Mistral Large 4 vs Mistral Large 3

What changed from Large 3 to Large 4

Ten months separate the two releases: Mistral Large 3 (2512) arrived on December 2, 2025 and Mistral Large 4 entered public preview on October 6, 2026. The headline changes are scale, reasoning and context.

• Scale: 1.05T total and 49B active parameters per token, up from 675B total and 41B active (+56% total, +20% active).
• Reasoning: Large 4 is a hybrid instruct-and-reasoning model controlled with reasoning_effort ("high" or "none"); the Large 3 model card describes it as not a dedicated reasoning model.
• Context: 524,288 tokens served by the API (Mistral's docs list 1M), up from 256K.
• Languages: training data covers more than 160 languages including every official EU language; the Large 3 card lists dozens of languages.
• Vision: both read images; Large 4 uses a 1.6B vision encoder, Large 3 a 2.5B one.
• License: Large 3 is Apache 2.0; Mistral has not announced the Large 4 license yet.

Spec comparison table

Large 4 wins on every capacity number except vision encoder size, and Large 3 still wins on availability of weights today.

Price difference per million tokens

Mistral Large 4 costs about 2.7–2.8× Mistral Large 3 at list price, and only about 1.4× during Mistral's temporary 50% sale. Ratios are Large 4 price divided by Large 3 price.

What a typical month costs on each model

A workload of 30M input and 10M output tokens a month costs $30.00 on Large 3, $41.30 on Large 4 at the sale price and $82.60 on Large 4 at list price.

The arithmetic: Large 3 = 30 × $0.50 + 10 × $1.50 = $30.00; Large 4 sale = 30 × $0.68 + 10 × $2.09 = $41.30; Large 4 list = 30 × $1.36 + 10 × $4.18 = $82.60. Reasoning output counts as output tokens, so leaving reasoning_effort on "high" for simple prompts raises the Large 4 bill further; Artificial Analysis rates Large 4 very verbose (200M output tokens for its index against a median of 81M).

If you only need the model for chat, you can skip token math entirely: our Pro plan is $39.90 a month or $199.90 a year for 3,000 credits a month, where one credit covers about one short reply.

Quality jump: what the published numbers show

Mistral did not put Large 3 in its Large 4 launch charts, and Large 3 is not in the current 44-model Vals Index, so there is no shared benchmark row for the two. What Mistral did publish is Large 4 against Mistral Medium 3.5, its other current model, and the gap there is wide.

Large 4 also costs less than Medium 3.5 at list price: $1.36 / $4.18 per 1M tokens against $1.50 / $7.50.

Self-hosting: Large 3 now, Large 4 needs more memory

Large 3 FP8 weights take about 675 GB and fit one 8× H200 node (1,128 GB) with about 453 GB to spare; Large 4 FP8 weights take about 1,050 GB and leave only about 78 GB on the same node. At 4-bit, Large 3 needs about 338 GB and Large 4 about 525 GB, both within 8× H100 (640 GB).

Large 3 runs today on vLLM 1.12.0 or newer. No inference engine has announced Large 4 support yet; the full memory breakdown is on /run-locally.

Switching your API calls from Large 3

Moving to Large 4 is mostly a model-name change to mistral-large-4 (alias mistral-large-4-0) on the same chat completions endpoint. Two things need code changes: with reasoning_effort "high" the message content becomes a list of thinking and text chunks instead of a string, and Mistral's docs ask you to send the full assistant message, thinking chunks included, back in multi-turn history.

Setting reasoning_effort to "none" keeps the plain-string response your Large 3 code already parses.

# Set API_KEY to your Mistral API key first
curl https://api.mistral.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $API_KEY" \
  -d '{"model":"mistral-large-4","messages":[{"role":"user","content":"Summarize this contract clause."}],"reasoning_effort":"none"}'