Kimi (Moonshot AI)

Kimi (Moonshot AI model family) is a family of large language models and multimodal foundation models developed by the Beijing-based Chinese company Moonshot AI. The family includes models for conversational generation, long-context document processing, mathematical and software-engineering tasks, visual understanding, and agentic systems that plan multi-step actions and invoke external tools.

The Kimi K2 branch uses a sparsely activated Mixture-of-Experts (MoE) Transformer architecture with approximately 1.04 trillion total parameters and 32.6 billion parameters activated per token; Moonshot AI model cards conventionally round these figures to 1 trillion and 32 billion. The original July 2025 Kimi K2 release had a 128,000-token context window. Kimi-K2-Instruct-0905 and subsequent K2-family releases increased the supported context to 256,000 tokens.

Kimi K2.5 introduced native image and video understanding, joint text–vision training, and Agent Swarm, a framework for parallel orchestration of dynamically created sub-agents. Kimi K2.6, released on 20 April 2026, extended the same multimodal architecture with stronger long-horizon coding, autonomous execution, and larger-scale swarm orchestration. Kimi K2.7 Code, released on 12 June 2026, is a coding-specialized agentic model built on K2.6; Moonshot AI reports that it uses approximately 30% fewer thinking tokens than K2.6 on its evaluations.

As of 10 July 2026, Moonshot AI describes K2.6 as its general-purpose multimodal Kimi model and K2.7 Code as its most capable coding model. The hosted kimi-k2.7-code-highspeed service uses the same model as K2.7 Code but a faster serving configuration; it is not a separate checkpoint. There is no separately documented general-purpose “K2.7” checkpoint in the official model list at that date.

The weights of the principal K2 checkpoints are distributed through Hugging Face under a modified MIT licence, while Moonshot AI also offers proprietary hosted access through the Kimi API Platform. The licence permits use, modification, distribution, sublicensing, and commercial use, but requires prominent display of the applicable Kimi model name when a derivative commercial product or service exceeds either 100 million monthly active users or US$20 million in monthly revenue. The term open-weight is therefore more precise than implying that the full training dataset and end-to-end training pipeline have been released.

History and background

Foundation of Moonshot AI

Moonshot AI was founded in Beijing in March 2023 by Yang Zhilin, Zhou Xinyu, and Wu Yuxin, who had academic ties to Tsinghua University. Yang, the company's chief executive, had previously conducted language-model research and worked at Google and Meta.

In February 2024, the company raised more than US$1 billion in a financing round led by Alibaba and HongShan, reportedly valuing Moonshot AI at approximately US$2.5 billion. In May 2026, media reports described a further US$2 billion round involving investors including Meituan's venture arm and China Mobile, at a valuation above US$20 billion. These figures were reported financing terms rather than audited operating metrics.

Release chronology

• October 2023 — Moonshot AI launched the first Kimi chatbot. The product was advertised as accepting approximately 200,000 Chinese characters per conversation, commonly associated with an approximately 128,000-token model context. Chinese-character counts and tokenizer token counts are different measures and should not be treated as interchangeable.
• March 2024 — a beta version of the Kimi product was advertised as accepting prompts of up to two million Chinese characters, up from 200,000 characters. This was a product-level character limit, not a claim of a two-million-token context window.
• 26 June 2024 — Moonshot AI published Mooncake, a KV-cache-centric disaggregated architecture for serving long-context language models.
• 20 January 2025 — release of Kimi k1.5, a multimodal reasoning model trained with scaled reinforcement learning and a context of up to 128,000 tokens. The accompanying report stated that the model matched OpenAI o1 on selected mathematical, coding, and multimodal benchmarks; it did not establish general equivalence across all tasks.
• 10 April 2025 — release of Kimi-VL, an open-weight multimodal MoE model with approximately 16 billion total language-model parameters, 2.8 billion active language-model parameters, a native-resolution MoonViT visual encoder, and a 128,000-token context window.
• 11 July 2025 — release of Kimi K2, an open-weight text-only MoE model with approximately 1.04 trillion total parameters, 32.6 billion active parameters, and an original 128,000-token context. The technical report followed on 28 July 2025.
• 5 September 2025 — release of Kimi-K2-Instruct-0905, which extended the context from 128,000 to 256,000 tokens and introduced improvements in agentic and front-end coding.
• 30 October 2025 — publication of Kimi Linear, a separate 48-billion-parameter hybrid linear-attention research model with three billion active parameters and support for contexts of up to one million tokens.
• 6 November 2025 — release of Kimi K2 Thinking, a reasoning-oriented version with a 256,000-token context, interleaved reasoning and tool invocation, and native INT4 quantization.
• 27 January 2026 — release of Kimi K2.5, a natively multimodal K2-family model supporting text, images, and video, with both thinking and instant generation modes.
• 9 February 2026 — public research-preview documentation for K2.5 Agent Swarm described product configurations with up to 100 parallel sub-agents and more than 1,500 tool calls.
• 20 April 2026 — release of Kimi K2.6, retaining the K2.5 multimodal architecture while emphasizing long-horizon software engineering, visual interface generation, proactive execution, and expanded Agent Swarm operation.
• 25 May 2026 — Moonshot AI discontinued the older hosted K2 API endpoints, including the July 2025, September 2025, and K2 Thinking services. The downloadable open-weight checkpoints remained available; the deprecation applied to the hosted API endpoints.
• 12 June 2026 — release of Kimi K2.7 Code, a coding-focused agentic model derived from K2.6. The accompanying model card reports improved completion of long-horizon software-engineering tasks and approximately 30% lower thinking-token usage than K2.6.

