MIDM K 3.0 Pro vs Kimi K3.1 Pro: Two Ways to Spend a Million Tokens

Moonshot and KT made opposite bets with the same year of engineering. Kimi K3.1 Pro, released October 23, 2026 with the checkpoint following on Hugging Face on November 2, is 2.8 trillion parameters that you can download as well as rent, scores 46 on the Artificial Analysis Intelligence Index, and costs $2.00 per million input tokens against $10.00 per million output. MIDM K 3.0 Pro, released October 29, 2026 on KT's Token Factory, is 120 billion parameters with 12 billion active, arrives with no weights, no published benchmark and no independent score, and costs $0.20 and $0.80 — $0.14 and $0.56 while KT's launch discount runs to November 30. One of these models is a capability purchase and one is a cost purchase. Which one you want depends entirely on which of those two words describes your workload, and the rest of this page is about how to tell.

What each vendor decided to optimise

Kimi K3.1 Pro is a post-trained descendant of Kimi K3 with substantially more reinforcement-learning compute behind it, and Moonshot's own account is that the gains come from longer, more deliberate reasoning rather than a bigger base. That shows up in the measurements the way you would expect: it is slow, at 32 output tokens per second and 37.1 minutes per Intelligence Index task, and it is verbose, emitting 113 million tokens across the suite against a 99.3-million median for open-weight models of similar size. Its cost per index task is $3.68. It is, in other words, a model that buys quality with tokens — and it can afford to, because it is also the model you can host yourself if the API bill becomes the problem.

MIDM K 3.0 Pro goes the other direction entirely. The architecture is built to spend as little as possible per token: Mamba layers mixed with mixture-of-experts and attention, 12 billion active parameters out of 120 billion total, MoE experts quantized to NVFP4 and the Mamba projections plus shared expert held at FP8. That is a serving-cost design, not a capability design, and KT says so by publishing it on the model page along with the quantization scheme. The context window is 261,795 tokens with a maximum output of 235,615 — an unusual ratio that suggests the model was sized for document-in, document-out work rather than for a chat turn.

• Parameters — Kimi K3.1 Pro 2.8 trillion total, 104 billion active, against MIDM K 3.0 Pro 120 billion total with 12 billion active

• Weights — Kimi K3.1 Pro published on Hugging Face on November 2, 2026, against MIDM K 3.0 Pro undisclosed, API only

• Context window — 1,048,576 tokens against 261,795, with 235,615 tokens of output

• List price — $2.00 in / $10.00 out per million, against $0.20 / $0.80, currently $0.14 / $0.56

• Cached input — $0.20 against $0.04

• Independent index — 46, against none published

• Measured cost per index task — $3.68, against not measured

• Output speed — 32 tokens per second as measured by Artificial Analysis, against KT's own contested dashboard

The one number on MIDM that is not a price

KT publishes latency and throughput on its own model page, and the page contradicts itself. The provider row reports 19.98 output tokens per second and 0.67 seconds end-to-end for the single KT endpoint. The three-day median chart above it reports 62.93 tokens per second on average, 0.27 seconds to first token, and 1.47 seconds end-to-end. Both describe the same endpoint over overlapping windows. Neither is a lie; the provider row looks like a snapshot and the chart looks like a rolling median, but a reader has no way to tell which figure to plan against, and a factor of three is not a rounding difference.

Against Kimi K3.1 Pro's measured 32 tokens per second, even the pessimistic MIDM reading is in the same band, and the optimistic one is twice as fast. That is the first point at which the cheap model looks adequate rather than merely cheap. It is also the only performance claim about it that is not a price, which tells you how thin the evidence is.

Why "cheap" and "open" are different arguments

The instinct when comparing these two is to reach for cost per million tokens and stop. That would miss the thing that actually separates them.

