MIDM K 3.0 Pro vs Mistral Large 4: A Bar of 26 and a Price of Seventy Cents

Mistral Large 4 scores 26 on the Artificial Analysis Intelligence Index. That is a modest number for a trillion-parameter flagship — three points above the median of 23 for open-weight models of its size, and twenty-nine behind GPT-6.1 Sol's 55 — and it is the single most useful fact in this comparison, because it is a ceiling that KT's MIDM K 3.0 Pro might actually clear. Every other cheap-model matchup on this blog runs an unmeasured newcomer against a model scoring in the forties or fifties, where the interesting question is how large the gap is. Here the measured side is beatable. MIDM K 3.0 Pro, released October 29, 2026 at $0.20 per million input tokens and $0.80 output, has no score at all; Mistral Large 4, previewed October 6 and open-weighted before the end of that month, has a score of 26, a measured $2.06 per index task, and 117 output tokens per second. For once, "we do not know" describes a genuinely open question rather than a polite way of saying "probably worse".

One shipped its weights, the other shipped a price list

The two releases ran within days of each other and pointed in opposite directions, which explains both products. Mistral put Mistral Large 4 into public preview on October 6 and had the checkpoints out before the month ended, carrying three further weeks of reinforcement learning, a 13-point Terminal-Bench 4.0 gain and a low reasoning level added, according to the company's own release post. KT opened Token Factory on October 29 with MIDM K 3.0 Pro as its headline domestic model. One company finished the month giving its flagship away; the other finished it selling access to a model nobody outside Korea can download.

Mistral Large 4 is a natively multimodal mixture-of-experts model: 1.05 trillion total parameters with 52 billion active, plus a 1.6-billion-parameter vision encoder, taking text and images in and returning text. Mistral's documentation puts the context window at 1M tokens and Artificial Analysis measures on that same 1M window, so unlike most vendor claims in this batch it is corroborated rather than merely asserted. The weights arrived under the Mistral Ecosystem Open-Weight License, which permits commercial use with restrictions — the part of "open" that most coverage skips.

MIDM K 3.0 Pro is the mirror image in almost every respect. 120 billion total parameters with 12 billion active, a hybrid of Mamba, mixture-of-experts and attention layers that is a serving-cost design rather than a capability one, text in and text out with no vision path at all, a 261,795-token context, and no weights. KT's differentiator is not the model but the delivery: domestic Korean infrastructure, KRW billing, and a platform built for enterprises that cannot send tokens abroad.

• Parameters — Mistral Large 4 1.05T total, 52B active, plus a 1.6B vision encoder; MIDM K 3.0 Pro 120B total, 12B active

• Modality — text and image in for Mistral Large 4; text only for MIDM K 3.0 Pro

• Context — 1M tokens for Mistral Large 4, confirmed both by Mistral's documentation and by the window Artificial Analysis tests; 261,795 documented for MIDM K 3.0 Pro

• Weights — published before the end of October 2026 under the Mistral Ecosystem Open-Weight License, against none

• List price — $0.70 in / $2.10 out per million for Mistral Large 4, against $0.20 / $0.80

• Price in force today — $0.14 / $0.56 for MIDM K 3.0 Pro while KT's discount runs to November 30, against Mistral's flat $0.70 / $2.10

• Independent index — 26, against none

• Cost per index task — $2.06, against not measured

What 26 actually means when you price it

A modest index score is not the same as a useless model, and Mistral Large 4's other measurements are the reason to say so. At 117 output tokens per second it is roughly twice as fast as GPT-6.1 Sol's measured 56 and more than three times Kimi K3.1 Pro's 32, and at 11.7 minutes per index task it beats the 13.1-minute median for its class outright. Its $2.06 per task is not cheap in absolute terms, but the number that matters more is the one underneath it: GPT-6.1 Sol costs $1.27 a task and scores 55, so Mistral Large 4 delivers roughly half the capability for about 1.6 times the price of the model that beats it. That is the awkward arithmetic of being a fast, cheap-per-token model that spends a lot of tokens.

Now put MIDM K 3.0 Pro's price against it. At $0.14 and $0.56 it is a fifth of Mistral Large 4's $0.70 input rate and a quarter of its $2.10 output rate, and after November 30 it settles at $0.20 and $0.80 — a little under a third of Mistral on input, a little over a third on output. If MIDM K 3.0 Pro scores anywhere near 26, it is the better buy for every workload that does not need vision or a million-token window. If it scores in the teens, the five-times-cheaper price is buying a genuinely weaker tool and the comparison collapses into "use the cheap one for tagging, the fast one for everything else".