Technical foundations and architecture

Mixture-of-Experts architecture

Kimi K2 and its descendants use a Transformer architecture in which the feed-forward components of most layers are implemented as sparsely activated experts. A simplified dense Transformer block may be written as:

${\mathbf{𝐇}}^{\prime }=\mathrm{MHA}(\mathbf{𝐇})+\mathbf{𝐇},{\mathbf{𝐇}}^{*}=\mathrm{MLP}({\mathbf{𝐇}}^{\prime })+{\mathbf{𝐇}}^{\prime },$

where $\mathbf{𝐇}\in {\mathbb{ℝ}}^{L\times d}$ contains the input representations, $L$ is the sequence length, and $d$ is the hidden-state dimension.

In an MoE layer, the single feed-forward network is replaced with experts $\{E_i{\}}_{i=1}^N$. A router selects a subset of experts for each token. In simplified notation:

$\mathrm{MoE}(\mathbf{𝐡})=\sum _{i\in {𝒮}(\mathbf{𝐡})}{g}_i(\mathbf{𝐡})E_i(\mathbf{𝐡}),$

where ${𝒮}(\mathbf{𝐡})$ is the selected expert set and the routing weights satisfy:

$\sum _{i\in {𝒮}(\mathbf{𝐡})}{g}_i(\mathbf{𝐡})=1.$

A common abstraction of the routing operation is:

$g(\mathbf{𝐡})=\mathrm{TopK}(\mathrm{Softmax}(\mathbf{𝐡}{W}_g)),$

where $W_g$ is a learned routing matrix.

K2 has 384 routed experts, of which eight are selected for each token, together with one shared expert. The ratio $384/8=48$ describes routed-expert sparsity, but it should not be interpreted as a 48-fold reduction in total inference cost: the shared expert, attention layers, embeddings, routing operations, communication, and memory movement also consume compute and bandwidth.

Architectural parameters of the K2 family

Parameter	K2-family value or qualification
Architecture	Transformer with Mixture-of-Experts
Total parameters	Approximately 1.04 trillion; rounded to 1 trillion in model cards
Activated parameters per token	Approximately 32.6 billion; rounded to 32 billion in model cards
Number of layers	61: one dense layer and 60 MoE layers
Hidden-state dimension	7,168
Attention heads	64
Routed experts	384
Experts selected per token	8 routed experts, plus 1 shared expert
Expert hidden dimension	2,048
Vocabulary size	160,000 tokens
Activation function	SwiGLU
Attention mechanism	Multi-head Latent Attention (MLA)
Original K2 context	128,000 tokens
K2-0905 and later K2-family context	256,000 tokens
Vision encoder in K2.5, K2.6, and K2.7 Code	MoonViT, approximately 400 million parameters
Native modalities in K2.5 and later multimodal checkpoints	Text, images, and video

The architecture remained substantially consistent from K2.5 through K2.7 Code. The main differences among these releases are therefore attributable primarily to continued pre-training, post-training, reinforcement learning, data construction, tool-use training, and product-level agent orchestration rather than to a newly disclosed backbone topology.

Multi-head Latent Attention

The K2 family uses Multi-head Latent Attention (MLA), an attention design that compresses the representations used to construct keys and values, reducing the size of the key–value cache relative to conventional multi-head attention.

Standard scaled dot-product attention is:

$\mathrm{Attention}(\mathbf{𝐐},\mathbf{𝐊},\mathbf{𝐕})=\mathrm{softmax}\left(\frac{\mathbf{𝐐}{\mathbf{𝐊}}^{\mathsf{T}}}{\sqrt{d_k}}\right)\mathbf{𝐕},$

where $\mathbf{𝐐}$, $\mathbf{𝐊}$, and $\mathbf{𝐕}$ are query, key, and value matrices. MLA uses lower-dimensional latent representations and learned projections to reconstruct head-specific components. This reduces KV-cache memory, which is particularly relevant for long-context decoding, although the exact end-to-end benefit depends on implementation, sequence length, batch size, and hardware.

