Moonshot AI has published the weights for Kimi K3.1 Pro. As of November 12 the Hugging Face repository is public and ungated: about 1.56 TB in 96 safetensors shards, created November 2 and last changed November 5 per the Hugging Face API. Moonshot’s late-October launch post said “Full model weights will be released by Nov 4, 2026.” No primary page states the day the repository opened. The hosted model lists at $2.00 input and $10.00 output per million tokens, one-third below Kimi K3. The license is K3’s: a model-as-a-service operator with more than $20 million in group revenue over twelve months needs a separate agreement with Moonshot.
S5 Labs has run neither the weights nor the API. Everything here is read from the linked pages: the license, the model cards, Moonshot’s docs, two inference-project pages, Hugging Face discussions and Artificial Analysis. No benchmark or speed figure is ours. The score table is Moonshot’s, and the hosted-model measurements are Artificial Analysis’s.
The model card describes K3.1 Pro as post-trained from Kimi K3, whose weights we covered when they shipped. It lists a mixture-of-experts model with 2.8T total and 104B activated parameters, 896 experts with 16 selected per token, a 1,048,576-token context, text and image input, and MXFP4 weights with MXFP8 activations. A footnote on the card says that, compared with K3, the full-attention Gated MLA layers are replaced with Gated DeepSeek Sparse Attention (Gated DSA) layers.
The license is K3’s license with the name changed
A diff of the K3.1-Pro LICENSE against the K3 LICENSE returns two lines: the title and the attribution string. The table is a plain reading of the text, not legal advice.
| Your situation | What the license says |
|---|---|
| Internal use: neither the software, its outputs nor its “underlying capabilities” reach third parties | Clauses 2 and 3 do not apply (clause 4a). |
| Access through Moonshot’s official products or “certified inference partners” | Clauses 2 and 3 do not apply (clause 4b). The license does not define the term, and we found no partner list on the model card, in the launch post or in the API docs we read. |
| An end-user product with the model “solely embedded within specific features or harnesses”, or “mere relaying of requests to models hosted by others” | Outside the Model as a Service definition (clause 2). |
| You or an affiliate give third parties inference or fine-tuning access with “meaningful control over the inputs, parameters, or training data”, and group revenue exceeds $20 million over any consecutive 12 months | A separate agreement with Moonshot “before using the Software or its derivative works for any commercial purpose”. |
| A commercial product with more than 100 million monthly active users, or more than $20 million in monthly revenue | “Kimi K3.1-Pro” must be prominently displayed in the interface (clause 3). |
The clause 2 test is “the aggregate revenue of the Licensee and its affiliates”, and nothing in the sentence limits it to revenue from hosting. The consequence covers any commercial purpose, so a hosting business that trips it reaches the rest of the company’s commercial use of the model.
The license does not define “meaningful control” or say what follows without an agreement. A provider near the line should ask [email protected] and get the answer in writing before deploying.
Other open-weight licenses set their own model-as-a-service revenue lines: $20 million for Muse Spark Open, $50 million for Qwen3.8-Max, $10 billion for GLM 5.3. Each test is worded differently, so read the one you are licensing.
API cost: a third below K3, more than double K2.7
Rates are per million tokens from Moonshot’s pricing page. The last column is S5 Labs arithmetic for one request of 100,000 input and 10,000 output tokens.
| Model and mode | Input | Output | 100K in + 10K out |
|---|---|---|---|
kimi-k3.1-pro | $2.00 | $10.00 | $0.30 |
kimi-k3.1-pro, all input read from cache | $0.20 | $10.00 | $0.12 |
kimi-k3.1-pro, Batch API at 60% | $1.20 | $6.00 | $0.18 |
kimi-k3.1-pro-highspeed | $4.00 | $20.00 | $0.60 |
kimi-k3 | $3.00 | $15.00 | $0.45 |
kimi-k2.7-code | $0.95 | $4.00 | $0.135 |
The cached row leaves out cache writes, which for kimi-k3.1-pro cost $2.00 per million for 5 minutes and $4.00 for an hour. The K3.1 Pro quickstart says there is “no tiering by context length”. K3 still lists at the $3/$15 our K3 launch article reported.
Per the pricing page, Moonshot closed K2 models to newly registered users on October 31 and plans to retire the series by the end of December 2026. At the table’s mix, moving from kimi-k2.7-code to K3.1 Pro costs about 2.2 times as much. Highspeed doubles the price for what Moonshot’s platform quickstart calls “about 2-3× the speed”; we found no independent measurement of it. Artificial Analysis measured the standard hosted model at 32 output tokens per second and scores it 46 on its Intelligence Index v5.0.1, which combines nine evaluations. Version 5 added, replaced and removed evaluations, so we would not set that 46 against the 57 our K3 launch article cited.
Migrating a K2 or K3 integration
Steps without their own link come from the K3.1 Pro quickstart.
- Top up. K3.1 Pro is “unlocked after a successful top-up (minimum $1)”, and cumulative top-up sets your rate-limit tier.
- Change the model ID to
kimi-k3.1-proathttps://api.moonshot.ai/v1. The platform offers Chat Completions, Responses and Anthropic-format Messages APIs. - Delete sampling parameters.
temperature,top_p,n,presence_penaltyandfrequency_penaltyare fixed and must be omitted. - Remove K2.x thinking configuration and set
reasoning_effort, a top-level field takinglow,highormax(the default). Thinking cannot be turned off. Our K3 launch article described K3 as locked to max, so the lower settings are a change. - Return the assistant message whole. Per the
reasoning_effortguide, multi-turn and tool-call requests must send backreasoning_contentandtool_callsunchanged. - Check
max_completion_tokens. It defaults to 131072 and goes up to 1048576. - Convert image inputs. Public image URLs are not supported; use base64 or
ms://<file-id>.
