Kimi K3.1 Pro: Benchmarks, Pricing, Open Weights and How to Run It

TL;DR

Kimi K3.1 Pro is Moonshot AI's open-weight flagship and the first model in its K3.1 series, a post-trained upgrade of Kimi K3. It is a 2.8T-parameter mixture-of-experts model with 104B parameters active per token, a 1,048,576-token context window, text and image input, and thinking that cannot be turned off. The API launched October 23, 2026. The weights went up on Hugging Face on November 2 under the Kimi K3.1 Pro License, an MIT-style license whose two conditions start at $20 million in revenue or 100 million monthly users. Moonshot charges $2 per million input tokens, $0.20 cached and $10 output, a third less than K3. In Moonshot's own table it scores 65.7 on Terminal-Bench 4.0 against Claude Opus 5.5's 66.4, and Artificial Analysis ranks it the top open-weights model on Intelligence Index v5 at 46. On Morph it runs as morph-kimik31 at $2.50 input and $14.00 output per million, or unlimited on the $200 Ultra plan.

Published and last updated November 9, 2026. Benchmark numbers are Moonshot-reported unless the row or sentence names Artificial Analysis.

Kimi K3.1 Pro at a Glance

Every row comes from Moonshot's model card, release post, platform docs, or the Hugging Face API, read November 9, 2026.

What Is Kimi K3.1 Pro?

Kimi K3.1 Pro is the first model in Moonshot AI's K3.1 series. Moonshot calls it "a post-trained version of Kimi K3, with substantially larger post-training compute." The base is the same 2.8T mixture-of-experts model. What changed is one attention swap and large-scale joint reinforcement learning across "hundreds of downstream domains." Moonshot says the gains land in coding, 3D creation, financial research, factual accuracy and agentic work.

There is no plain "Kimi K3.1" model. Moonshot shipped one model in two API variants: kimi-k3.1-pro and kimi-k3.1-pro-highspeed, which Moonshot says runs 2 to 3 times faster at twice the per-token price. Only K3.1 Pro has open weights. It is available in Kimi.ai, Kimi Work, Kimi Code and the Kimi API, and through hosts that run the open weights. For the original model from July 2026, see our Kimi K3 guide.

On Morph the model ID is morph-kimik31, on the same OpenAI-compatible and Anthropic-compatible endpoints as every other Morph model. The older morph-kimik3 ID still works as an alias and now serves the K3.1 weights; new code should use morph-kimik31.

What Changed From Kimi K3

Moonshot did not publish a full K3 vs K3.1 table. K3 is not a column in its main benchmark table. What it did publish is a set of head-to-head numbers scattered through the release post. Artificial Analysis published its own before-and-after runs. Both are below, labeled by source.

Three changes matter more than any single row. First, the attention swap: Moonshot replaced K3's full-attention Gated MLA layers with Gated DeepSeek Sparse Attention "for more efficient long-context modeling." The architecture section below covers what that means. Second, the price: Artificial Analysis confirms the first-party rates are "33% lower than Kimi K3 on all three rates." Third, the model talks more. Artificial Analysis counted 55K output tokens per index task, 14% more than K3's 48K, so part of the per-token saving goes back into token volume.

One change runs the other way. For the open-weight release, Moonshot says it "identified and removed cyber-related data from both training stages, while retaining the general-purpose coding tasks needed for model usefulness," and that K3.1 Pro's cyber capability "remains behind that of frontier proprietary models that have been publicly available for at least three months." The xdev-bench-hard row above still more than doubled.

Kimi K3.1: Confirmed Facts vs Loose Claims

Several numbers about K3.1 circulate in two versions. Each pair below is real; they just measure different things.

Architecture: KDA Plus Gated DeepSeek Sparse Attention

Kimi K3.1 Pro keeps K3's mixture-of-experts layout. Each token routes to 16 of 896 experts plus 2 shared experts, for 104B active parameters out of 2.8T. The model has 93 layers (one dense), a 7,168 attention hidden size with 96 heads, a 3,584 latent MoE dimension, 3,072 hidden units per expert, a 160K vocabulary and SiTU-GLU activations. Vision comes from MoonViT-V2, a 401M-parameter encoder.

The change is in attention. 69 of the attention layers are Kimi Delta Attention (KDA), carried over from K3. In K3 the other 24 were full-attention Gated MLA layers, which read every earlier token for every new one. K3.1 replaces them with Gated DeepSeek Sparse Attention. A lightweight indexer (32 index heads, head dimension 128) scores the earlier tokens, and the attention step reads only the top 2,048 of them instead of all of them. Moonshot's stated reason is "more efficient long-context modeling." vLLM's recipe describes the same change as adding "sparse attention (DSA, top-2048 indexer) to 24 of its MLA layers."