MuonClip and QK-Clip

Kimi K2 was trained with MuonClip, a modification of the Muon optimizer that combines Muon's matrix-oriented update procedure with QK-Clip, a stabilisation mechanism intended to prevent extreme attention logits during large-scale training.

For an attention head $h$, the monitoring statistic can be expressed as:

$S_h^{\max }=\underset{b,i,j}{\max }\left(\frac{{\mathbf{𝐪}}_{b,h,i}{\mathbf{𝐤}}_{b,h,j}^{\mathsf{T}}}{\sqrt{d_h}}\right),$

with a scaling coefficient:

${\gamma }_h=\min (1,\frac{\tau }{S_h^{\max }}),$

where the K2 report uses $\tau =100$.

Despite its name, QK-Clip is not described as directly clipping every attention logit during the same forward pass. Instead, excessive per-head logits are used as a signal for post-update rescaling of the relevant query and key projection weights. In K2's MLA implementation, the content-related query and key components are scaled by factors derived from ${\gamma }_h$, while the shared rotary-key component is handled separately. This distinction is important because direct logit clipping and projection-weight rescaling have different optimisation behaviour.

Moonshot AI reported that K2 completed pre-training on 15.5 trillion tokens without an observed loss spike. This is a developer-reported result from a single training programme; public evidence does not independently isolate how much of the stability came from QK-Clip rather than the rest of the optimiser, data pipeline, numerical controls, or distributed-training system.

Multimodality and MoonViT

Kimi-VL introduced the MoonViT native-resolution visual encoder. Its design uses variable-resolution image packing rather than forcing every input to a single fixed resolution and was initialised from the SigLIP-SO-400M line of visual encoders.

For an image $\mathbf{𝐗}$, a simplified multimodal projection may be written as:

${\mathbf{𝐙}}_{\mathrm{v}\mathrm{i}\mathrm{s}}={\mathrm{\Phi }}_{\mathrm{v}\mathrm{i}\mathrm{s}}(\mathbf{𝐗}),$

${\mathbf{𝐇}}_{\mathrm{v}\mathrm{i}\mathrm{s}}={\mathrm{MLP}}_{\mathrm{p}\mathrm{r}\mathrm{o}\mathrm{j}}\left({\mathbf{𝐙}}_{\mathrm{v}\mathrm{i}\mathrm{s}}\right),$

where ${\mathrm{\Phi }}_{\mathrm{v}\mathrm{i}\mathrm{s}}$ is the visual encoder and the projector maps visual features into the language model's hidden space. The projected visual tokens are then combined with text tokens for processing by the Transformer.

K2.5 uses a video-capable MoonViT variant, referred to in the technical report as MoonViT-3D. The model card lists approximately 400 million visual-encoder parameters. The architecture is retained by K2.6 and K2.7 Code.

The native model modalities are text, images, and video. PDF support in the Kimi product and API is implemented through file upload, document parsing, and conversion of pages or extracted content into model inputs; a PDF container is not a separate neural input modality of the base checkpoint.

Pre-training and post-training

Kimi K2 pre-training

Kimi K2 was pre-trained on approximately 15.5 trillion tokens spanning web text, source code, mathematics, and knowledge-oriented data. Moonshot AI reported the use of MuonClip throughout the run and stated that no loss spike was observed.

The public report describes data categories and training methods at a high level but does not release the complete training corpus, all filtering rules, language distribution, deduplication pipeline, or the full distributed-training configuration. Consequently, the training run cannot be reproduced exactly from the public artefacts alone.

Kimi K2.5 continual pre-training

Kimi K2.5 was developed through continual pre-training from a Kimi K2 checkpoint rather than by training an unrelated model from scratch. The published training stages include:

• approximately one trillion tokens of standalone visual-encoder training;
• approximately 15 trillion mixed text and visual tokens for joint pre-training;
• a long-context mid-training stage of approximately 500 billion tokens at a 32,000-token sequence length;
• a further approximately 200 billion tokens at a 256,000-token sequence length.

The technical report presents K2.5 as an early-fusion system in which visual and textual data are jointly included during pre-training. It does not describe a simple schedule in which the vision-token share is merely increased monotonically throughout the run. Its ablation results instead favour introducing visual data early with a moderate visual-token ratio over injecting a high proportion of visual data only at a late stage.

