MIDM K 3.0 Pro: KT's 120B Hybrid Lands With a Price and No Benchmarks

Mi:dm K 3.0 Pro — spelled MIDM K 3.0 Pro in most model catalogues, and listed internationally as KT: Mi:dm K 3.0 Pro — is the first Korean telecom model to arrive on an international price list with a frontier-shaped spec sheet and no public evaluation behind it. KT is serving it from its own Token Factory platform with a release date of 29 October 2026, and the international listing is dated 5 November 2026. It is a 120-billion-parameter mixture of experts that activates 12 billion of those parameters per token, it accepts 261,795 tokens of context, and it costs $0.20 per million input tokens and $0.80 per million output tokens. What it does not have, as of 11 November 2026, is a single published benchmark score from anyone — not from KT, not from an independent evaluator, not from a rival lab's comparison table. That gap is not a footnote to this page. It is the main thing a buyer needs to price in.

Three dates, and which one is the release

KT's own materials give three different dates for the same model, and they are all correct in their own frame, which is why the "when did it come out" question has been so hard to answer.

• 15 October 2026 — Mi:dm K 3.0 is shown publicly for the first time at KT's AID-X Day in Seoul, alongside a next-generation voice agent, an internal agent platform called KT-Claw, and the company's prompt-compression and model-routing work. Korean trade press covering the event describe it as the group's "next-generation model," demonstrated rather than sold. Nothing is priced, and no model identifier exists yet.

• 29 October 2026 — the date KT's own Token Factory model page carries as 출시일, the release date. This is when the model became something a customer could send a request to and be billed for, on KT's own platform.

• 5 November 2026 — the date on the international catalogue listing, the point at which developers outside KT's own billing relationship could reach it. Every one of those catalogue entries reproduces the same one-sentence description — mixture of experts, 120B total, 12B active, hybrid Mamba/MoE/attention — which originates from KT and has not been independently verified.

The reason the distinction is worth three bullets rather than one sentence is that "released 29 October" and "available internationally 5 November" mean different things to a procurement team. The first is a vendor relationship. The second is a line item you can evaluate this week. Both are recent enough that the model has not accumulated any published third-party testing, so a reader arriving in November 2026 is early, not late.

What the architecture actually is

The hybrid design is the interesting part and the part most catalogue copy compresses into a single clause. KT's description is that Mi:dm K 3.0 Pro mixes Mamba layers, mixture-of-experts layers and ordinary attention layers in one stack. Each of those does a different job. Attention gives the model exact recall of arbitrary earlier positions, which is what long-document tasks and precise instruction-following depend on. Mamba gives it a linear-time path through long sequences, which is what keeps a 262K-token prompt from becoming quadratic in cost and latency. The mixture-of-experts layer is what makes a 120B model affordable to serve at all: only 12B parameters — a tenth of the network — run for any given token, so the compute bill tracks the active count rather than the total.

Putting all three in one stack is a real engineering position, not marketing. It is the same instinct behind several 2026 releases that have tried to escape the cost of pure attention at long context, and it is a harder thing to train than picking one. Whether KT executed it well is exactly the question no one has answered publicly.

KT also discloses the serving format, which is more than most labs publish: on its Token Factory page the model is described as being served from a checkpoint whose MoE experts are quantized to NVFP4, with the Mamba projections and the shared expert held at FP8. That is a serving decision, not a training one, and it has a practical consequence — the weights a customer is calling are not the weights KT trained, and any benchmark run against this endpoint measures the quantized artifact. For a 120B model at $0.20 per million input tokens, aggressive quantization is how the price is reached. It also sets an expectation: a 4-bit expert path is not where you look for exact reproduction of a full-precision reference.

The envelope, in the numbers KT publishes

• Parameters — 120B total, 12B active per token, per KT's own model card copy.

• Context — 261,795 tokens. That is 256K in round terms, and about a quarter of what the frontier models in this price band's neighbourhood offer.

• Maximum output — 235,615 tokens in a single response. This is unusually generous relative to the context window: the model can write almost as much as it can read.

• Modality — text in, text out. No image, audio or file input, and no multimodal output. Catalogue listings confirm text-only on both sides of the request.

• Reasoning — the model exposes a reasoning control and returns reasoning content separately from the answer, so a caller can suppress or display the thinking trace.

• Structured output — JSON and schema-constrained output are supported, which is the capability that matters most for anyone wiring the model into a pipeline.

• Tool calling — not listed among the model's supported request parameters on its catalogue entry, in contrast to most 2026 frontier models, which expose tools and tool_choice explicitly. If an agent loop needs function calling, this is the first thing to test rather than assume.

• Pricing — $0.20 per million input tokens, $0.80 per million output tokens, and $0.04 per million cached input tokens.

• Measured serving behaviour — roughly 196 output tokens per second and 0.62 seconds to first token at the median, with availability reported at effectively 100% over the trailing day on the single serving provider.

The price, and the discount that is not on the international listing

KT's own platform currently sells the model below its list price. The Token Factory page shows input at ₩190 per million tokens against a struck-through ₩271, and output at ₩759 against ₩1,085, which is a 30% promotional discount applied to the same $0.20 and $0.80 rates the international listing quotes. Cached input follows the same pattern: $0.03 at the promotional rate, $0.04 at list.

Two things follow. First, the effective price today is $0.14 input and $0.56 output if you buy from KT directly, and $0.20 / $0.80 if you reach the model through a third-party catalogue that passes the list rate through. Second, promotional pricing on a model this new is a launch instrument, not a rate card, and it will end without a great deal of notice. A cost model built on $0.14 should be re-run on $0.20 before it is used to justify anything.

