Title card reading 'One rate card. One keynote date.' over the label 'Mistral Large 4 · Qwen 4 Max', with a filled Mistral Large 4 rate card for price, context and weights beside a dashed Qwen 4 Max card marked 'not published' on every row.
Guides & Insights

Mistral Large 4 vs Qwen 4 Max: One Has a Rate Card, the Other Has a Keynote Date

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Mistral Large 4 and Qwen 4 Max are not two products at the same stage of existence. Mistral Large 4 is a shipping 1.05-trillion-parameter open-weight model — public preview on 6 October 2026, weights on 29 October — with a published price, a 1M-token context and a benchmark sheet. Qwen 4 Max is a tier name Alibaba put on a stage at its Apsara conference on 22 September 2026. As of 4 November 2026 it has no model card, no API identifier, no price, no context window, no weights and no published benchmark score. The Qwen flagship you can actually call today is Qwen3.8-Max, which shipped on 3 August 2026 and is a different, older model.

So the useful comparison is not the one in the title. It is a three-way decision: buy Mistral Large 4 now, buy Qwen3.8-Max now, or wait for Qwen 4 Max. This piece prices the first two against each other, then sets the five thresholds Qwen 4 Max would have to clear to be worth the wait.

The state of play, dated

• Mistral Large 4 — preview 6 October 2026; open weights 29 October 2026; 1.05 trillion total parameters with 49 billion active per token; text and image in, text out; 1M context claimed and 524,288 tested by Artificial Analysis; $1.36 input and $4.18 output per million at list, or $0.68 and $2.09 on a standing half-price promotion with no published end date.

• Qwen 4 Max — announced 22 September 2026 as the top of a four-tier family alongside Qwen 4 Flash, Qwen 4 Plus and an open-weight Qwen 4 27B. Alibaba's written material described the family as in training as of 2 November. On 29 October 2026 at Apsara Amsterdam the company's CTO said it would arrive "next week"; that week opened on 2 November with nothing shipped, and it is still empty on 4 November.

• Qwen3.8-Max — generally available 3 August 2026; Alibaba reports 2.4 trillion total parameters with 95 billion active; text, image and video in, text out; 1M context; $2.00 input and $6.00 output per million, with a $0.25 cached-input rate.

The only Qwen 4 artifact with a number attached is a client-side default. Qwen Code merged commits on 28 and 31 October that size every qwen4 identifier, including the Qwen 4 Max spelling, at 1M tokens and give the family a reasoning-effort ladder. That is a CLI configuration written before the model shipped, not a published context window. Treat it as a hint, not a spec.

One number to retire before it spreads further: the 5-to-10-trillion-parameter figure that circulates with Qwen 4 coverage belongs to a later roadmap discussion about Qwen 4.5 and Qwen 5. It is not a Qwen 4 Max specification, and Alibaba has published none.

What Qwen 4 Max would have to clear

Five thresholds, in order of how much they change the decision.

• Output price at or below $2.09 — Mistral Large 4's current promotional output rate. At $6.00, Qwen3.8-Max is nearly three times that. A Qwen 4 Max that arrives at the same $6.00 does not undercut the incumbent it replaces, let alone the Mistral promotion.

• A served context of at least 1M tokens, with a documented maximum output. Mistral Large 4 claims 1M and Artificial Analysis tests 524,288. Qwen3.8-Max claims 1M with a 131,072 output ceiling on the dated snapshot. A Qwen 4 Max with a 256K window would be a downgrade on paper.

• Video input, retained. Qwen3.8-Max accepts text, images and video; Mistral Large 4 accepts text and images. This is the one capability where the Qwen line has a hard advantage, and Qwen 4 Max would need to keep it.

• Open weights under a licence that permits commercial self-hosting. Mistral's modified MIT licence does, with a model-as-a-service carve-out. Alibaba has said a Max-class open-weight release is coming; the licence terms decide whether it is usable.

• A shared benchmark where it beats 51% on the Artificial Analysis Cyber Index, 48% on the Coding Agent Index or 41% on Terminal-Bench 4.0 — on the same harness, not a vendor-selected eval.

None of these is published, so none can be scored. The list is useful anyway: it tells you what to read first on launch day, and it stops "Qwen 4 Max is newer" from becoming an argument by itself.

The real head-to-head: Mistral Large 4 vs Qwen3.8-Max

On the two models that exist, the cards diverge in a way that makes a single winner unlikely.

• Price per million — Mistral Large 4 $1.36 in and $4.18 out at list, $0.68 and $2.09 promoted; Qwen3.8-Max $2.00 and $6.00. Mistral is cheaper on both lines even at list, and roughly a third of the Qwen output price during the promotion.