The checkpoint was trained quantization-aware: MXFP4 weights with MXFP8 activations. vLLM's recipe is specific: "The checkpoint ships MXFP4 routed-expert weights; everything else is BF16." That is why a 2.8T-parameter model downloads as 1.56 TB. Moonshot recommends vLLM, SGLang and TokenSpeed as inference engines.

Benchmarks: Moonshot's Table, Then Independent Runs

The table below is Moonshot-reported, from the Hugging Face model card. All K3.1 Pro results use max reasoning effort, temperature 1.0 and top-p 0.95, and most coding rows ran K3.1 Pro in the Claude Code harness. The Claude and GPT columns are labeled "max, w/ fallback" and come from vendor posts or official leaderboards, so harnesses differ row to row. The table compares against Claude Fable 5.1, which is now a legacy model; Claude Fable 5.2 launched November 9, 2026 and is not in it.

Read it as two stories. K3.1 Pro leads on long-horizon agent work: SWE Marathon, DeepSWE, HLE with tools, BrowseComp, Automation Bench-AA and OfficeQA Pro v2. It trails on the hardest program synthesis and science tasks: 31.5 on Program Bench against Opus 5.5's 65.0, and 38.1 on Terminal-Bench-Science against 63.3. The Terminal-Bench 4.0 headline is a near tie, 65.7 to 66.4. Moonshot's footnote says competitor scores there are "the maximum values reported in model blog posts or on the official Terminal-Bench leaderboard," and K3.1 Pro "is evaluated with the Claude Code harness." Moonshot reports no SWE-bench Verified, SWE-bench Pro or LiveCodeBench score.

Independent signal: Artificial Analysis

Artificial Analysis ran its own evaluations, and the result depends on which version of its index you read. At launch on October 23, on the index version then current, it scored Kimi K3.1 Pro 55, fourth overall and 11 points above K3's 44. Only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56) were ahead. On October 27 it released Intelligence Index v5. On v5, Kimi K3.1 Pro scores 46, the top open-weights model, with Opus 5.5 leading at 58, then GPT-6 Astra at 57 and GPT-6.1 Sol at 55. The 55 and the 46 come from different index versions and cannot be compared.

On Terminal-Bench 4.0, Artificial Analysis's own run measured 67%, which it calls "the highest score of any model," up from 13% for Kimi K3 and ahead of Claude Sonnet 5.5 at 64%. Anthropic's Sonnet 5.5 system card reports 70.6% on the same benchmark, run in Claude Code in --bare mode. Who runs the harness moves the ranking. On GDPval-AA, Artificial Analysis puts K3.1 Pro third at 1,846 Elo, up 312 from K3, behind Opus 5.5 (1,866) and Muse Wonder Full (1,857).

Against Claude Fable 5.2, the only overlap so far is between two vendor posts. Anthropic reports Fable 5.2 at 74.6% on Terminal-Bench-Science 0.1 and 74.2% on Humanity's Last Exam with tools. Moonshot reports K3.1 Pro at 38.1 and 70.0 on the same-named benchmarks, with different harnesses. The full matchup is on Kimi K3.1 vs Fable 5.2.

Throughput: What Kimi K3.1 Pro Actually Runs At

A 104B-active model that always thinks first has two speeds that matter: how fast tokens stream once they start, and how long you wait before the answer begins. The public measurements below use different methods, so each one names its source.

Kimi K3.1 Pro on Morph

Both numbers come from Morph's public hourly charts at ultra.morphllm.com: the median of hourly medians over the 24 hours ending November 9, 2026. Morph measures them itself. Morph is not one of the hosts Artificial Analysis tracks, so there is no third-party number to set beside them.

Hosted providers (Artificial Analysis)

Artificial Analysis tracks 15 hosts for Kimi K3.1 Pro. Read November 9, 2026, their end-to-end output speed ranges from 32 tokens per second on Moonshot's own API to 251 on LithosAI's ULTRA endpoint, about an 8x spread on the same weights. Its speed figure counts time to first token in the denominator. Time to the first answer token, which includes the model's thinking, ranges from 17 to 167 seconds. Time to the first token of any kind is 1.4 to 6.3 seconds. The per-host table is in the next section.

