Mistral Large 4 vs Kimi K3.1 Pro is the defining open-weights matchup of this ten-day window — and the two trillion-parameter-class contenders bet on opposite strategies. Mistral shipped first: Mistral Large 4 (“Le Chonk”) went live on Hugging Face on October 29, weights downloadable, license text published. Moonshot went the other way: Kimi K3.1 Pro posted the strongest agentic coding numbers Artificial Analysis has published this season — and, as of this writing (Nov 3), had shipped zero weight files to back them up.
That asymmetry is the entire comparison. One model you can hold, the other you can only rent; one tops the vendor’s own charts, the other tops the independent board. So this head-to-head keeps every number labeled with who measured it, flags where vendor and independent accounts disagree, and ends with a decision guide that survives contact with those disagreements.
(We’ve covered both models in depth solo — see our Mistral Large 4 open-weights deep dive, the Kimi K3.1 Pro open-weights situation, and the Mistral 3 deep dive (675B MoE, Apache 2.0). The cyber-disclosure angle stays there; this is capability, pricing, architecture, and the license.)
Mistral Large 4 vs Kimi K3.1 Pro at a glance
| Mistral Large 4 | Kimi K3.1 Pro | |
|---|---|---|
| Total parameters | ~1T (Mistral’s docs list 1.05T; Artificial Analysis lists 1T) | 2.8T |
| Active parameters | 49B per token (52B including embeddings and output layers, per Mistral’s model card) | 104B (per Artificial Analysis) |
| Context window | 1M tokens | 1M tokens |
| Modalities | Text + image in, text out | Text + image in, text out |
| Released | Preview Oct 6; weights Oct 29, 2026 | Oct 23, 2026 (API/apps); weights promised by Nov 4 |
| Weights today | LIVE — bf16/Fp8, NVFP4, GGUF | Not published (verified below) |
| License | Modified MIT + MaaS carve-out (text public) | TBD until weights drop |
| API list price (per 1M) | $0.70 in / $0.07 cached in / $2.10 out | $2 in / $10 out / $0.20 cached in |
One param-number conflict resolves here: “49B vs 52B active” isn’t an error. Mistral’s HF card states both — 49B active per token, 52B counting the embedding and output layers. Third-party write-ups pick whichever their source used — a small taste of why every figure below wears a source label.
What each model actually is
Mistral Large 4 is a natively multimodal Mixture-of-Experts generalist, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European datacenters. Per the company’s model card: 128 routed experts with 4 active per token plus 1 shared expert, a 1M-token context, a Pixtral-style vision encoder, natively fluent in 160+ languages (including every official EU language), and a reasoning-effort dial (low/medium/high, default high). An unusual detail: the checkpoint ships with an embedded DSpark speculative decoding draft head — three dense MLA draft layers in BF16 proposing three-token blocks — which Mistral measured at up to 1.9× higher throughput without changing the model’s outputs.
Kimi K3.1 Pro is not a new pretrain. Moonshot’s blog describes it as “a post-trained version of Kimi K3, with substantially larger post-training compute,” using reinforcement learning jointly across “hundreds of downstream domains” with re-engineering of the RL framework, long-context inference, and sandbox infrastructure. It keeps Kimi K3’s 2.8T scale plus a 1M context and text+image input, and adds a sibling — K3.1 Pro Highspeed — that Moonshot says runs 2–3× faster at equivalent intelligence.
The scale gap matters less than where the parameters live. Kimi spends ~104B active of 2.8T (~3.7% per token); Mistral spends 49–52B of ~1T (~5%). Kimi’s per-token compute budget is roughly double Mistral’s — which shows up in throughput, memory, and every hosting quote you’ll get.
Who each is for
Mistral Large 4 is aimed at organizations that need to own the model: sovereign-AI buyers, EU-language workloads, vision-heavy document work (Mistral says it surpasses frontier closed models on visual grounding), and anyone whose roadmap says “self-hosted” in ink.
