MIDM K 3.0 Pro vs Qwen 4 Max: One Ships, One Is Still a Slide

There is a version of this comparison that has been written four times already on this blog, and it goes like this: Qwen 4 Max was announced at Alibaba's Yunqi conference on September 22, 2026 as the top of a four-tier family, has published no price, no context window, no weights and no API identifier since, and therefore cannot be compared with anything. All of that is still true in November. What is new is the other side of the pairing. MIDM K 3.0 Pro, which KT put on sale on October 29, 2026 at $0.20 per million input tokens and $0.80 output, went on sale with a published rate card on domestic infrastructure — and that makes the Qwen comparison a question about two national strategies rather than two models. One company shipped a deliberately small, deliberately cheap model and priced it. The other announced a family and a roadmap reaching ten trillion parameters, and has not yet put a price on a token of it.

What Qwen 4 Max is, on paper

Alibaba's September announcement laid out four tiers — Qwen 4 Max, Qwen 4 Flash, Qwen 4 Plus and Qwen 4 27B — described the family as training on a new-generation architecture, and projected a roadmap reaching five to ten trillion parameters in later generations. Since then the only publicly inspectable artefact has been Qwen3.8-Flash-Next, open-sourced in late August and labelled on its own card as a preview of the Qwen 4 architecture: 125 billion main parameters plus 51 billion of N-gram embeddings, roughly 6 billion active per step, a 262,000-token native context, and a training-cost claim about 90 per cent below Qwen3.7-Plus. In early November the open-source promise itself came into question, with reporting suggesting the 27B tier may ship behind a licence rather than as a plain release. As of November 11 there is no price for any Qwen 4 tier, no API identifier, no context window, and no benchmark.

What Alibaba does sell is Qwen3.8 Max, the model Qwen 4 Max will replace. That is a real product with a real rate card — $2.00 input and $6.00 output per million, a 1,000,000-token context — and an independent score. So the usable comparison is not MIDM K 3.0 Pro against a slide. It is MIDM K 3.0 Pro against Qwen3.8 Max, with Qwen 4 Max as the thing you might be waiting for.

Two small models with the same idea

Put MIDM K 3.0 Pro next to Qwen3.8 Max and the family resemblance is striking. Both are Asian flagships built on sparse architectures and optimised for cost per token rather than peak capability. Both are text-only at the API surface that matters here. Both are priced to undercut the American and European labs by a wide margin, and both are sold partly on the argument that your tokens should not leave the region.

• Parameters — MIDM K 3.0 Pro 120B total with 12B active, hybrid Mamba/MoE/attention; Qwen does not publish Qwen3.8 Max's parameter count

• List price — $0.20 in / $0.80 out per million, against $2.00 / $6.00

• Price in force today — $0.14 / $0.56 until November 30, 2026, against the list rate

• Cached input — $0.04, against $0.25 per million

• Context — 261,795 tokens with 235,615 of output, against 1,000,000

• Independent index — none published, against 34 on Artificial Analysis v5.0.1

• Measured cost per index task — not measured, against $10.02

• Output speed — KT's own dashboard reports both 19.98 and 62.93 tokens per second, against 51 measured

Read the price lines and MIDM K 3.0 Pro looks like the bargain of the year: a tenth of Qwen3.8 Max's input rate and a shade under a seventh of its output rate. Read the two measurement lines and the picture changes. Qwen3.8 Max scores 34 — below the median of 40 for its price band, and a genuinely modest result — but it costs $10.02 per index task, which is one of the worst cost-to-capability ratios in this bracket. It uses 150 million tokens to get through a suite whose median is 72.8 million, and it takes 55.8 minutes per task. The model is roughly eight times as expensive per task as GPT-6.1 Sol and scores twenty-one points lower.

That is the gap MIDM K 3.0 Pro is aiming at. It does not have to be good to beat Qwen3.8 Max on economics; it has to be not-terrible. Thirty-four is a low bar for a 120-billion-parameter model to clear, and that is the first time in this series of matchups that the unmeasured side has a plausible path to winning on both axes.

The three-currency problem

Both vendors sell the same political product, and it is worth being precise about what that product is. KT's Token Factory pitch is domestic processing on Korean infrastructure with KRW billing and usage consolidated onto a KT telecom bill. Alibaba's is sovereign capability — a Chinese model, Chinese weights, Chinese datacentres, and a roadmap that promises not to depend on anyone else. Neither is a technical claim; both are procurement claims, and they are aimed at buyers who cannot route a compliance questionnaire to a California address.