Kimi K3.1 Pro's open weights are a hedge against its own price. A team that commits to it can, in principle, stop paying $2.00 and $10.00 and start paying for GPUs — and at 2.8 trillion parameters that is a serious capital question, not a weekend project. The point is that the option exists, and Moonshot shipping the weights within ten days rather than merely promising them is what makes the option real. The licence terms and the practical serving cost are a separate article; the existence of a checkpoint is not.

MIDM K 3.0 Pro has no such hedge. It is a closed API on one provider, and KT's own platform documentation states that models served by a single provider do not receive its automatic routing. If KT's endpoint degrades, there is no second place to call. What KT offers instead is a different kind of hedge: domestic hosting. The Token Factory platform is explicitly positioned on Korean infrastructure, KRW billing and data staying in Korea, which for a Korean enterprise with a procurement committee is a requirement that neither a Chinese open-weight model nor an American API satisfies.

That is the real matchup. Kimi K3.1 Pro is for teams whose constraint is capability or independence. MIDM K 3.0 Pro is for teams whose constraint is jurisdictional or budgetary. They are not competing for the same buyer, which is why a straight benchmark comparison between them would be misleading even if one existed.

What a million tokens of your own work costs on each

Abstract rates hide the decision, so put a shape on it. A legal-review pipeline that reads a 200,000-token contract and produces a 4,000-token summary, run two hundred times a month, is 40 million input and 800,000 output tokens.

On Kimi K3.1 Pro that is $80 of input and $8 of output — $88 a month, and the long context is comfortably inside its window. On MIDM K 3.0 Pro it does not run at all, because 200,000 tokens of contract plus a 4,000-token answer is close enough to the 261,795-token ceiling that a longer document breaks it. Drop the contract to 120,000 tokens and MIDM K 3.0 Pro handles it for 24 million input tokens at $0.14 — $3.36 — plus 800,000 output at $0.56, or $0.45. Under four dollars against Kimi's $88 — about $5.40 once the discount expires on November 30.

Roughly twenty-three times cheaper, on a task the cheap model can barely fit. That is the entire value proposition in one line, and it is also the entire risk: nothing in that arithmetic says the summary is any good. Kimi K3.1 Pro's 46 and its $3.68 per index task come with a measured quality level attached. MIDM K 3.0 Pro's four dollars come with no quality claim at all beyond KT's own specification sheet.

Running the comparison on your own documents

Both models are callable today and both take the same request shape, which makes a side-by-side cheap to arrange. Kimi K3.1 Pro is in the OrcaRouter catalogue at Moonshot's published $2.00 and $10.00 with nothing added on top — the vendor's list price passed through, so a Moonshot price change reaches your key the same day rather than at the next billing cycle — and it sits behind the same key and endpoint as the other 200-plus models we route, which means an A/B against MIDM K 3.0 Pro is a model-string change on the Kimi side and a second base URL on the KT side. MIDM K 3.0 Pro itself we do not serve; KT's Token Factory is where it lives.

The useful setup for a team that wants the cheap model but cannot afford to be wrong is not a choice between them. It is MIDM K 3.0 Pro on the bulk of the traffic with Kimi K3.1 Pro behind it, and a routing rule that sends anything over 250,000 tokens straight to Kimi because the MIDM window will reject it anyway. Failover covers the other failure mode: a single-provider endpoint that degrades resolves to a measurable model instead of to an error page. Neither of those is a substitute for knowing which model is better. Both are ways to keep working while you find out.

What would settle it is a public evaluation, and the asymmetry in who has one is the story here. Kimi K3.1 Pro arrived with a score, a speed measurement, a cost-per-task figure and a downloadable checkpoint on day twelve. MIDM K 3.0 Pro arrived with a rate card, a specification sheet and a latency chart that disagrees with itself. Both are legitimate releases; only one of them lets a buyer do arithmetic before signing. If you are choosing this month, choose Kimi K3.1 Pro when the answer has to be right and MIDM K 3.0 Pro when the answer has to be cheap and Korean — and if you genuinely cannot tell which case you are in, that is the argument for running two hundred of your own documents through both before the discount ends on November 30.