The uncomfortable part is that nobody can currently distinguish those two worlds. KT has not published benchmark results for MIDM K 3.0 Pro. Artificial Analysis lists KT's earlier Mi:dm K 2.5 Pro but has never filled in an intelligence score for it, and carries no entry at all for this model. The model page is a specification sheet, a price table, a launch date, and a live latency chart — and that chart disagrees with itself, reporting 19.98 tokens per second in the provider row and a 62.93 average in the three-day median directly above it.

The licensing question, which is not the open-weights question

Mistral Large 4's weights being downloadable does not mean they are free to serve. The checkpoint is published, and Mistral's terms require a licence for commercial hosting above a threshold — this is the "open weights, gated hosting" pattern that has become standard for European labs, and it is why the practical choice for most teams is still the vendor's API. What the weights buy you is an exit: if Mistral's pricing moves or the EU data-residency rules change, the model does not disappear.

MIDM K 3.0 Pro offers no exit at all. It is a closed API on a single provider, and KT's own platform documentation is explicit that single-provider models do not receive its automatic routing. If the endpoint degrades, there is no second place to call and no checkpoint to fall back on. What KT substitutes for that is jurisdiction — the whole Token Factory pitch is domestic processing, Korean-language competence and KRW settlement — and for a Korean enterprise that is not a marketing line, it is the reason the model exists.

So the licence comparison is really a question about what kind of risk you are carrying. With Mistral Large 4 you carry capability risk (26 is not high) but not vendor risk. With MIDM K 3.0 Pro you carry capability risk in full and vendor risk on top, in exchange for a price and a jurisdiction that nothing else in this bracket offers.

Three workloads, three answers

Rather than a verdict, the useful thing is to name where each model is the obvious pick, because the price gap is wide enough that the answer flips more than once.

Document extraction from scanned Korean paperwork is MIDM K 3.0 Pro's case, and it is not close. The work is text-in, structured-out, the volumes are large, and the 261,795-token window is ample. At $0.56 per million output tokens the whole pipeline costs less than the storage it writes to, and Mistral Large 4's vision encoder is irrelevant because the OCR has already happened upstream.

Screenshot triage, chart reading, and anything else where the input is an image is Mistral Large 4's case, and MIDM K 3.0 Pro simply cannot enter it. A text-only model is not a cheaper option for a vision task; it is not an option.

Long-context work — a 400,000-token codebase, a full regulatory filing — splits on a number rather than a preference. Mistral Large 4 carries a 1M-token window that both its documentation and its measured configuration agree on, which is enough. MIDM K 3.0 Pro stops at 261,795 and will return an error, not a degraded answer.

The awkward fourth case is the one most teams actually have: high-volume text work under 200,000 tokens where quality matters and the volume makes the price gap decisive. That is where the missing score costs real money, because a team that guesses wrong either pays ten times too much for a year or ships a pipeline that quietly underperforms.

How to buy the answer instead of guessing

There is a cheap way to settle it. Two hundred requests from your own logs with known-good outputs, run through both models, scored by whatever rubric your team already uses. At MIDM K 3.0 Pro's rates that costs pennies; on Mistral Large 4, with its token appetite, a few dollars. Two days of work buys a number that is more relevant than any index because it is measured on your distribution.

For the Mistral side, the practical detail is that a rate card moving under you is where markup does the most damage: a promotion read as a list price is a bill that doubles without anyone shipping anything. Mistral Large 4 is reachable through Mistral's own API and several third-party platforms; it is not in our catalogue, and we do not route it. The discipline we apply to the 200-plus models we do route is the same one that makes a promotion visible rather than late: provider list prices passed through at 0% markup, so a vendor change lands the same day rather than waiting on a reseller to notice. MIDM K 3.0 Pro is not in our catalogue either; it is served by KT's Token Factory, and the honest advice for that side is to try it on KT's own key while the 30 per cent discount runs to November 30.

What the comparison comes down to, in the end, is that Mistral Large 4 is a known quantity with a modest score and a fast token stream, and MIDM K 3.0 Pro is a cheap unknown with an unusually low bar to clear. Twenty-six is not a high number. A 120-billion-parameter model with 12 billion active could plausibly reach it — Korean-language benchmarks are kind to models trained on Korean, and KT has been building this family since 2023. It could also fall well short, and there is currently no way for anyone outside KT to know. That uncertainty, not the price, is what a buyer should be pricing in.