Kimi K3.1 Pro is aimed at people chasing the agentic coding ceiling: repository maintenance, terminal workflows, financial-document agents, 3D creation. Today you use it through Moonshot’s apps and API at production rates; API customers get Zero Data Retention. Moonshot explicitly frames the upcoming release as available “in a non-discriminatory way” — but that’s a promise, not a file. One data-location nuance for the sovereignty-minded reader: per Moonshot’s international privacy policy (as relayed by VentureBeat’s launch coverage), the Kimi API processes data on servers in Singapore — Zero Data Retention is not the same thing as on-region or on-prem. One more trust item, stated with its own caveats intact: VentureBeat’s launch-week coverage documented undisclosed paid-influencer allegations around the K3.1 Pro rollout — including a promotional brief allegedly directing creators “DO NOT put on the paid partnership label” — while explicitly noting the document is unauthenticated and that no Moonshot authorization is established. These are allegations, not findings; we note them because a post premised on “who measured it” can’t skip the week’s credibility fight.
Kimi K3.1 Pro vs Mistral Large 4: benchmarks, with every number source-labeled
Here is the full set, per-source, never averaged. (Read the last column first.)
| Benchmark | Mistral Large 4 | Kimi K3.1 Pro | Who measured / stated |
|---|---|---|---|
| AA Coding Agent Index | 48 (ML4) — “effectively tied” with GLM-5.3, using 47% fewer tokens per task — but AA’s Oct 29 assessment (relayed by VentureBeat) lists GLM-5.3 at 49 and ML4 at 48: second, not tied | Not found in sources we could render | Mistral’s model card, citing Artificial Analysis; AA’s assessment relayed by VentureBeat |
| (Mistral preview bundle) Coding Agent Index | 49.8, with DeepSWE v1.1 61.7%, SWE-Atlas-QnA 59.4%, and Terminal-Bench 4.0 at 28.3% | — | Mistral’s preview post — whose archived Oct 6 footnote credits the component scores to “the numbers reported by the ArtificialAnalysis coding index” (AA measured privately ahead of the harness’s public launch) and Vibe Code Bench to vals.ai; the footnote was later removed from the live page |
| Terminal-Bench 4.0 | 41% (vendor, release build, +13pp over the 28.3% self-reported preview, charted at 41.4% — and itself an AA-measured run in the Mistral Vibe agent: AA’s Coding Agent Index lists 41.41 at High effort / 26.26 at Medium per AA page data read by The AI Realist) vs ~25% in a minimal agent (AA’s mini-swe-agent evaluation; the integer 25% is what VentureBeat’s Oct 29 ML4 coverage relays, while the 24.7% decimal comes from readings of AA’s Terminal-Bench 4.0 evaluation page by MegaBrain/The AI Realist on Nov 1) | 65.7% (vendor, max effort, Claude-Code harness) vs 67% (AA-measured) — highest AA has measured on the board, ahead of Claude Sonnet 5.5’s 64; second board Vals.ai: 63.64% (#3 overall) | Mistral news post; Moonshot blog; AA launch article; AA’s TB4.0 evaluation page (25% via VentureBeat’s relay; 24.7% via MegaBrain/The AI Realist); AA’s coding-agent page data via The AI Realist; vals.ai |
| AA Intelligence Index (v5) | 26 for the release (“the highest among western open-weight models” per Mistral’s card); 38 for the October preview per AA’s chart — but 38 is the debut-era pre-v5 scale; AA’s v5 rescore of the preview is 25, so the like-for-like v5 comparison is 25 → 26 (see caveat below) | 55 — 4th overall, behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), Muse Wonder Full (56) | Mistral’s card (AA v5); tess.im citing AA; AA launch article |
| GDPval-AA (Elo) | No figure we could render | 1,846 — 3rd, behind Claude Opus 5.5 (1,866) and Muse Wonder Full (1,857); +312 Elo over Kimi K3 | AA launch article |
| AutomationBench-AA | 59.9% (Mistral-reported, preview) | 73% (AA; up from 58 for K3) | Tess citing Mistral; AA |
| SWE-Marathon v1.1 | Inside Mistral’s preview bundle (61.7% was DeepSWE, not Marathon) | 51.7 (Moonshot-reported table) | Moonshot’s comparison table (via VentureBeat) |
| AA Cyber Index | 51 (release, cited by Mistral); 50 for the preview (AA) | 64 (release, per AA; up from 41) | Mistral’s news post; tess.im; AA |
A caveat on granularity: Mistral’s card gives the Coding Agent Index as a whole percent (48) and quotes AA calling the matchup “effectively tied.” More precise decimals — [UNVERIFIED: rendered AA Coding Agent Index values at decimal precision (Mistral Large 4 vs GLM-5.3), e.g. 48.3 vs 48.9] — still can’t be pulled from an HTML-renderable AA page. But the rank order we do have: AA’s October 29 assessment (relayed by VentureBeat) lists GLM-5.3 at 49 and Mistral Large 4 at 48 — second, not tied. Read “effectively tied” as the vendor’s framing of a one-point AA gap.
