Mistral is BACK! (Le Chonk)

Main topic

• Mistral Large 4 (“Lechonk”) is presented as a frontier open-weight model “baked completely in Europe from scratch” (i.e., not derived from a Chinese open-source base model).
• The video covers:
  • Capabilities
  • Pricing
  • Benchmarks (coding, agents, multimodal, cyber/finance/legal)
  • Infrastructure/sovereignty
  • A hands-on usability critique focused on integrating the model into agentic coding workflows

Product / technical highlights

Model type & size

• Open-source / open weights model
• Claimed scale: ~1T parameters total
• 49B active parameters (uses a subset of parameters per task via Mixture-of-Experts (MoE))

Hybrid capabilities

• Combines instruction + reasoning
• Uses Mixture-of-Experts (MoE)
• Multimodal input
  • Unifies instruction/reasoning with “agent/aenta” capabilities in one model

Cybersecurity focus

• Marketed as “frontier” in cybersecurity for an open-weight model
• Benchmarked strongly on security-focused evaluations

Public preview + access

• Public preview launching for Mistral Large 4
• Users can obtain access via a Mistral API key
• Not yet positioned as plug-and-play for typical chat/UI products
  • The speaker notes it’s not easily available in Mistral’s chat app (where they’d prefer to use it)

Pricing (as quoted)

• $1.36 per million input tokens
• $4.18 per million output tokens
• The speaker notes the pricing is unusually precise

Benchmarks & reported results (key numbers mentioned)

Coding / developer benchmark (Deep 1.1)

• Mistral Large 4 preview: 62
• GLM (open-source): GLM 5.3 (referenced as previously top)
• Kimmy K3 (with “Kimmyode CLI”): 68
  • Speaker’s note: this is 2nd place overall for the chart

Terminal Bench

• Mistral Large 4 preview: ~22
  • Speaker notes it’s below GLM 5.3

Cybersecurity (AA Cyber Index)

• Mistral Large 4 preview: 50, tied for first (with GLM 5.3 Flash)
• Other models (e.g., Kimmy K3) score lower

Automation / agent benchmarks

• “Automation bench” score:
  • Mistral Large 4: 59.9
  • GLM 5.3: 62.2
  • Kimmy K3: 58.3
• Finance agent V2 described as competitive
• Harvey’s legal agent described as very strong/dominant on legal benchmarks
  • Speaker note: Mistral is not the top there

Artificial Analysis Intelligence Index

• Speaker’s take:
  • Mistral Large 4 preview ranked very low (25th/25th)
  • Score around 38
  • Compared to top Anthropic/OpenAI models

Key takeaway from the benchmarks: strong in cybersecurity and some agent/coding evaluations, but weaker/less “overall” on broad intelligence index metrics compared with current closed-source frontier systems.

Training & “sovereignty” (Europe-only origin)

• Speaker emphasizes sovereignty:
  • conceived, designed, created, trained, and run for inference from Europe
• Trained from scratch
  • not post-trained on an existing foundation model (e.g., “take Qwen/other and fine-tune”)
• Training infrastructure:
  • 3,800 NVIDIA Grace Blackwell GPUs
  • in a European data center

Comparison to closed-source frontier + open-source tradeoffs

Closed-source model positioning (as described)

• OpenAI/Anthropic-like models are portrayed as:
  • about ~6 months ahead (speaker estimate)
  • often cheaper/faster via new releases and price reductions

Speaker’s argument for open-source value

• Full control (including hosting anywhere)
• Privacy
  • with zero data retention guarantees (relative to managed APIs)
• Ability to tune guardrails and agent behavior

Usability / integration critique (hands-on tutorial-like complaints)

The speaker tests Mistral Large 4 via API and demonstrates integration problems with an agentic coding harness:

• Not easy to plug into agent frameworks
  • described as not “plug-and-play”
• Excessive chain-of-thought / thinking tokens
  • with high thinking effort, the model emits many “thinking” tokens
  • this burns context space and can cause the model to run out of room
  • context/window management becomes the user’s responsibility, not automatic
• Rubik’s cube simulation workload example
  • First run: chain-of-thought/output explosion caused context issues
  • Then functionality checks:
    • scramble showed planned moves, but 3D animation didn’t move
    • later iteration: spinning worked, but colors disappeared
    • autosolve seemed correct briefly, then colors reset at the end
• Conclusion from the demo:
  • many failures may come from harness interaction, not necessarily the base model
  • performance may break down without a harness optimized for the model
• Praise for Cursor
  • the speaker claims Cursor can make other models work well in its agentic harness
  • and is easier there than in “open-code” setups

Limitations highlighted

• Context window only ~500k tokens
  • contrasted with a common ~1M context window in many other models
• Integration friction remains the main blocker for broader adoption

Speaker/source identification (end)

• Main speaker: the YouTube reviewer/host (not named in the provided subtitles)
• Sources referenced for benchmarks/tools:
  • Artificial Analysis (cyber/Intelligence index evaluations)
  • Deep 1.1 benchmark
  • Terminal Bench
  • Cyber Gym
  • AA Cyber Index
  • Harvey’s legal agent, finance agent V2, automation bench
  • Cursor (used for integration comparison)
  • Hugging Face, OpenAI, Anthropic (for comparison context)