Self-hosted (vLLM recipe)

vLLM's recipe says plainly that its baseline layout, one tensor-parallel group of 8 GPUs, was not tuned for throughput. It does publish three measured tradeoffs on 8x B300:

• Long prompts: tensor plus expert parallelism prefills a single 340K-token prompt in about 10 seconds, against about 15 seconds at plain TP8 (measured on a pre-release checkpoint).
• Small batches: the low-latency decode path cuts the decode step by 1.6 to 1.7x at batch 1 and 1.2 to 1.3x at batch 32 to 64.
• Many long sessions: decode context parallelism (DCP8) holds about 18.5M tokens of KV cache against 2.6M at TP8 with FP8 KV, at the cost of 7 to 25% slower decode steps.

A K3.1 draft model for speculative decoding, Inferact/Kimi-K3.1-Pro-DSpark-preview, is on Hugging Face, and vLLM's MI355X agentic preset uses a K3.1 DSpark draft. See speculative decoding for why its batch-1 gains shrink as concurrency rises.

Which Provider Should You Use for Kimi K3.1?

Speed, cost per task and reliability do not line up across hosts. Pick on all three. The table is sorted by output speed.

Source: Artificial Analysis, Kimi K3.1 Pro providers. Output speed is end to end, with time to first token included. Cost per task is the weighted average cost of one Intelligence Index task at each host's prices. Morph is not in this set; see the throughput section for Morph's own measurements.

Four things fall out of the table. Moonshot's own standard endpoint is the slowest of the 15 at 32 tok/s, with a 167-second wait for the first answer token. Its Highspeed route runs 61 tok/s, about 1.9x faster on this measure, and has the highest error rate in the set at 9.1%. The two fastest hosts, LithosAI (ULTRA) at 251 tok/s and Systalyze at 227, cost $6.32 and $8.10 per task, two to two and a half times Modal's $3.19. And the cheapest host per task, PrimaLabs at $2.94, also has a 5.1% error rate. Modal (118 tok/s, 0.0% errors, $3.19 per task) is the closest thing to a balanced pick in this snapshot.

OpenRouter: check tool_choice, not just price

OpenRouter lists 24 endpoints for moonshotai/kimi-k3.1-pro, read November 9, 2026, and Morph is one of them. All 24 advertise the full 1,048,576-token context and support tools. Quantization labels vary from host to host (mxfp4, fp4, bf16 and "unknown" all appear). Input prices run from under $0.20 to $4.00 per million. The differences that break agents are in the declared tool support:

Morph's OpenRouter route declares all four tool_choice modes. Whichever host you pick, run your own prompt shape against it before committing volume.

The Token Economics: Always-On Thinking

K3.1 Pro always thinks. Moonshot's quickstart: "K3.1 Pro always has thinking mode enabled," and there is no switch to turn it off. You choose how much with reasoning_effort set to low, high or max, and the default is max. That default has a price. Artificial Analysis counted 55K output tokens per Intelligence Index task, 14% more than K3.

Max is not automatically the wrong default. Moonshot's Terminal-Bench 4.0 cost chart, built on tbench.ai data, puts K3.1 Pro at 52.8% for $4.16 per task on high effort and 65.7% for $5.25 per task on max. Going from high to max adds $1.09 per task and 12.9 points. The same chart plots Claude Opus 5.5 at 64.85% for $14.30 per task. That 64.85 is the chart's figure; the main table uses 66.4. Drop to high or low for small, frequent calls, not for the hard terminal tasks where max earns its cost.

Moonshot rates from the Kimi platform pricing page; Claude rates from Anthropic's pricing docs; Morph rates from the same constants that bill the API. Moonshot also bills cache writes: $2.00 per million for the default 5-minute TTL and $4.00 for the 1-hour TTL.

Two readings of that table. K3.1 Pro's list price matches Claude Sonnet 5.5's ($2/$10), and on Artificial Analysis's launch-day index the two sat one point apart, 55 to 56. And Moonshot's own API is cheaper per token than Morph: $2.50 vs $2.00 input and $14.00 vs $10.00 output. If per-token list price is the only thing you optimize, buy from Moonshot.

Morph's case is elsewhere. First, measured speed: about 115 tok/s on Morph's charts, against 32 tok/s for Moonshot's endpoint on Artificial Analysis. Those are different methods, so test with your own prompts. Second, 50% off any K3.1 request sent with service_tier: "standby". Third, one key for K3.1 Pro, DeepSeek V4.1 Flash and GLM-5.3. Fourth, the Ultra plan: a flat $200 per month with unlimited Kimi K3.1 and DeepSeek V4.1 Flash and no token meter. Each subscription is rate-limited to 24,290 uncached input and 6,162 output tokens per minute on K3.1 (701,079 and 76,559 on V4.1 Flash), and the plan is capped at 1,000 subscriptions. It pays off once your monthly K3.1 usage at per-token rates would pass $200.