And a scale caveat on the Intelligence Index row above: the preview’s 38 is a debut-era score on the pre-v5 scale, while the release’s 26 is on the current v5 scale. AA’s v5 rescore of the preview is 25 — on like-for-like v5, that’s 25 → 26 (per AA’s assessment as relayed by VentureBeat’s ML4 coverage, which relays AA’s assessment as: “the earlier 38 and today’s 26 belong to different scales; the relevant v5 comparison is 25 to 26”). Juxtaposing 38 and 26 in one breath would be exactly the cross-version comparison this post otherwise refuses to make.
The release’s independent Terminal-Bench 4.0 score, meanwhile, does exist — but it needs a harness label before it can be compared to anything. AA’s October 29 assessment of the final build in a minimal (mini-swe-agent) harness puts it at ~25% — the integer 25% is what VentureBeat’s ML4 coverage relays, while the 24.7% decimal comes from readings of AA’s Terminal-Bench 4.0 evaluation page by MegaBrain/The AI Realist on Nov 1. And the vendor-side 41% is not a purely vendor figure either — it is AA’s own coding-agent run of ML4 inside Mistral’s Vibe agent. The ~16 points between AA’s own two measurements are the gap section’s real story.
The vendor-vs-independent gap is not noise — it’s the story
Head-to-head posts usually reconcile this quietly. Don’t. Here’s the same benchmark, Terminal-Bench 4.0 (which we benchmarked the top coding agents on in depth), traced through three provenance chains — vendor, independent, and the collision between them:
- Kimi: the vendor reports 65.7%, and Artificial Analysis independently measures 67% — the highest score AA has put on the board, above Claude Sonnet 5.5’s 64. A second independent board agrees in kind: vals.ai puts K3.1 Pro at 63.64% (#3 overall) — lower than AA’s 67, but still top-tier. Vendor and independent agree, and the independent number is the bigger one. That’s about as clean as benchmark cross-checking gets.
- Mistral: the vendor’s preview bundle put TB4.0 at 28.3%, and the release announcement claims 41% — a 13-point improvement “taking the lead among open-weight models” — with the same announcement’s chart showing the preview’s self-reported bar at 28% and the release at 41.4%. But “the independent number for the release build” is not one number — it’s two, and they differ by harness. In AA’s Coding Agent Index, AA itself ran the final ML4 inside Mistral Vibe, Mistral’s own coding agent (AA runs each model in its maker’s agent or a neutral one: Opencode for GLM-5.3, Claude Code for Claude, Mistral Vibe for Large 4) and logged 41.41 at High effort, 26.26 at Medium (AA page data as read by The AI Realist, Nov 1). In the minimal mini-swe-agent evaluation, the same build scores ~25% — the integer 25% relayed by VentureBeat’s Oct 29 coverage, the 24.7% decimal read on AA’s Terminal-Bench 4.0 evaluation page by MegaBrain/The AI Realist — and vals.ai’s second board, running the same TB4.0 with the Mini-SWE-agent harness, shows 26.26% ±5.13 (#8 of 22 open-weight models), with the preview at 27.78% ±1.82: no gain. So the ~16-point spread below the vendor’s charted 41.4% is not one board contradicting another — it is an agent-harness premium: per The AI Realist’s reading of AA’s page data, Large 4’s premium of +16 points is the largest of the 13 models AA compares, while the other twelve sit between −4 and +7, median 0. The same vendor-vs-AA split recurs on FrontierCode 2.0: Mistral’s chart shows 19.7% taking the open-weight lead, ahead of DeepSeek’s 19.3%, but AA’s evaluation puts ML4 at 15 — behind DeepSeek’s 29 and GLM’s 22.
