Kimi K3.1 vs Kimi K4: What K4 Has to Beat, Now That the Weights Are Out

Kimi K4 does not exist. Kimi K3.1 Pro does, and since this week it exists as a download as well as an endpoint: 96 shards, 1.56 TB, a licence file and three serving recipes. That asymmetry is the entire comparison. One side is a fixed, inspectable artefact. The other is a slogan photographed on an internal banner and a chip-procurement report that named no model. So the question worth answering is not which one is better. It is what Kimi K4 would have to demonstrate before waiting for it became a rational decision.

That question has a real answer now, because the baseline stopped moving. Until 4 November, "wait for K4" and "wait for the weights" were the same instruction with different deadlines. They are no longer the same instruction, and only one of them still has a date attached.

The family's dated artefacts, in order

Moonshot's cadence is unusually legible for a lab at this scale, which is what makes the absence of a K4 artefact informative rather than merely disappointing.

• Kimi K3 — announced 16 July 2026 as the first open model in the 2.8-trillion-parameter class; weights followed on 27 July, eleven days later.

• Kimi K3.1 Pro — announced and sold through the API on 23 October 2026, described as a post-trained successor with a 1M-token context and the same 2.8 trillion total parameters. The tech blog is dated 24 October.

• Kimi K3.1 Pro weights — promised "by Nov 4"; the repository's only commit is timestamped 2 November 2026, and the download is live.

• Kimi K4 — no blog post, no model card, no repository, no API identifier, no rate-card row, no licence, no Artificial Analysis page.

Two generations since July 2025 — K2 in July 2025, K3 in July 2026 — with point releases between them is the pattern. Kimi K3.1 Pro is exactly the kind of release that pattern predicts, and a K4 inside 2026 would be the exception rather than the continuation.

The bar K4 has to clear

A successor is measured against the thing it replaces, and that thing is now a file you can read. These are the numbers Kimi K4 would inherit, improve on or lose to.

• Size — 2.8 trillion total parameters, 104 billion active, 93 layers, 896 experts with 16 selected per token and 2 shared.

• Context — 1,048,576 tokens, with a 943,718-token maximum output on our model page.

• Weight footprint — 96 safetensors shards, 1.56 TB, MXFP4 weights and MXFP8 activations from quantization-aware training.

• Price — $2.00 per million input and $10.00 per million output, $0.20 cached input, $1.20 and $6.00 on the Batch API, $4.00 and $20.00 on the Highspeed variant.

• Independent standing — 46 on the Artificial Analysis Intelligence Index v5, ninth overall, and the highest-ranked model its vendor calls open-weight.

• Vendor benchmarks — 65.7 on Terminal-Bench 4.0 at maximum reasoning effort, 51.7 on SWE Marathon v1.1, 92.7 on BenchCAD with tools, 51.1 on HLE-Full.

• Soft spots — 38.1 on Terminal-Bench-Science and 39.4 on HLE-Diamond, both far below the closed frontier.

• Licence — the Kimi K3.1 Pro License: permissive use, modification, distribution and fine-tuning, with a separate agreement required for model-as-a-service businesses above $20 million in annual revenue.

The soft spots are the interesting rows. A K4 that arrives with another two points on Terminal-Bench and the same Terminal-Bench-Science score would be a point release with a new number. The place where the open-weight frontier is genuinely behind is scientific and diamond-hard reasoning, and that is where a real generation change would have to show up.

Three tests that would justify waiting

Waiting is a bet, so it should have conditions. These are the three we would apply, in order of how much they would change a decision.

The first is a capability move larger than the measurement noise. Benchmark scores swing with the harness, the effort setting and the judge, which is why the same model can read 55 on one index version and 46 on the next without changing a weight. A K4 that beat Kimi K3.1 Pro by two points on a re-scored index would not clear that bar. One that moved Terminal-Bench-Science from 38 to the high fifties, or HLE-Diamond from 39 into the sixties, would.

The second is openness at the same speed. Kimi K3.1 Pro's weights arrived ten days after the API launch, under a licence whose terms were published with them. A K4 that shipped API-first with an open-ended promise, or with a licence that added restrictions the K3 line did not carry, would be a worse deal than the model it replaced regardless of its scores. The artefact order matters here: a model card and a licence file are harder to walk back than a blog post.

The third is price, and it is the hardest to clear. Kimi K3.1 Pro already cut its input, output and cache prices by a third against its predecessor. A K4 at the same $2 and $10 would be a capability upgrade with no cost relief; a K4 above it would need to be dramatically better to justify the increase, in a market where the open-weight tier is converging on price rather than diverging.