The difference is that one of them has a price today. A Korean enterprise can sign for MIDM K 3.0 Pro this week, at a known rate, with a discount that runs to November 30. A Chinese enterprise planning around Qwen 4 Max is planning around a date nobody has published, and in the meantime is buying Qwen3.8 Max at ten times the Korean model's input rate while it waits.

There is an irony in that worth naming. Qwen 4 Max is the more ambitious project by any measure — a trillion-parameter family with an open-weight tier and a stated roadmap to ten trillion. MIDM K 3.0 Pro is a 120-billion-parameter model with 12 billion active and no published benchmark, which by the standards of the frontier is a small model. Ambition and availability have come apart, and for the next quarter at least the buyer's decision is made by availability.

What you can actually run this week

If the choice is "MIDM K 3.0 Pro or Qwen", the honest answer in November is that you are choosing between a shipping model and a shipping model plus a promise. Qwen3.8 Max is real, it is callable, it carries a dated checkpoint — Qwen3.8 Max (0902) — and it has a published performance record. MIDM K 3.0 Pro is real, callable, cheaper by an order of magnitude, and has no performance record at all.

That combination is not a reason to avoid it. It is a reason to size the commitment correctly. The workloads where an unmeasured model is safe are the ones where a wrong answer is cheap and a slow answer is not: bulk classification in Korean, form-field extraction, first-pass summarisation where a human reads the output anyway. The workloads where it is not safe are the ones where the model's own judgement is the deliverable. Nothing about KT's specification sheet tells you which side of that line MIDM K 3.0 Pro falls on, and KT has not published a benchmark that would.

Nor does the architecture settle it. Sparse activation and NVFP4 quantization tell you the model is cheap to serve; they do not tell you it reasons well. The 235,615-token output cap is unusual and useful for long-form generation, and the 261,795-token input window is small enough that a lot of 2026 agentic work will not fit. Those are engineering facts, not quality claims.

A routing answer to a timing problem

The awkward part of waiting for Qwen 4 Max is that waiting has a cost, and the awkward part of adopting MIDM K 3.0 Pro today is that nobody has measured it. Both problems have the same shape: you want to route work to whichever model is currently the right one, without rewriting your integration each time the answer changes.

That is what the catalogue structure is for. Qwen3.8 Max is in the OrcaRouter catalogue at Alibaba's published $2.00 and $6.00 with nothing added, alongside Qwen3.8 Max (0902) as a dated sibling and Qwen3.8 Flash at $0.15 and $0.47 for the cheap tier — one key, one endpoint, one bill, and a model string is the only thing that changes when you swap. Qwen 4 Max is not in the catalogue and will not be until Alibaba ships something with an identifier and a price; that is the correct state for a keynote. MIDM K 3.0 Pro is also not in our catalogue — it is served by KT Token Factory, and that is where you would call it.

Where a router earns its place in this particular decision is the failover path. A single-provider model behind one key has no redundancy: if KT's endpoint degrades, or a request exceeds 261,795 tokens, there is no second place to send it. Keeping a measured model on the same key as the fallback turns "we are trialling a cheap Korean model" from a production risk into a routing rule. That is a smaller claim than it sounds, and it is the one that lets a team act on this page rather than filing it under "revisit when benchmarks land".

The date to watch, and the one nobody has given

The concrete dates are these. MIDM K 3.0 Pro's 30 per cent discount ends November 30, 2026, after which the model costs $0.20 and $0.80 — still the cheapest flagship-class rate on the board, still seven and a half times cheaper than Qwen3.8 Max on output. Qwen 4 Max has no date. Alibaba has said the family is in training, has published a roadmap, and has not published a release window; the most that can be said is that rumour has pointed at a launch repeatedly and the launch has not happened.

Which means the decision this month is not between two models. It is between a model and a wait. MIDM K 3.0 Pro, for all that it arrives without a benchmark, arrives: a price, a context window, a quantization scheme, a latency dashboard that contradicts itself, and an endpoint you can call today. Qwen 4 Max arrives as four names on a slide. When the slide becomes a product the comparison will be worth writing properly — and it will probably be worth writing about Qwen 4 Flash and Plus too, since those are the tiers that will meet this price band. Until then, the useful thing to do with Qwen 4 Max is exactly what the Qwen side has done with MIDM K 3.0 Pro: note that it exists, and wait for a number.