On Moonshot, caching is automatic with a 5-minute default TTL, and partial cache blocks bill as misses. On Morph, pass a per-conversation prompt_cache_key in the request body or an x-session-id header so repeated prefixes bill at $0.29 per million; see the caching docs. Neither API tiers price by context length.

Is Kimi K3.1 Free?

Not through an API. Moonshot's quickstart: "Kimi K3.1 Pro is a flagship model: it is unlocked after a successful top-up (minimum $1). Your cumulative top-up amount also determines your account tier and rate limits." OpenRouter's catalog, read November 9, 2026, has no free K3.1 Pro route. Every host in the provider table charges.

The weights are free. The Hugging Face repo is ungated, and the license grants use, modification and redistribution with no fee. Its two conditions start at $20 million in revenue or 100 million monthly users, and internal use is exempt from both. What you pay for is hardware: a 1.56 TB download and, on vLLM's baseline, eight B300-class GPUs. The local section below has the details.

The closest thing to unmetered hosted use on Morph is the $200/month Ultra plan: no token meter on K3.1 or V4.1 Flash, within the per-minute limits above.

Kimi K3.1 API: Pricing and How to Call It

On Morph, Kimi K3.1 Pro is morph-kimik31 at https://api.morphllm.com/v1 (OpenAI-compatible) and https://api.morphllm.com (Anthropic Messages). Create a key at dashboard/api-keys. The same key works for every model on Morph's model list. Full setup per client is on the Kimi K3.1 API guide.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.morphllm.com/v1",
    api_key="YOUR_MORPH_API_KEY",
)

resp = client.chat.completions.create(
    model="morph-kimik31",
    reasoning_effort="high",  # low | high | max (default max)
    messages=[
        {"role": "user", "content": "Find the race condition in this worker pool."},
    ],
)
print(resp.choices[0].message.content)

Claude Code, Codex and Morph's agent setup

Claude Code talks to Morph's Anthropic Messages endpoint with four environment variables:

export ANTHROPIC_BASE_URL="https://api.morphllm.com"
export ANTHROPIC_AUTH_TOKEN="YOUR_MORPH_API_KEY"
export ANTHROPIC_MODEL="morph-kimik31"
export ANTHROPIC_SMALL_FAST_MODEL="morph-dsv41flash"
claude

Codex uses the Responses wire API. In ~/.codex/config.toml, with MORPH_API_KEY exported:

model = "morph-kimik31"
model_provider = "morph"

[model_providers.morph]
name = "Morph"
base_url = "https://api.morphllm.com/v1"
env_key = "MORPH_API_KEY"
wire_api = "responses"

Morph's agent setup command:

curl -fsSL https://www.morphllm.com/install.sh | bash && morph install

Morph also documents three features for agent workloads on K3.1: dynamic tool loading, Agent Runs keyed by a run_id, and session caching with prompt_cache_key or x-session-id.

Moonshot's first-party API

Moonshot serves kimi-k3.1-pro at https://api.moonshot.ai/v1 (OpenAI-compatible) and https://api.moonshot.ai/anthropic (Anthropic Messages, with streaming, tool use, image input, thinking and structured output). Its quickstart documents strict JSON Schema output, a partial mode for prefix continuation, its own dynamic tool loading (a tool definition placed in a system message with no content becomes available from that message on), and image input as base64 or ms:// file IDs. Public image URLs are not supported. It says its built-in web search "is being updated and is not recommended for production workflows in the near term." Zero Data Retention is available for API customers.

messages = [{"role": "user", "content": "Find the bug in this function: ..."}]

first = client.chat.completions.create(model="morph-kimik31", messages=messages)
assistant = first.choices[0].message

# Append the complete assistant message, reasoning and tool calls included.
# Do NOT rebuild it as {"role": "assistant", "content": assistant.content}.
messages.append(assistant)
messages.append({"role": "user", "content": "Now write the failing test for it."})

second = client.chat.completions.create(model="morph-kimik31", messages=messages)
print(second.choices[0].message.content)

Tool calls, sampling and the request checklist

Moonshot's quickstart gives two more rules. For tools: "Use tool_choice="required" on the first turn to require at least one tool call," then return one tool result per call with the matching tool_call_id. For sampling: temperature=1.0, top_p=0.95, n=1, presence_penalty=0 and frequency_penalty=0 are fixed on Moonshot's API, and the docs say to omit them. Many OpenAI-compatible clients send temperature: 0 by default.