Even at list, the position is stark. Twenty cents per million input tokens is roughly one two-hundredth of what the frontier charges — Claude Opus 5.5 is $4 and $20, Gemini 4 Argon is $2 and $10 — and it is cheaper than D​eepSeek's own flagship tier at $0.66 and $1.98 off-peak. Whether that is a bargain or a warning depends entirely on capability, and capability is the one number nobody has.

The measurement gap, stated precisely

Three things are commonly mistaken for benchmarks, and none of them is one.

The first is KT's own announcement. The AID-X Day materials and the model card describe architecture, quantization and context, and do not publish a table of scores against named evaluations. There is no vendor-reported Terminal-Bench number, no SWE-bench figure, no Korean-language benchmark result to argue about. That is unusual for a 2026 launch — most labs ship a table even when the table is self-serving — and it means there is nothing to discount.

The second is Artificial Analysis, the independent evaluator most of this blog's comparisons lean on. As of 11 November 2026 it has no page for the model: the model path returns a 404 and the model does not appear in its index. That is the cleanest signal available that the model has not been independently run through a standard suite.

The third is the aggregator composite. At least one model-comparison site publishes a 40-out-of-100 "overall" score and a coding rank of #185 of 445 for Mi:dm K 3.0 Pro. Those numbers are not benchmark results. They are a weighted blend of price, recency, context length, output capacity and a capability signal, and on that site price alone carries 25% of the weight and context 15%. A model this cheap and this new scores well on two of the five inputs by construction. Read as a spec-sheet summary it is fine; read as a capability measurement it is circular.

So the honest state of play is: one architecture description from the vendor, one price, a context window, and nothing that says whether the model is good. Anyone publishing a comparison table with Mi:dm K 3.0 Pro's benchmark scores in it this week either invented them or copied someone who did.

Where it actually fits in the market

The nearest thing to a positioning statement comes from KT's own history rather than from this model's card. KT has run the Mi:dm line as a Korea-centric family: Mi:dm 2.0 shipped as a 12B base-instruct model and a 2B mini, both released as open weights on Hugging Face in July 2025, and the enterprise materials describe the family as tuned for Korean society, culture and domain knowledge rather than for English-language coding benchmarks. A 120B hybrid at this price is that programme scaled up, not a pivot to the global frontier.

That framing tells a buyer what to test first. If the workload is Korean-language customer service, domestic regulatory text, or a domain corpus where the vocabulary is Korean, a model trained for that is a different proposition from a frontier model with Korean competence bolted on. If the workload is English-language agentic coding against a large repository, this is a cheap model with no evidence behind it and a context window a quarter the size of the leaders'.

It is also worth being clear about what is not open. Mi:dm 2.0's weights are public; Mi:dm K 3.0 Pro's are not. There is no repository on KT's Hugging Face organisation carrying the 3.0 name, and the model is served only through KT's platform. That closes off self-hosting, fine-tuning on your own hardware, and the whole class of evaluation that requires downloading the checkpoint.

How to call it, and what the single provider means

Mi:dm K 3.0 Pro is served by exactly one provider — KT — and OrcaRouter does not currently route it. That matters for a specific reason rather than as a disclaimer. A single-provider model cannot fail over. If KT's endpoint degrades, there is no second host to route around, and the availability figure quoted for the model is a measurement of one company's infrastructure rather than of a market. Teams that need a hard availability guarantee usually want either a second provider or an application-level fallback, and for this model only the second option exists.

The comparison set is a different story. OrcaRouter carries more than 200 models behind a single key at provider list price with 0% markup on top, so the cheap models this one is actually competing with — D​eepSeek's tiers, the G​LM line, Q​wen — and the frontier models it is being measured against sit on the same endpoint and the same invoice. That makes a head-to-head trial a routing change rather than a procurement project, which is the sensible way to answer the question this page cannot: send the same 50 prompts to Mi:dm K 3.0 Pro and to two models you already trust, and score the outputs yourself. At $0.20 in and $0.80 out, that experiment costs cents.

What would change this page

Four developments would move Mi:dm K 3.0 Pro from "interesting price, unknown model" to something a team can plan around, and each is checkable.

An independent evaluation. A single Artificial Analysis entry with an Intelligence Index score and a per-task cost would do more for this model's standing than any press release, because it would put the number beside the models it is priced against — the same index puts Claude Opus 5.5 at 58 and DeepSeek V4 Pro at 36, and a figure in that band would settle the argument about whether the discount is compensation or arbitrage.

A published vendor benchmark sheet. The absence is conspicuous enough that its appearance would itself be news, and would at least give the model's own designers' claims something to be checked against.

A second serving provider. Two independent hosts would make failover possible and would let a buyer treat the price as a market rate rather than a promotional one.

Weights. KT open-sourced Mi:dm 2.0 and has not open-sourced this. If that changes, the model becomes evaluable by anyone with a GPU, and the measurement gap closes without waiting for an evaluator.

Until at least one of those happens, the defensible statement about Mi:dm K 3.0 Pro is narrow and specific: a 120B hybrid mixture-of-experts model, 12B active, 262K context, text-only, served from a quantized checkpoint by one provider, at a price that undercuts every frontier model by an order of magnitude — and no published evidence of how well it works. That is enough to justify a cheap test. It is not enough to justify a migration.