Supervised fine-tuning and reinforcement learning

The K2 post-training pipeline combines several stages:

Supervised fine-tuning (SFT):

• instruction-following and conversational data;
• code, mathematics, and reasoning examples;
• synthetic agent trajectories generated from tool specifications and simulated or executable environments;
• verification and filtering of tool-call results.

Zero-vision SFT for K2.5:

• a text-only supervised-fine-tuning phase performed after joint vision–language pre-training;
• intended to preserve and activate multimodal agentic capabilities without requiring every post-training example to include an image;
• followed by multimodal reinforcement learning rather than serving as the final visual-training stage.

Reinforcement learning:

• verifiable-reward training for mathematics, code, and other tasks with automatically checkable outcomes;
• preference and rubric-based rewards for less directly verifiable agentic tasks;
• interactions with real and synthetic environments;
• joint text–vision reinforcement learning for K2.5.

The K2.5 report states that visual reinforcement learning also improved several text-only evaluation results in its ablation experiments. These gains are reported by the model developers and do not by themselves establish that visual data will improve every textual task or deployment setting.

Quantization-aware training

Kimi K2 Thinking introduced a native weight-only INT4 checkpoint produced through quantization-aware training. K2.5, K2.6, and K2.7 Code use the same general native-INT4 approach, with a group size of 32 described in the K2.5 technical materials.

Native INT4 reduces weight-storage and memory-bandwidth requirements relative to higher-precision checkpoints. It does not make the trillion-parameter model equivalent to a 32-billion-parameter dense model in storage: all experts must still be stored or distributed, even though only a subset is activated for each token.

Agent Swarm

Agent Swarm is Moonshot AI's framework for self-directed parallel agent orchestration. A primary model decomposes a request into heterogeneous subtasks, creates or assigns specialised sub-agents, allows them to invoke tools concurrently, and synthesises their results.

The K2.5 technical report describes Parallel-Agent Reinforcement Learning (PARL), which separates execution from optimisation. In the reported setup, the orchestrator is trainable while sub-agents can be held fixed, and sub-agent trajectories are not treated identically to the orchestrator trajectory during optimisation. Moonshot AI reported up to a 4.5-fold latency reduction relative to a sequential single-agent baseline on highly parallelisable search tasks.

The February 2026 K2.5 product preview advertised configurations with up to 100 sub-agents and more than 1,500 tool calls. The April 2026 K2.6 product description increased the advertised orchestration scale to as many as 300 sub-agents and approximately 4,000 coordinated steps. These are product limits and demonstrations, not a claim that every request benefits from maximum parallelism. Sequential tasks, tightly coupled subtasks, rate limits, duplicated work, and synthesis overhead can reduce or eliminate the advantage.

Main models

Model	Type	Parameters	Maximum context	Native inputs	Release
Kimi k1.5	Multimodal reasoning model trained with scaled RL	Not publicly disclosed	128,000 tokens	Text and images	20 January 2025
Kimi-VL	Multimodal MoE VLM	Approximately 16B total language-model parameters / 2.8B active, plus visual encoder	128,000 tokens	Text, images, and video	10 April 2025
Kimi K2	Text-only MoE LLM	Approximately 1.04T total / 32.6B active	128,000 tokens	Text	11 July 2025
Kimi-K2-Instruct-0905	Updated text-only MoE LLM	Approximately 1.04T / 32.6B	256,000 tokens	Text	5 September 2025
Kimi K2 Thinking	Reasoning and tool-use MoE LLM	Approximately 1.04T / 32.6B	256,000 tokens	Text	6 November 2025
Kimi K2.5	Native multimodal agentic MoE model	Approximately 1.04T / 32.6B; approximately 400M vision encoder	256,000 tokens	Text, images, and video	27 January 2026
Kimi K2.6	Native multimodal general-purpose and agentic MoE model	Approximately 1.04T / 32.6B; approximately 400M vision encoder	256,000 tokens	Text, images, and video	20 April 2026
Kimi K2.7 Code	Coding-specialised multimodal agentic MoE model	Approximately 1.04T / 32.6B; approximately 400M vision encoder	256,000 tokens	Text, images, and video	12 June 2026

Kimi k1.5

Kimi k1.5 is a multimodal reasoning model developed to study scaling reinforcement learning and inference-time reasoning. Its report describes long-context reinforcement learning, improved policy optimisation, and “long-to-short” distillation methods that transfer some capabilities of long reasoning traces to shorter outputs.

Moonshot AI reported 77.5 on AIME, 96.2 on MATH 500, a 94th-percentile Codeforces result, and 74.9 on MathVista in its long-reasoning configuration. The paper's statement that the model matched OpenAI o1 refers to selected reported benchmarks and evaluation settings, not to an independently established equivalence across all capabilities, safety properties, languages, or deployment scenarios.