Before you decide K3.1 is slow or weak in your agent, log one outgoing request body and check four things: reasoning_effort is set to what you meant, temperature is absent, max_tokens is a number you chose, and the previous assistant message went back with its reasoning and tool calls.

How to Run Kimi K3.1 Pro Locally

"Locally" means a GPU server. The Hugging Face repo holds 96 safetensors shards totaling 1.56 TB, and vLLM's recipe runs it on one tensor-parallel group of 8 GPUs. It needs vLLM 0.34.0 or later, which the recipe says is not released yet and ships as a nightly wheel, or the vllm/vllm-openai:kimi-k31 Docker image. The single-node B300 command from the recipe:

export VLLM_ENGINE_READY_TIMEOUT_S=3600
vllm serve moonshotai/Kimi-K3.1-Pro \
  --trust-remote-code \
  --tensor-parallel-size 8 \
  --load-format fastsafetensors \
  --enable-auto-tool-choice \
  --tool-call-parser kimi_k3 \
  --reasoning-parser kimi_k3

GGUF and community builds

As of November 9, 2026, Unsloth's own unsloth/Kimi-K3.1-Pro repo is a 1.56 TB safetensors mirror, not a GGUF. The Unsloth-style dynamic GGUFs published so far sit in a preliminary repo, danielhanchen/Kimi-K3.1-Pro-GGUF-Preliminary, with two builds: UD-Q4_K_XL at 1,509 GB and UD-Q8_K_XL at 1,561.5 GB. There is no 1-bit or 2-bit build in it yet. Other community GGUF, MLX and pruned uploads exist on Hugging Face; we have not evaluated them.

So the practical floor today is about 1.5 TB of combined memory for llama.cpp, or 8 datacenter GPUs for vLLM. The 1.56 TB download also needs disk and bandwidth before anything runs. Sizing a dedicated deployment against a speed and concurrency target is a conversation for Morph's private deployments team.

Open Weights and the Kimi K3.1 Pro License

Moonshot's release post promised that "full model weights will be released by Nov 4, 2026." The moonshotai/Kimi-K3.1-Pro repo was created on November 2, 2026, last modified November 5, and is not gated. It is the only K3.1 repo in the moonshotai organization; there are no Highspeed weights.

The license is the Kimi K3.1 Pro License (the LICENSE file is headed "Kimi K3.1-Pro License"). The base grant is MIT-style: "the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software; to run, deploy, fine-tune, or otherwise modify the Software." Two conditions sit on top:

• Model as a Service above $20M. If the licensee or its affiliates run a Model-as-a-Service business and their aggregate revenue exceeds $20 million over any consecutive 12 months, they "must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose." The license defines Model as a Service as giving a third party access to inference or fine-tuning "in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data."
• Attribution at scale. A product with "more than 100 million monthly active users, or more than 20 million US dollars ... in monthly revenue" must prominently display "Kimi K3.1-Pro" in its user interface.

Neither condition applies to "(a) internal use of the Software ... or (b) any use of the Software accessed through Moonshot AI's official products or certified inference partners." Self-hosting for internal use triggers neither condition.

Kimi K3.1 Pro vs GLM-5.3 and DeepSeek V4.1 Flash

These are the three open-weight models Morph serves on one key. Moonshot's own table includes GLM-5.3 and DeepSeek V4.1 Flash-1017 columns, so the benchmark rows below come from one source. Specs, licenses and list prices come from each vendor's own docs.

The row to read twice is DeepSWE v1.1. In Moonshot's own table, DeepSeek V4.1 Flash-1017 beats K3.1 Pro, 77.0 to 75.5, and on Morph it costs $0.12 input and $0.80 output against $2.50 and $14.00. For short, scoped coding tasks that is the better default. K3.1 Pro earns its price on long-horizon work: Terminal-Bench 4.0, SWE Marathon, HLE with tools and office documents, where it leads both by 7.5 to 33.3 points. GLM-5.3 is text-only and trails both on most rows here. Z.ai's own card reports Terminal-Bench 2.1 and 3.0 instead, and those versions are not comparable to 4.0.

Deeper matchups: GLM-5.3 vs Kimi K3.1, DeepSeek V4.1 Flash, DeepSeek V4.1 Flash 1017 and Kimi K3.1 vs Claude.

Kimi K3.1 Pro: Pros and Cons

When to Use Kimi K3.1 Pro (and When Not)

Frequently Asked Questions