- And the leaderboard collision: Mistral’s announcement says its release takes the open-weight TB4.0 lead at 41%; Artificial Analysis’s K3.1 Pro launch article reports an open-weight model at 67 on the same benchmark name. Both statements appeared within a week of each other.
That last bullet is the most important passage in this post — and so is the number just added to the Mistral bullet. A benchmark score without a scorer, a date, and an effort setting is marketing data. Moonshot ran its K3.1 Pro numbers at max effort with the Claude-Code harness, and the caution about reproducing them comes — per VentureBeat’s launch coverage — from Moonshot’s own API parameter reference, not from any AA document: don’t assume lower-effort API calls reproduce max-effort launch scores. Moonshot’s API defaults to max, and changing effort mid-conversation invalidates prefix-cache hits, which also changes the bill.
Pricing: Mistral wins the token bill, decisively
| Mistral Large 4 | Kimi K3.1 Pro | Kimi K3.1 Pro Highspeed | |
|---|---|---|---|
| Input / 1M tokens | $0.70 | $2.00 | $4.00 |
| Cached input / 1M | $0.07 | $0.20 | $0.20 |
| Output / 1M tokens | $2.10 | $10.00 | $20.00 |
Per the vendors’ published rates: Kimi’s standard tier charges about 2.9× more per input token and ~4.8× more per output token than Mistral. The Highspeed tier, pitched as 2–3× the speed at equivalent intelligence, doubles the token bill again, matching Claude Opus 5.5’s list rates. For an output-heavy workload — which agent work usually is — Mistral is several times cheaper, with a standing cached rate (7¢) against Kimi’s 20¢.
Kimi’s side has structural sweeteners: Moonshot’s platform bills K3-series batch inference at 60% of standard rates, enables context caching automatically (five-minute TTL by default), and per AA costs roughly $1.98 per Intelligence Index task. For Mistral, AA priced the debut preview at $1.13 per index task on the pre-release list price — a stale basis — but AA’s later like-for-like measurement of the release (per VentureBeat’s relay of the October 29 assessment) puts the shipped model at $2.06 per Intelligence Index task, within pennies of K3.1 Pro’s $1.98. On AA’s cost-per-task board, then, the two are nearly even — the decisive Mistral price advantage is on the token rates in the table above. Don’t read the $1.13 debut figure or the preview’s $5.54 release-basis price against Kimi’s $1.98; $2.06 vs $1.98 is the like-for-like pair. And one final asymmetry: Mistral’s list price is also the fallback for self-hosters — you can stop paying it entirely.
Architecture: three trillion, two philosophies
Mistral’s MoE is deliberately granular — 128 routed experts, 4 firing per token, plus 1 shared expert — keeping the active-parameter budget low (49B) while the 1T total provides breadth. Add the 1.6B vision encoder and the DSpark head, and the design is serving-engineering-first: cheap and fast to run for an operator who owns the hardware. Mistral ships official vLLM (recommended) and SGLang configs, transformers support, GGUF builds, and fine-tuning recipes from Axolotl, Unsloth, and TRL.
Kimi’s approach inverts it: 2.8T total, 104B active — more than twice Mistral’s per-token compute — plus a Highspeed variant that buys back latency with raw speed. Moonshot’s engineering effort went into the training layer: reworking the RL stack, long-context inference, and rollout scheduling to scale joint RL across hundreds of domains. Hosting Kimi means a bigger machine footprint than Mistral’s — and until the weights drop, that planning stays hypothetical.
Weights and license, as of November 3
Mistral Large 4: live, with the license text published. The Mistral Ecosystem Open-Weight License is a modified MIT license, free for commercial and non-commercial use, with one carved-out clause:
- Section 2: If the Licensee or any affiliate operates a Model-as-a-Service business, it must enter a separate commercial license with Mistral before commercial use. “Model as a Service” is defined as giving third parties access to the Model or derivatives for inference or post-training in a way that lets them meaningfully control inputs, parameters, or training data — and it explicitly excludes (a) end-user products with model capabilities embedded in specific features or harnesses, and (b) mere relaying of requests to models hosted by others.