What the K4 evidence actually is, checked again this week

Nothing on this list has changed since we audited it in October. The one thing that changed — Kimi K3.1 Pro shipping — moved the baseline, not the evidence.

• The name comes from a leaked internal celebration banner from late July 2026 carrying the phrase "push K4 to the absolute limit". A phrase on a slide is where a product name can start; it is not an announcement.

• The only substantive report is The Information's, from around 29 July 2026, that Moonshot was seeking additional Nvidia Blackwell-class GPUs to train a successor larger than K3. It named no model, no parameter count and no date. It is a procurement story.

• The 5-to-6-trillion-parameter figure appeared in Chinese-language coverage in the second week of October, attributed to no named source and present in no Moonshot document. The word doing the work in those reports is "probably".

• One leaker has claimed a Moonshot flagship codenamed "x4" has finished pre-training and is in mid-to-late post-training. Whether x4 is K4 is unconfirmed, and a single unverified account is a lead rather than evidence.

• The checkable surfaces are all empty. Moonshot's blog has no post after 24 October. The Hugging Face organisation's newest text model is Kimi K3.1 Pro; a search for K4 returns third-party re-uploads. The API's model list carries kimi-k3.1-pro, kimi-k3.1-pro-highspeed, kimi-k3, kimi-k2.7-code and kimi-k2.6. The pricing page has no K4 row. Artificial Analysis has no K4 page.

There is also a policy risk that attaches to a training run rather than to an API. Moonshot has faced reported allegations about how it obtained Blackwell-class chips, and a threatened Entity List designation would constrain the GPU supply a K4 depends on far more than it would constrain a model that has already shipped and been downloaded. That is a reason to discount an aggressive K4 timeline, not a reason to doubt Kimi K3.1 Pro's availability.

What waiting actually costs

The cost is not zero, and it is easier to state than the benefit. Every month spent waiting is a month without a model that scores 46 on the current index, runs a million-token context, costs $2 and $10, and can be downloaded and inspected. If your workload is ordinary agentic coding, document extraction or long-context retrieval, the thing you would be waiting to replace is already better than what most teams are running.

The second cost is architectural drift. A team that defers a decision tends to build against whatever is convenient in the meantime, which makes the eventual migration more expensive than the one it postponed. The way to keep the option cheap is to put the model behind an interface you control. OrcaRouter's routing DSL does that concretely: you address a policy rather than a model, so when a K4 identifier eventually exists, adding it is a new route with its own weights and fallback rather than a rewrite of the calling code. Automatic failover across providers runs behind the same endpoint, which matters for a model this new — a production path pinned to a single upstream is the failure mode a router exists to remove.

The third cost is harder to quantify but real: the open-weight field is moving. Kimi K3.1 Pro is the current leader among models described as open-weight, but DeepSeek V4.1 Flash and GLM-5.3 sit behind it on the same index, and a six-month wait is long enough for that order to change. Waiting for K4 is not waiting in place.

The signals that would actually move the date

If you want to monitor this rather than guess at it, watch artefacts in the order a lab usually produces them.

• An API model identifier. This is the cheapest thing to publish and usually the first — a string in the model list, often before any announcement.

• A rate-card row. A price is the second cheapest artefact, and the one that most clearly signals a commercial launch rather than a demo.

• A Hugging Face repository with a licence file. For Moonshot this has meant an eleven-to-twelve-day gap after the API launch, so a repository appearing is a late signal, not an early one.

• A model card and a benchmark table with harness settings. The Kimi K3.1 Pro card is unusually specific — it names the Claude Code harness, the reasoning effort and the judge model — and that level of disclosure is itself a signal of a finished release.

• A product surface. Kimi Code, Kimi Work and the mobile apps have historically exposed new model identifiers before the developer documentation catches up.

What does not count: parameter counts with no document behind them, benchmark numbers extrapolated from K3, and posts that begin "K4 is coming" without one of the five artefacts above. Those have been circulating since July and none of them has aged into a specification.

The decision rule

Wait only if your workload lives in the two places where Kimi K3.1 Pro is demonstrably weak — scientific terminal work and diamond-hard reasoning — and you can absorb a wait measured in quarters rather than weeks. For that team, the case for waiting is real: the gap to the closed frontier on those tasks is wide enough that a real generation change would matter, and no point release is likely to close it.

For everyone else, the answer is to build on what exists. Kimi K3.1 Pro is priced below the proprietary models it sits beside, it is downloadable, and its failure modes are being documented by third parties in public. Kimi K4 is a scheduling question, and the schedule is the one thing about it nobody has published. Betting a roadmap on a photographed slogan is not patience. It is a guess with better branding.