claude Opus 5.5 which is insane that it beats Fable."

Mistral Large 4 preview currently occupies the 25th position in these intelligence rankings. When considering intelligence versus cost benchmarks, both Anthropic and OpenAI models lead in the 'cheap and good' quadrant, indicating superior value.

This market positioning highlights their ability to deliver high-performing models at competitive price points, setting a high bar for other developers.

Recent releases, such as Claude Opus 5.5 from Anthropic and GPT 6.1 Soul from OpenAI, represent significant advancements in AI technology. These companies appear to be 'pacing the frontier,' opting to release highly capable, faster, and cheaper versions of existing models rather than immediately launching major new versions like GPT-7 or Fable 6.

Matthew Berman noted that if "pacing the frontier actually means releasing incredibly good models, the best in the world, cheaper, faster than anything else out there, then I'm for pacing." This strategy allows them to continuously enhance their offerings while maintaining a lead in the market.

These benefits are the primary drivers for open-source adoption.

The expectation is that community building and optimization will enhance its performance for specific use cases over time. Matthew Berman noted the trade-off, stating, "But as compared to Mistraw Large 4 preview, what you're getting is a less capable model and more expensive, but you get full control."

Unlike competitors such as GLM53, MIMO, and Opus 5.5 that support 1 million tokens, Mistral Large 4 and Grok 4.7 operate within a 500,000 token limit. Matthew Berman noted, "its context window is only a half of a million tokens as compared to what seems to be the standard of a million tokens."

This limitation contrasts with models like GLM53, MIMO, and Opus 5.5, which provide larger context windows, potentially impacting the model's ability to process extensive information.

At 116 tokens per second, Mistral Large 4 (preview) offers high-speed performance that trails the leader, DeepSeek v4.1 Flash, at 227 tokens per second. This speed places it behind DeepSeek v4.1 Flash, which leads at 227 tokens per second.

Performance is anticipated to improve further through ongoing community optimization efforts, which typically enhance the efficiency of open-source models over time.

Matthew Berman explicitly highlighted it as "the number one model of all the open source models including the Chinese models."

This ranking indicates its superior capability in cybersecurity tasks, outperforming models such as MIMO, Grok, GLM 5.3, and Kimmy K3, though it excludes comparisons with closed-source models from OpenAI and Anthropic.

Cursor is cited as a prime example of a harness that effectively optimizes diverse models, demonstrating how critical this component is.

Many current harnesses struggle to adapt to new models without specific optimization, suggesting that unless the model builder also provides an optimized harness, achieving high performance remains challenging. A model's failure in specific tasks, such as visualization, can often be attributed to harness interaction rather than the model itself.