- Section 3: The Section 2 requirement doesn’t apply to internal use, or to use through Mistral’s official products.
In plain terms: embed it in your product or relay it onward, fine. Stand up a hosted-API business with the weights, you owe a commercial license conversation.
Kimi K3.1 Pro: not out. This section carries the post’s date for a reason. We verified live against the moonshotai Hugging Face org while writing: the newest repository is Kimi-K3 (2.8T), updated September 2 — no K3.1 Pro repo exists as of Nov 3. Moonshot’s own blog states: “Full model weights will be released by Nov 4, 2026.” So K3.1 Pro is an open-promised-weights model: sold through a paid API while the download is a dated promise. Artificial Analysis still lists weights as not published.
That makes license self-hosting a half-side comparison by construction. Mistral’s license text is public and we can dissect clause by clause; Kimi’s license text does not exist yet to read, so the “what license will open weights come with?” column is TBD, not assumed. The same-day caveat, stated plainly: if Moonshot published weights on Nov 4 as promised, this section — and the TBD license column above — is stale by a day; read the license file before drawing deployment conclusions. Our companion post on the K3.1 Pro open-weights situation will carry the license autopsy when the text lands.
The caveated verdict
If forced to one line each:
- Agentic coding ceiling, today: Kimi K3.1 Pro — an AA-measured 67 on Terminal-Bench 4.0 (the highest AA has measured), 3rd on GDPval-AA at 1,846 Elo, 73 on AutomationBench, 4th on the Intelligence Index. But that ceiling is API-only until Nov 4’s promised drop, and as the gap analysis above shows, how much of it you capture depends on harness and effort settings.
- Price and shipping certainty: Mistral Large 4 — ~2.9× cheaper input and ~4.8× cheaper output at list, weights you can download today, a license text you can read end-to-end, and a hardened serving stack (vLLM/SGLang/GGUF, fine-tuning recipes).
- Self-hosting: Mistral, actionable now — 49–52B active on a ~1T model is a materially smaller budget than K3.1’s 104B on 2.8T. Kimi becomes self-hostable only after the Nov 4 weights and when its license text is read, not before.
No single-number verdict survives this matchup — but the two cross-checks resolve differently, and for different reasons. Kimi’s is a clean validation, with one asterisk: vendor 65.7 vs AA 67 (vals.ai’s second board: 63.64), with a one-clause caveat — Moonshot’s own table puts 65.7 narrowly behind Claude Opus 5.5’s 66.4, while AA places the model ahead of Sonnet 5.5’s 64, so the ordering against the best proprietary models depends on whose table you trust. Mistral’s is not a refutation, despite first appearances: both headline TB4.0 numbers for the release are AA’s own measurements in different harnesses. AA ran ML4 in Mistral’s Vibe agent and logged 41.41 at High effort; AA ran it in the minimal mini-swe-agent and scored it 24.7/25 (the integer 25% as relayed by VentureBeat; the decimal as rendered on AA’s TB4.0 evaluation page per the readings cited above); vals.ai’s 26.26 ±5.13 agrees with the minimal-harness reading, and neither minimal-agent board shows a gain over the preview. In a purely generic agent, GLM-5.3 leads ML4 by roughly 17 points (42 vs 25). The ~16-point spread is a harness-and-effort finding, not a vendor-refuted-by-board finding: an agent premium of +16 points, the largest of the 13 models AA compares (the other twelve range from −4 to +7). This post’s own caution — which agent ran the test? — applies to its own comparison: same benchmark, same evaluator, 16 points apart depending on the harness. There is a genuine vendor-vs-AA contradiction in Mistral’s release, but it lives on FrontierCode 2.0 (19.7 vs 15) rather than in the Vibe-vs-minimal spread. And the model whose independent number flattered its vendor is still the one you can only rent until Nov 4. So the honest conclusion is unchanged: re-benchmark both on your workload — in your harness, at your effort setting — before believing any of these numbers.