Kimi-VL

Kimi-VL is an open-weight multimodal MoE model designed for image and video understanding, optical character recognition, visual grounding, document analysis, screen interaction, and multimodal reasoning. Its language decoder activates approximately 2.8 billion parameters per token, while its visual front end uses the native-resolution MoonViT encoder.

The family includes an instruction-following checkpoint and Kimi-VL-Thinking, a variant post-trained for longer multimodal reasoning with supervised fine-tuning and reinforcement learning. Kimi-VL is architecturally and historically distinct from the later trillion-parameter K2.5 branch.

Kimi K2

Kimi K2 is the first trillion-parameter open-weight model in the K2 line. It has approximately 1.04 trillion total parameters, 32.6 billion active parameters, 384 routed experts, and an original context limit of 128,000 tokens.

The technical report describes a multi-stage post-training pipeline based on large-scale synthesis of agentic trajectories, interaction with real and synthetic environments, and reinforcement learning for tool use, coding, mathematical reasoning, and instruction following. The July 2025 model was a non-thinking checkpoint: it was not designed to emit an extended reasoning trace before every answer.

Kimi-K2-Instruct-0905 retained the same backbone while increasing the context to 256,000 tokens and improving coding-agent and front-end-generation performance. Moonshot AI's hosted July and September 2025 K2 API endpoints were discontinued on 25 May 2026, although their open-weight checkpoints remain downloadable.

Kimi K2 Thinking

Kimi K2 Thinking was post-trained to interleave model-generated reasoning with tool calls over long agentic episodes. Its model card reports stable operation over approximately 200–300 sequential tool invocations in the developer's test settings.

APIs and inference frameworks may expose generated reasoning text through a field such as reasoning_content. This field is an output generated by the model and should not be treated as guaranteed access to an otherwise hidden or complete internal computational process.

K2 Thinking supports a 256,000-token context and was released as a native INT4 checkpoint. Its former hosted Kimi API endpoint was among the older K2 endpoints discontinued on 25 May 2026.

Kimi K2.5

Kimi K2.5 is a natively multimodal continuation of Kimi K2. It was trained on approximately 15 trillion additional mixed text and visual tokens and supports text, image, and video inputs within a 256,000-token context.

The model supports two principal response modes:

• Thinking mode — produces an explicit model-generated reasoning stream and supports long-horizon agentic generation;
• Instant mode — suppresses the explicit thinking stream and targets lower latency.

K2.5 combines visual understanding with software engineering, document analysis, graphical-interface interaction, and tool use. The official API can support uploaded files, including PDFs, but PDF ingestion is performed by the surrounding file-processing system rather than by adding “PDF” as a separate base-model modality.

Kimi K2.6

Kimi K2.6 retains the same disclosed architecture as K2.5 but uses further post-training to improve long-horizon coding, front-end and lightweight full-stack generation, visual debugging, proactive agents, and swarm orchestration.

Moonshot AI demonstrated K2.6 in extended software-engineering runs involving thousands of tool calls and many hours of execution. These demonstrations indicate the intended operating regime but remain developer-selected case studies rather than controlled estimates of success rates across arbitrary repositories.

K2.6 supports both thinking and non-thinking modes and is listed by Moonshot AI as the general-purpose Kimi model for agentic, coding, visual, and conversational work as of July 2026.

Kimi K2.7 Code

Kimi K2.7 Code is a coding-focused agentic checkpoint built on K2.6. It retains the approximately one-trillion-parameter MoE architecture, 256,000-token context, MoonViT visual encoder, native INT4 quantization, and support for interleaved tool use.

Unlike K2.6 and K2.5, K2.7 Code is documented as a thinking-only model: instant mode is not supported, and the official API preserves the generated thinking state for multi-step interactions. Video chat through third-party self-hosted inference stacks was described as experimental at release, with full support initially limited to Moonshot AI's hosted API.

The hosted kimi-k2.7-code-highspeed endpoint uses the same underlying model but allocates a faster serving tier. Moonshot AI documents typical output around 180 tokens per second and up to approximately 260 tokens per second in short-context cases, subject to service load. These are serving-speed figures, not changes to the model's weights or intelligence.

Reported benchmark results

Benchmark scores in this section are developer-reported unless otherwise stated. Results can depend substantially on the evaluation harness, sampling parameters, context-management policy, tool implementation, step budget, judge model, repository state, and baseline version. Scores from different release reports should therefore not be interpreted as a single controlled longitudinal experiment.