The docs put no number on how many more output tokens max produces than low; measure it on your own tasks before fixing a budget.
Self-hosting: eight B300s, and what is left for KV cache
The card recommends vLLM, SGLang and TokenSpeed and states no GPU model, count or memory requirement. The sizing comes from other pages:
| Quantity | Figure | Source |
|---|---|---|
| Baseline deployment | One tensor-parallel group of 8 GPUs, “either two GB300 trays or one B300 node”, vLLM 0.34.0+ | vLLM recipe |
| KV cache at TP8, FP8 KV, 8x B300 | About 2.6M tokens, roughly two and a half 1,048,576-token contexts | vLLM recipe; division ours |
| KV cache in the recipe’s DCP8 mode | About 18.5M tokens, between 17 and 18 full contexts; decode steps 7-25% slower | vLLM recipe; division ours |
| One user’s 8x B300 server | 2,304 GB of HBM3e, about 208 GB per GPU in use, about 80 GB per GPU free after bring-up | HF discussion #18, self-reported |
The vLLM page labels only its prefill benchmark as run on a pre-release checkpoint, and it says “Throughput was not tuned.” SGLang’s cookbook says “No cell has a cookbook-format speed sweep yet”, and the discussion post has no speed numbers.
The only measured throughput on these pages is in SGLang’s cookbook: about 2,583 decode tokens per second at batch 64 and 8K context on a two-node B300 prefill/decode run with a draft model. That is not a sustained long-context figure, so a break-even still has to start from the price side. At $10.00 per million output tokens, each $10 an hour of node cost has to be earned back with 1 million output tokens an hour, about 278 tokens per second summed across concurrent requests. That counts output only, at list price and full utilization. At the $30 to $60 an hour our K3 weights article estimated for an eight-B300 node, a July figure we have not re-priced, the bar is roughly 830 to 1,670 tokens per second.
Moonshot’s own scores are mixed
The card reports these at max effort, temperature 1.0 and top-p 0.95, with most Kimi coding runs on the Claude Code harness. Its footnotes say the Terminal-Bench and DeepSWE comparator scores come from vendor posts, system cards or the leaderboard.
| Benchmark (Moonshot-reported) | K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| Terminal-Bench 4.0 | 65.7 | 66.4 | 58.2 |
| DeepSWE v1.1 | 75.5 | 74.2 | 74.1 |
| OSWorld V2.1 | 79.6 | 81.8 | 75.8 |
| HLE-Diamond | 39.4 | 60.9 | 65.8 |
| HLE-Full with tools | 70.0 | 67.7 | 57.2 |
| ExploitBench | 74.7 | 91.0 | 100 |
The table shows three of the card’s six model columns; Claude Fable 5.1, DeepSeek-V4.1-Flash-1017 and GLM-5.3 are left out. In those columns, DeepSeek-V4.1-Flash-1017 scores 77.0 on DeepSWE v1.1, above all three shown, and Claude Fable 5.1 is ahead of K3.1 Pro on OSWorld V2.1 (80.7), HLE-Diamond (55.1) and ExploitBench (83.0) and behind it on the other three rows.
Moonshot’s launch post says that for the open-weight release it “identified and removed cyber-related data from both training stages”, and it reports 81.0% on its internal xdev-bench-hard against 90.5% for a checkpoint without those measures. Read the ExploitBench row with that in mind.
The post does not say whether the hosted kimi-k3.1-pro is the same checkpoint as the download. In a Hugging Face discussion a user asks whether the card’s benchmarks are for the hosted model or the weights; it showed no Moonshot reply when we read it. Artificial Analysis measured the hosted API. Run your own evaluation on the weights you serve.
LongCat-2.5 is plain MIT, with NPU-only deployment notes
Meituan’s LongCat-2.5 repository was created on November 4. Its license line reads “The model weights are released under the MIT License”, with no revenue clause, and adds that it “does not grant any rights to use Meituan trademarks or patents.”
The card says 1.6 trillion total parameters and approximately 48 billion activated, while Hugging Face’s safetensors metadata totals 1,779,752,648,320, a gap the card does not explain. The repository holds about 3.56 TB, roughly 2.3 times K3.1 Pro’s. The card’s only deployment instructions target NPU platforms through an SGLang-FluentLLM branch; we read the card, not the linked repository. The Hugging Face API showed no inference provider on November 12. The scores Meituan measured itself are labeled “indicative only” and used different harness settings from Moonshot’s, so they cannot rank the two.
Choosing between the API, your own node and LongCat-2.5
On the API, K3.1 Pro undercuts K3 by a third. kimi-k2.7-code users should budget about 2.2 times their current spend at the table’s mix and move before the December retirement.
If data cannot leave your network, internal use sits outside clauses 2 and 3. Plan for an eight-GPU B300-class node and hold your measured throughput against 278 tokens per second for each $10 an hour of node cost.
Anyone selling inference should add up twelve months of group revenue before loading the weights. Above $20 million, the agreement comes first. LongCat-2.5 offers MIT terms with no revenue test, provided you have the hardware and serving stack its card documents.
