Kimi K3.1 vs Kimi K3: What the Open Weights Say That the Launch Post Did Not

Kimi K3.1 Pro and Kimi K3 now ship the same 96 safetensors shards, the same 1.56 TB on disk, and a different attention stack. The configuration in Moonshot's new repository describes 93 layers as "69 KDA + 24 Gated DSA". The configuration in Kimi K3's repository, published in July, says "69 KDA + 24 Gated MLA". Moonshot's launch post for Kimi K3.1 Pro, dated 24 October 2026, calls the newer model "a post-trained version of Kimi K3" and does not mention the swap at all. Everything else in this comparison sits downstream of that one line, including the reason the newer model is the cheaper one.

The weights themselves arrived without an announcement. Moonshot's blog has no entry after 24 October. The repository's only commit is timestamped 2 November 2026, two days before the 4 November deadline the company had set for itself; whether the files were publicly readable before that deadline is not something the repository records. What it does record is the architecture, the licence and the inventory — which is exactly what the launch post left out.

The swap the announcement did not mention

The model card carries the admission in a footnote, one line under an otherwise familiar specification table: "Compared with Kimi K3, we replace the full-attention Gated MLA layers with Gated DeepSeek Sparse Attention (Gated DSA) layers for more efficient long-context modeling."

Twenty-four of the 93 layers change. The other 69 remain Kimi Delta Attention, the linear-attention mechanism that Kimi K3 introduced, and the layer count, the expert count and the hidden dimensions are untouched. Gated MLA is full attention with a compressed key-value cache; Gated DSA selects a sparse subset of past tokens for each query and gates the result. The practical difference is in how the model spends compute and cache as the context grows, which is the part of a 1M-token window that shows up on your serving bill rather than in a benchmark table.

What the card does not provide is a number. There is no tokens-per-second figure, no cache-memory comparison, no long-context latency measurement — only the phrase "more efficient long-context modeling." Moonshot's launch post gives no architecture detail whatsoever. So the claim that the swap makes anything faster is, at this point, a design statement rather than a measurement, and nobody outside Moonshot has published a serving comparison between the two checkpoints.

That matters more than it sounds. If you tuned batching, cache allocation or context management around Kimi K3's attention profile, the newer checkpoint is not a drop-in at the serving layer even though it is a drop-in at the API layer. The API hides it; a self-hosted deployment does not.

What the two checkpoints genuinely share

The identical download is the tell that this is not a larger model. Every headline dimension below is the same on both cards, and the file inventory confirms it:

• Total and active parameters — 2.8 trillion total, 104 billion active on both models.

• Depth — 93 layers, of which one is dense, on both.

• Experts — 896 total, 16 selected per token, 2 shared, on both.

• Vocabulary — 160K tokens on both.

• Vision — MoonViT-V2, 401 million parameters, on both.

• Context — 1,048,576 tokens on both; our own model page lists a 943,718-token maximum output for Kimi K3.1 Pro.

• Quantisation — MXFP4 weights and MXFP8 activations, quantization-aware training, on both.

• Download — 96 safetensors shards totalling 1.56 TB for each repository.

• Input modalities — Kimi K3's card lists text, images and video; Kimi K3.1 Pro's card lists text and image, while the API documentation advertises video as well. Treat the video claim as unverified until the processor code is exercised.

A point release that keeps the parameter count, the expert routing and the file size, and changes 24 attention layers, is a different kind of release from a new generation. It is closer to a mid-cycle architecture revision plus a large post-training run than to a new base model, and the pricing reflects that.

Two rate cards, and an invoice that does not fall by a third

On paper the newer model is cheaper on every line. Kimi K3.1 Pro is $2.00 per million input tokens and $10.00 per million output; Kimi K3 is $3.00 and $15.00. The cache-hit rate falls from $0.30 to $0.20, cache writes from $3.00 and $6.00 to $2.00 and $4.00 at five-minute and one-hour TTL, and a Highspeed variant at $4.00 and $20.00 trades price for two to three times the output speed. Batch inference on Kimi K3.1 Pro bills at 60% of standard, or $1.20 and $6.00.

The invoice is a different story. Kimi K3.1 Pro reasons by default at maximum effort, and maximum effort writes more tokens. Bito's agentic-coding comparison found the two models tied at 72% overall while Kimi K3.1 Pro emitted roughly 98,000 output tokens against Kimi K3's 44,000, took 28 minutes against 17, and cost $4.28 against $2.46 for a long session. Plotly's data-analysis run went the other way on quality and the same way on cost: 42 of 43 tasks solved against 35, with spend rising from $5.01 to $13.47 and median time from 43 seconds to two and a half minutes. Ultralytics, testing auto-annotation, measured a 33% improvement in mAP50, 0.444 against 0.334, moving from 19th to 9th of 39 models.

Those three are vendor-published tests on small samples, not peer-reviewed evaluations, and they disagree about how much better the new model is. They agree on the direction of the cost: a one-third cut in price per token does not survive a model that thinks for longer. Budget per completed task, not per million tokens.

Both sides of this comparison are callable on one key. Kimi K3.1 Pro is routed as kimi/kimi-k3.1-pro at Moonshot's $2.00 and $10.00, and Kimi K3 as kimi/kimi-k3 at $3.00 and $15.00, with the provider's list price passed through and no markup added, so a change to Moonshot's rate card is live here the same day. That is the specific reason it belongs in a pricing section rather than a footer: the entire argument for switching is a price gap, and any spread added on top of the list price would eat part of the saving before you had measured a single task.