Kimi K2 and K2-0905

The following Kimi K2-Instruct results were reported in the K2 technical report without extended thinking:

Benchmark	Kimi K2-Instruct	Reported setting
SWE-bench Verified	65.8%	Agent harness, Pass@1
SWE-bench Multilingual	47.3%	Agent harness
LiveCodeBench v6	53.7%	Pass@1
Tau2-Bench	66.1%	Micro-average
ACEBench English	76.5%	English subset
AIME 2025	49.5%	Avg@64
GPQA-Diamond	75.1%	Avg@8

For Kimi-K2-Instruct-0905, Moonshot AI reported 69.2% on SWE-bench Verified, 55.9% on SWE-bench Multilingual, and 44.5% on Terminal-Bench. The coding results were generally averaged over five full-dataset runs, and the model card notes that the evaluation harness modified tool contexts and prompts relative to some standard implementations.

Kimi-VL

The Kimi-VL report provides separate results for the base instruction-following and thinking variants. Selected developer-reported results include:

Benchmark	Score	Variant or qualification
LongVideoBench	64.5	Kimi-VL
MMLongBench-Doc	35.1	Kimi-VL
InfoVQA	83.2	Kimi-VL
ScreenSpot-Pro	34.5	Kimi-VL
MMMU	61.7	Kimi-VL-Thinking
MathVision	36.8	Kimi-VL-Thinking
MathVista	71.3	Kimi-VL-Thinking

These figures should not be compared directly with later K2.5 results without accounting for different model scales, checkpoints, prompts, and evaluation dates.

Kimi K2.5 and K2.6

Selected results reported by Moonshot AI for K2.5 and K2.6 are shown below. Both columns use thinking mode, but individual benchmark harnesses and context-management rules vary.

Category	Benchmark	Kimi K2.5	Kimi K2.6
Agentic reasoning	HLE-Full with tools	50.2	54.0
Agentic search	BrowseComp with context management	74.9	83.2
Agentic search	BrowseComp with Agent Swarm	78.4	86.3
Agentic search	DeepSearchQA, F1	89.0	92.5
Agentic search	DeepSearchQA, accuracy	77.1	83.0
Coding	Terminal-Bench 2.0	50.8	66.7
Coding	SWE-bench Pro	50.7	58.6
Coding	SWE-bench Multilingual	73.0	76.7
Coding	SWE-bench Verified	76.8	80.2
Coding	LiveCodeBench v6	85.0	89.6
Reasoning	GPQA-Diamond	87.6	90.5
Vision	MMMU-Pro	78.5	79.4
Vision	MathVision	84.2	87.4

For K2.5, the model card reports temperature 1.0, top-p 0.95, and a context of 256,000 tokens unless otherwise specified. K2.6 used temperature 1.0, top-p 1.0, and a 262,144-token context for most reported experiments. Some competitor scores in both reports were re-evaluated by Moonshot AI rather than copied from an official external result.

Kimi K2.7 Code

The K2.7 Code model card reports the following selected scores:

Category	Benchmark	Kimi K2.6	Kimi K2.7 Code
Coding	Kimi Code Bench v2	50.9	62.0
Coding	Program Bench	48.3	53.6
Coding	MLS Bench Lite	26.7	35.1
Agentic	Kimi Claw 24/7 Bench	42.9	46.9
Agentic tool use	MCP Atlas	69.4	76.0
Agentic tool use	MCPMark Verified	72.8	81.1

Kimi Code Bench v2 and Kimi Claw 24/7 Bench are Moonshot AI internal or developer-curated evaluations. At the publication of the K2.7 Code model card, the human-verified MCPMark edition was described as planned for public release. These scores therefore offer evidence about the developer's target workloads but have less independent reproducibility than fully public benchmarks with fixed artefacts and third-party runs.

Applications

Conversational assistance, search, and document analysis

Kimi products are used for conversational assistance, web-assisted research, extraction and synthesis from long documents, comparison of reports, and production of structured written outputs. The long context windows allow large document collections or extended tool histories to be processed, although retrieval, context selection, and file parsing remain necessary when the source collection exceeds the model limit.

File-based question answering can accept office documents and PDFs through the Kimi API workflow. The surrounding platform extracts text, images, or page representations before model processing.

Software engineering

Kimi K2 and later releases target repository-level software-engineering tasks, including:

• locating and correcting defects;
• implementing features across multiple files;
• generating and refactoring code;
• executing tests and shell commands;
• analysing build or runtime failures;
• front-end implementation from textual or visual specifications;
• performance profiling and optimisation;
• integration with coding agents and tool protocols.

K2.7 Code is specifically optimised for long-horizon coding agents. It can be integrated with compatible coding frameworks through OpenAI-style or Anthropic-style APIs or deployed locally using supported inference engines.