Mistral Large 4 vs Kimi K3.1 Pro: which one should you pick?
- “I need an on-prem, vision-capable, EU-languages-capable model this week” → Mistral Large 4. It’s the only one of the two that can answer this, at any price.
- “I’m chasing the agentic coding lead and can use the API while I wait” → Kimi K3.1 Pro. Test it at the effort level you’ll actually run.
- “My workload is output-heavy and margin-sensitive” → Mistral Large 4. $2.10 vs $10 per output-token settles this before any benchmark table does.
- “Self-hosting plus license clarity is a hard requirement” → Mistral today. Kimi only after Nov 4’s weights arrive and the license text is published and read.
- “I just want the single strongest open model, full stop” → wait until Nov 4: if the weights ship, Kimi K3.1 Pro — independent-evaluated, 2.8T, license permitting — will be the strongest answer to that question in open weights. If they don’t ship, the promise was the product, and Mistral Large 4 remains the frontier you can actually download.
FAQ: Mistral Large 4 vs Kimi K3.1 Pro
Q: Is Mistral Large 4 or Kimi K3.1 Pro better for coding agents?
On the independent numbers, Kimi K3.1 Pro leads agentic coding — 67 on Terminal-Bench 4.0 as measured by Artificial Analysis, the highest score on that board — Mistral Large 4’s 41% — charted at 41.4% by the vendor — is itself an AA-measured run inside Mistral’s Vibe agent; the same build scores ~25% (24.7% on AA’s live page) in a minimal mini-swe-agent, so much of that 16-point spread is harness, not model. Effort settings move these scores materially on both sides; re-benchmark in your own harness at your intended effort level.
Q: Which is cheaper, Mistral Large 4 or Kimi K3.1 Pro?
Mistral Large 4, by a wide margin: $0.70/$2.10 per million input/output tokens versus Kimi K3.1 Pro’s $2.00/$10.00 — roughly 2.9× cheaper input and ~4.8× cheaper output.
Q: Is Kimi K3.1 Pro actually open-weights yet?
Not as of November 3, 2026. Moonshot has committed to releasing full model weights by Nov 4; the moonshotai Hugging Face org had no K3.1 Pro repository when we checked. Mistral Large 4’s weights, by contrast, have been downloadable since October 29.
Bottom line: Mistral Large 4 vs Kimi K3.1 Pro boils down to certainty vs ceiling. Mistral Large 4 is the frontier you can download, self-host, and afford today; Kimi K3.1 Pro is the agentic-coding leaderboard you can only rent until Nov 4 — and every number in between depends on who did the measuring.
(Back to the Mistral Large 4 open-weights deep dive, the Kimi K3.1 Pro open-weights status and license watch, and the Mistral 3 deep dive — the 675B MoE underpinnings.)
References and further reading
- Mistral Large 4 — Hugging Face model card (Mistral-Large-4-1T-A52B)
- Mistral Ecosystem Open-Weight License — raw license text
- Mistral Large 4 1T NVFP4 checkpoint
- ggml-org/Mistral-Large-4-1T-A52B — GGUF builds
- unsloth/Mistral-Large-4-1T-A52B-GGUF
- Mistral AI — Mistral Large 4 preview announcement
- Mistral AI — Mistral Large 4 open-weights release
- Moonshot AI — Kimi K3.1 Pro launch post
- Moonshot AI — developer platform, pricing and API parameter reference
- Moonshot AI on Hugging Face (org page — no K3.1 Pro repo as of Nov 3)
- Artificial Analysis — model intelligence, coding agent index and evaluations
- vals.ai — independent coding-benchmark boards
- tess.im — coverage citing Artificial Analysis data
- VentureBeat — Kimi K3.1 Pro and Mistral Large 4 launch coverage
- vLLM — GitHub
- SGLang — GitHub
- Hugging Face Transformers documentation
- Axolotl — fine-tuning recipes, GitHub
- Unsloth — fine-tuning recipes, GitHub
- TRL — Hugging Face, GitHub
Please let us know if you enjoyed this blog post. Share it with others to spread the knowledge! If you believe any images in this post infringe your copyright, please contact us promptly so we can remove them.