The benchmark table moved under both models

Artificial Analysis rebuilt its Intelligence Index on 27 October 2026, and the rebuild is more consequential to this comparison than anything Moonshot published. On the current v5 index, Kimi K3.1 Pro scores 46 and sits ninth, twelve points behind Claude Opus 5.5 at 58, with Gemini 4 Carbon at 59, GPT-6 Astra at 57 and GPT-6.1 Sol at 55 above it. On the previous v4.3.2 index the same model scored 55 and ranked fourth, three points off the lead.

Those two figures are not a regression, and they are not comparable. Version 5 moved roughly 60% of the index onto private held-out sets, added FrontierCode 2.0 and Terminal-Bench-Science, removed SciCode and AA-LCR, and raised the median output length per task from about 35,000 tokens to about 59,000. A model that answers briefly loses ground on an index that now rewards long agentic work. Kimi K3.1 Pro remains the highest-ranked model its vendor describes as open-weight.

The predecessor's number is the trap. Kimi K3 scored 57 on the v4 index at launch and ranked third. The 43.6 that circulates for Kimi K3 (max) comes from aggregator snapshots taken before the v5 cutover, and we could not find a published v5 score for Kimi K3 itself. So "46 against 43.6" is not a controlled pair, and anyone presenting the difference as a measured capability gain is subtracting across two instruments. What can be said is narrower: on the current index, the newer model is the one that has been scored, and it is ninth.

The vendor's own table, which is a different instrument again, has Kimi K3.1 Pro at 65.7 on Terminal-Bench 4.0 against Claude Opus 5.5's 66.4, 51.7 on SWE Marathon v1.1, 92.7 on BenchCAD with tools, 94.2 on GPQA-Diamond and 51.1 on HLE-Full. Its soft spots are equally visible: 38.1 on Terminal-Bench-Science against GPT-6 Astra's 68.1, and 39.4 on HLE-Diamond against 60.9 for Opus 5.5. Moonshot ran the coding evaluations through the Claude Code harness at maximum reasoning effort. None of it has been reproduced independently.

The cleanest paired numbers in the launch are the ones Moonshot measured against its own predecessor: Kimi Code Bench 3.0 Long-horizon at 61.2% against 55.95%, a claim-level hallucination rate of 8.3% against 23.4% with tools, and 81.0% against 33.3% on xdev-bench-hard. The same internal benchmark belongs to Moonshot, so the deltas are consistent but not external. The hallucination figure is the one with the widest practical consequence and the least corroboration.

The licence is Kimi K3's licence with one name changed

Both repositories ship a licence file of about three kilobytes, and a line-by-line comparison finds a single substantive difference: "Kimi K3.1-Pro" is substituted for "Kimi K3" in the attribution clause. Everything else is the same document.

That means the terms are already known rather than pending. Use, copying, modification, distribution, sublicensing and sale are permitted, and fine-tuning is explicitly allowed. A business that provides model-as-a-service and whose aggregate revenue exceeds $20 million over any consecutive twelve months must sign a separate agreement with Moonshot before commercial use. Products above 100 million monthly active users or $20 million in monthly revenue must display the model name prominently. Internal use, and use through Moonshot's official products or its certified inference partners, are exempt from both requirements.

The exemption for certified partners is the clause worth reading twice if you plan to resell access. The $20 million threshold is measured over a rolling twelve months, not per product line, and "model as a service" is defined broadly enough to cover an API that gives customers control over inputs, parameters or training data — while carving out end-user products where the model is embedded in a feature, and plain relaying of requests to models hosted elsewhere.

What it takes to run either one

A 1.56 TB MXFP4 checkpoint is not a workstation download, and the two models are equally heavy. Moonshot points at three serving stacks — vLLM recipes, an SGLang cookbook and TokenSpeed — and all three assume a multi-node deployment at this size. Quantization-aware training from the supervised fine-tuning stage is what keeps the file at 1.56 TB rather than the roughly 5.6 TB a BF16 copy would occupy, and that quantisation is baked into the scores above rather than applied afterwards.

The practical consequence of identical footprints is that the architecture swap buys no hardware relief. If you can serve Kimi K3, you can serve Kimi K3.1 Pro on the same class of cluster; you cannot serve either on a single node.

Which one to choose

• You self-host and your stack is tuned to Kimi K3 — re-benchmark before migrating. The attention change alters cache behaviour, and the only efficiency claim attached to it is a sentence in a footnote.

• You call the API for agentic coding, long-horizon tool use or document verification — Kimi K3.1 Pro, on the strength of the paired internal deltas and the lower rate card, with a shadow evaluation first.

• You need a model with a settled failure profile — Kimi K3. It has been in production since July and its quirks are documented by other people. The newer model changes reasoning behaviour by design.

• You need determinism — check before you switch. Kimi K3.1 Pro fixes temperature at 1.0 and top_p at 0.95 and rejects alternatives, and changing reasoning effort mid-conversation invalidates prefix-cache hits.

The honest summary is that Kimi K3.1 Pro is a better model than Kimi K3 on almost every axis the vendor measured, a modestly better one on the only independent index that has scored it, and a more expensive one per completed task than its rate card implies. What changed this week is not the verdict but the evidence: the checkpoint can now be read, served and tested instead of taken on trust, and the first thing reading it produced was a correction to the launch post.