• Cached input — Mistral Large 4 $0.14 per million; Qwen3.8-Max $0.25. For an agent that re-sends a stable prefix, Mistral's cache is the larger saving.

• Context — 1M claimed on both. Mistral's tested configuration is 524,288 tokens; Qwen3.8-Max lists 1M with a 131,072-token output ceiling on the 0902 snapshot.

• Modality — Mistral Large 4 takes text and images; Qwen3.8-Max takes text, images and video. If your pipeline ingests video, the comparison ends here.

• Weights — Mistral Large 4 is downloadable at roughly 1.06 TB under a modified MIT licence. Qwen3.8-Max's open-weight terms are the thing to check before planning a self-host, because a 2.4-trillion-parameter checkpoint is a different procurement category from a 1-trillion-parameter one.

• Benchmarks — Mistral's headline numbers are vendor-reported: Artificial Analysis Intelligence Index v5 at 26, Coding Agent Index at 48%, Terminal-Bench 4.0 at 41%, Cyber Index at 51%. Qwen3.8-Max has a wider public benchmark record, but the figures move between index versions, so compare only within a single run.

Scoreboard titled 'Mistral Large 4 vs Qwen3.8-Max', listing input, output, cached input, context, modality and weights, with a note that Qwen 4 Max has no column because it has no published figures and a verdict strip on price, cache and video input.

Cost per task, with the promotion flagged

Per-token prices flatter models that write more. Here is one workload: 1 million input tokens and 100,000 output tokens.

• Mistral Large 4 at list: $1.36 plus $0.418 — about $1.78.

• Mistral Large 4 at promotion: $0.68 plus $0.209 — about $0.89.

• Qwen3.8-Max: $2.00 plus $0.60 — about $2.60.

On that mix Mistral Large 4 is 32% cheaper at list and 66% cheaper on promotion. If Mistral's claim of 47% fewer tokens per task on the Coding Agent Index holds, the gap widens, because output is where agent runs spend. The promotion is the fragility: it has no published end date, and at list the gap narrows to about a third.

The Qwen counter is not price. It is video input, a broader public benchmark record and the possibility that Qwen 4 Max resets the price. None of those is a per-token number today.

Where each one breaks

• Mistral Large 4 self-hosting — roughly 1.06 TB of weights, an eight-GPU serving class, and a licence that forbids model-as-a-service use without a commercial agreement. The API is easier than the download.

• Mistral Large 4 context — 1M claimed, 524,288 tested. If your workload is a single 800K-token codebase, test before you commit.

• Mistral Large 4 modality — no video input.

• Qwen3.8-Max price — $6.00 output is the highest line in this comparison, and long agent runs are output-heavy.

• Qwen3.8-Max certainty — its benchmark numbers are reported across several index versions; do not subtract one run from another.

• Qwen 4 Max — no product, no price, no date. Architecting around it is a bet, not a plan.

Screenshot of the OrcaRouter model page for Qwen3.8-Max showing the model id qwen/qwen3.8-max, a 1M-token context, text, image and video input, $2.00 input and $6.00 output per million tokens, and 5.00 s p50 time to first token.

How to decide this week

Three rules, no hedging.

• Need video input, or a model with a public benchmark trail you can audit: Qwen3.8-Max. It is callable today.

• Need downloadable weights, a readable licence and the lowest output price: Mistral Large 4, with the promotion's end date treated as unknown.

• Can wait: put Qwen 4 Max on a watchlist and do not build a migration plan until the five thresholds have published values. The CTO's "next week" has already slipped once; a stage date is not a release date.

If you want Qwen3.8-Max and a second family behind one key, that is the routing case. OrcaRouter carries Qwen3.8-Max at $2.00 and $6.00 with 0% markup on the provider list price, plus 200+ other models, automatic failover and a routing DSL for sending different task classes to different models. Mistral Large 4 is not in the catalogue — it is a Mistral API or self-host decision — so a two-vendor setup is exactly where a gateway earns its place.

Screenshot of the Mistral Large 4 model card on Hugging Face showing the Le Chonk title, 1 trillion total and 49 billion active parameters, a 1M-token context, text-and-image input, the NVFP4 checkpoint note and evaluation results including Terminal-Bench at 41.

The answer that does not change

Mistral Large 4 vs Qwen 4 Max is not a close call today, because only one of them is a product. The comparison that matters is Mistral Large 4 vs Qwen3.8-Max, and it splits on capability rather than quality: Mistral on price, cached input, weights and licence; Qwen on video input and public benchmark breadth. Qwen 4 Max could change that in a single launch, but it has not, and the five thresholds above are the checklist for the day it does.

Compared in this article1

Detected from this article · Benchmarks: Artificial Analysis · updated daily