Multimodal applications

Kimi-VL, K2.5, K2.6, and K2.7 Code support visual question answering, document-image understanding, OCR, chart and diagram interpretation, video analysis, interface reconstruction, and vision-grounded coding.

K2.6 materials describe “coding-driven design”, in which the model uses textual and visual inputs together with code-execution, image-generation, or browser tools to produce interactive interfaces. The quality of such systems depends on the external harness and execution environment as well as on the model itself.

Multi-agent research and production workflows

Agent Swarm is intended for workloads that can be divided into largely independent tasks, such as broad web search, collection of many documents, parallel review of multiple files, comparison from several professional perspectives, and generation of multi-section reports.

Parallel execution can reduce wall-clock latency, but it can also increase token use, tool cost, duplicated searches, and synthesis complexity. A higher number of sub-agents is therefore not inherently equivalent to better output.

API access and local deployment

As of July 2026, the principal multimodal models listed by the Kimi API Platform are:

• kimi-k2.7-code;
• kimi-k2.7-code-highspeed;
• kimi-k2.6;
• kimi-k2.5.

All are documented with a 256,000-token context. The platform also retains older Moonshot V1 generation models with 8,000-, 32,000-, and 128,000-token contexts. The hosted July 2025, September 2025, and K2 Thinking endpoints are deprecated.

The Kimi API supports OpenAI-compatible interfaces, and current K2-family model cards also document Anthropic-compatible access. Available functions depend on the selected model and include tool calls, streaming, visual inputs, and model-generated thinking streams.

For self-hosting K2.5, K2.6, and K2.7 Code, Moonshot AI recommends:

• vLLM;
• SGLang;
• KTransformers.

The earlier K2-0905 documentation additionally listed TensorRT-LLM.

Even with MoE activation and INT4 quantization, local deployment remains resource-intensive because all trillion-scale expert weights must be stored or distributed. As an example, Moonshot AI's deployment guide for the FP8 July 2025 K2 checkpoint specified 16 H200 or H20 GPUs as the smallest documented mainstream configuration for a 128,000-token context; its 256,000-token K2-0905 example likewise used 16 H200 GPUs. Actual requirements vary with precision, quantization format, context length, KV-cache allocation, batch size, expert parallelism, offloading, and inference software.

Limitations and open problems

Resource requirements

Sparse activation reduces per-token arithmetic relative to a dense trillion-parameter network, but it does not remove the need to store the full expert set. Large K2 deployments therefore require substantial accelerator memory, interconnect bandwidth, and careful expert-parallel or tensor-parallel execution.

Long contexts also increase prefill cost and KV-cache memory. Native INT4 checkpoints lower weight-memory and bandwidth demands but do not eliminate the computational and operational complexity of multi-node inference.

Hallucinations and error propagation

Like other generative language models, Kimi models can produce unsupported statements, fabricated citations, incorrect code, or faulty interpretations of visual material. Tool access can improve grounding, but it also introduces new failure modes, including:

• selection of an unsuitable tool;
• malformed arguments;
• misinterpretation of tool output;
• repeated or unnecessary calls;
• incorrect state carried across a long trajectory;
• compounding errors during autonomous execution.

Long reasoning traces and large tool budgets do not guarantee correctness. High-impact outputs require external validation, testing, source inspection, and permission controls.

Dependence on synthetic data and model-based evaluation

The K2 post-training pipeline relies heavily on synthetic tool environments, generated agent trajectories, automated verifiers, and LLM-based rubric judges. This permits scaling beyond manually authored examples but creates open questions about:

• bias in self-evaluation or model-based judging;
• overfitting to synthetic tool specifications;
• transfer from simulated environments to real systems;
• error reinforcement through iterative bootstrapping;
• sensitivity to the distribution of generated tasks;
• the extent to which benchmark improvements generalise to uncontrolled workflows.

Benchmark comparability

Reported results frequently use different context-management strategies, reasoning budgets, tool sets, judge models, and numbers of sampling runs. Some K2.5, K2.6, and K2.7 Code evaluations use internal benchmarks or developer-re-evaluated baselines. Direct ranking from isolated table cells can therefore be misleading.

For agentic benchmarks in particular, measured performance reflects the combined model, prompt, tool interface, execution environment, context compressor, stopping rule, and retry policy.

Transparency and reproducibility

Moonshot AI releases weights, model cards, selected code, architectural details, and technical reports for several Kimi models. However, the complete training datasets, all data-cleaning rules, full language and domain distributions, every post-training environment, and the production training infrastructure are not public.

As a result, third parties can reproduce inference and conduct independent evaluations but cannot exactly reproduce the original K2, K2.5, K2.6, or K2.7 Code training runs solely from the released materials.

