Enjoying this issue?
Get tomorrow's AI & engineering digest in your inbox — hand-picked, summarized, and always spam-free.
TLDR
Mistral Large 4, a 1.05 trillion parameter mixture-of-experts model, will not fit in consumer hardware: its 4-bit weights need about 525-590 GB, far exceeding the 512 GB max of the largest Mac Studio. The claimed 82% cyber score is real but narrow, based on a 131-task exploit benchmark where many competing models refuse to participate. For most users, the model is impractical at home, but it may be worth testing for security tasks if other models refuse.
Key points
Mistral Large 4 has 1.05 trillion parameters and uses a mixture-of-experts architecture.
At 4-bit quantization, the model's weights require about 525-590 GB of memory.
Mistral says it will release the weights by the end of October, possibly October 27.
The model was trained on 3,800 Nvidia GPUs, according to Mistral.
On the Cyberjim E2EA benchmark, Mistral Large 4 scores 81.7% among models that attempt the tasks.
Claude Opus 5.5, GPT6 Astra, and four other models refuse at least 98% of the benchmark tasks.
The model's inference speed is limited by memory bandwidth, not compute.
Holding 560 GB of memory costs roughly $8,000-$10,300 in consumer DDR5 or $16,400-$23,000 in server DDR5.
Mistral Large 4 wrote 200 million output tokens in testing, compared to a median of 81 million for other models.
The model's license is custom and should be read before use.
Tools mentioned
Techniques
- Mixture of experts
- 4-bit quantization
- Router-based expert selection
- Memory bandwidth limited inference
Stop scrolling. Start reading smarter.
Receive the day's most important AI & engineering updates in one concise email. No spam.
Transcript (captions)
Mistro large 4, the one trillion parameter model won't fit in the biggest Mac studio which holds 512 gigabytes. At 4bit, its weights need about 525 to 590 GB. Mistrol says it will release the
weights by the end of October. Can you run them at home? And is the 82% cyber score real? In a mixture of experts model, a router picks two expert networks for each token and only those
two work. Mix trial stores 47 billion parameters but uses 13 billion. Mistrol hasn't published a large four expert count. Four bit means half a bite per weight. So Mistl's 1.05 trillion
parameters take about 525 GB or about 590 GB at 4.5 bits. The router can pick any expert, so all stay in memory. Each token reads only the 49 to 52 billion active parameters, about 25 to 29 GB.
speed tops out at memory bandwidth divided by bytes read per token. A Mac at 1.2 terabytes per second reading 27 gigabytes would reach about 44 tokens a second, but it can't hold the weights.
Holding 560 GB costs about $8,000 to $10,300 in consumer DDR5 or $16,400 to $23,000 in server DDR5 that buys memory only. Mistl says it trained the
model on 3,800 Nvidia GPUs. The 3.8% claim divides that by 100,000 for GPT6 Astra from an unclear NVIDIA post. Artificial analysis and independent group scores the preview 38 level with
GPT6 Luna Max. Astra scores higher. Artificial analysis measures $1.13 per intelligence index task against 25 cents for GLM 5.3 flash. Mistral large 4 wrote 200 million output tokens and the median
model wrote 81 million. Cost is price times tokens. So more writing costs more. The test is Cyberjim E2EA with 131 tasks. The agents exploit must crash the unpatched project. Its patch
must fix the crash and the patched project must pass its own tests. A refusal score zero. Claude Opus 5.5. GPT6 Astra and four other models refuse at least 98% of these tasks. Among
models that try, Mistro large 4 scores 81.7%. MIMO V2.6 Pro 78.6% and GPT6 Luna Max 77.9%. That lead is about four tasks of 131.
For home use, skip it because the weights alone need about 525 GB. If other models refuse your security tasks, test it when the weights ship. Plan for the 27th of October. Read the custom
license first. Will the lead hold on the final