Safety and misuse

An independent preliminary safety evaluation of Kimi K2.5 published in April 2026 examined chemical, biological, radiological, nuclear, cyber, sabotage, self-replication, political-bias, and harmlessness risks. The authors reported capabilities comparable to contemporary closed models in some dual-use domains, comparatively low refusal rates for some hazardous requests, competitive but not frontier-level autonomous cyberoffensive performance, and concerns in selected sabotage and self-replication tests. The study also described its findings as preliminary and called for more systematic safety evaluations of frontier open-weight releases.

Open-weight availability allows independent research and local control but also shifts more responsibility to deployers. Production systems may require sandboxing, least-privilege tool permissions, network restrictions, audit logs, content safeguards, human approval gates, and application-specific testing.

Ethical and regulatory considerations

Kimi model cards state that developers and integrators are responsible for evaluating model behaviour in their intended use case and for applying appropriate safeguards. Multimodal inputs may include personal, confidential, copyrighted, or otherwise regulated material, requiring lawful collection, secure processing, and suitable retention controls.

Public generative-AI services offered in China operate within the applicable Chinese regulatory framework for algorithms, data, and generative-AI services. International deployments may additionally be subject to privacy, copyright, consumer-protection, sector-specific, export-control, and AI-governance rules in other jurisdictions.

On 23 February 2026, Anthropic alleged that DeepSeek, Moonshot AI, and MiniMax had used fraudulent accounts and large numbers of interactions with Claude in unauthorised model-distillation campaigns. Anthropic described more than 16 million exchanges across approximately 24,000 accounts collectively attributed to the three companies. The allegation was a claim by Anthropic, not an adjudicated finding, and the collective figures should not be attributed to Moonshot AI alone. Public reporting available by July 2026 did not establish an independent investigation resolving the allegation.

Research directions

Published Moonshot AI work suggests several continuing research directions:

• Long-horizon agents. Improving reliability, recovery from failed actions, memory management, tool selection, and verification over hundreds or thousands of steps.
• Parallel agent orchestration. Developing Agent Swarm beyond simple task fan-out, including communication among sub-agents, dynamic parallel width, conflict resolution, cost control, and stronger synthesis.
• Efficient MoE inference. Reducing expert-routing overhead, balancing expert loads, improving multi-node communication, and exploiting low-precision weights without excessive quality loss.
• Native multimodality. Jointly training text, images, video, code, interfaces, and tool-use trajectories rather than attaching vision only after text pre-training.
• Long-context architecture. Improving attention, memory, retrieval, and context compression so that nominal context length produces useful retention rather than merely accepting more input tokens.
• Independent evaluation. Increasing reproducibility of benchmark claims, safety assessment, multilingual testing, and real-world agent evaluation.

A separate Moonshot AI research line, Kimi Linear, combines Kimi Delta Attention with periodic global MLA layers in a 3:1 ratio. The released checkpoints have approximately 48 billion total parameters, three billion active parameters, and a one-million-token context. The authors report up to a 75% reduction in KV-cache use and up to a sixfold decoding-throughput improvement at very long contexts relative to their full-attention comparison. These are results for the Kimi Linear research architecture and should not be presented as features of the trillion-parameter K2 checkpoints.

Moonshot AI has not published a detailed, binding roadmap for future Kimi version numbers. Statements about unreleased general-purpose K2.7 or later models would therefore be speculative as of 10 July 2026.

External links

• Moonshot AI official website
• Kimi research articles and release log
• Kimi API model list
• Moonshot AI on Hugging Face
• Moonshot AI on GitHub
• Kimi K2.6 model card
• Kimi K2.7 Code model card

See also

• Large language model
• Mixture of experts
• Transformer (deep learning architecture)
• AI agent
• DeepSeek
• Qwen

Literature

• Kimi Team. Kimi K2: Open Agentic Intelligence. arXiv:2507.20534, 2025.
• Kimi Team. Kimi K2.5: Visual Agentic Intelligence. arXiv:2602.02276, 2026.
• Kimi Team. Kimi k1.5: Scaling Reinforcement Learning with LLMs. arXiv:2501.12599, 2025.
• Kimi Team. Kimi-VL Technical Report. arXiv:2504.07491, 2025.
• Kimi Team. Kimi Linear: An Expressive, Efficient Attention Architecture. arXiv:2510.26692, 2025.
• Qin, R. et al. Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving. arXiv:2407.00079, 2024.
• Vaswani, A. et al. Attention Is All You Need. Advances in Neural Information Processing Systems 30, 2017.
• Yong, Z.-X. et al. An Independent Safety Evaluation of Kimi K2.5. arXiv:2604.03121, 2026.