Moonshot AI is retiring its Kimi K2 series by the end of December 2026, so teams running Kimi K2.6 need to know exactly what changes when they move to Kimi K3. This page compares the two models side by side: API identifiers, dates, architecture, parameter counts, context windows, thinking controls, published prices, availability, provider-reported benchmarks, licences and lifecycle status. It is written for developers migrating K2.6 workloads, architects who self-host the open weights and buyers checking cost before the K2 retirement. Models are listed newest first, an objective order that implies no ranking. Every fact comes from Moonshot AI's Hugging Face model cards, blog posts and API documentation, cited under Standards and references.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses Moonshot AI's public documentation and Hugging Face model cards as of November 2026. Model details change, so confirm against the provider's current documentation.
Key takeaways
- K2.6 is on a retirement path. From 31 October 2026 K2 series models are no longer available to newly registered users, and Moonshot AI plans to retire all K2 series models by the end of December 2026; existing users are not affected until then.
- K3 is almost three times larger. K3 lists 2.8T total and 104B active parameters with 896 experts; K2.6 lists 1T total and 32B active with 384 experts.
- Context grows from 262,144 to 1,048,576 tokens. The Kimi pricing page lists 1,048,576 tokens for the K3 series and 262,144 for the K2 series.
- Thinking can no longer be switched off. K2.6 offers a Thinking mode and an Instant mode; K3 always thinks, takes reasoning_effort low, high or max (default max) and requires the full reasoning history to be passed back.
- Licence and quantisation change. K2.6 uses a Modified MIT License with native INT4 weights; K3 uses the Kimi K3 License with native MXFP4 weights and MXFP8 activations.
What is Kimi?
Kimi is the model family from Moonshot AI. Moonshot AI publishes open weights for each Kimi model on Hugging Face and serves them through the Kimi API Platform, the Kimi apps and Kimi Code. Model names follow a generation and point-release pattern: K2, K2.5, K2.6, then K3 and K3.1. Kimi K2.6 is a K2-series multimodal Mixture-of-Experts model whose card states that its architecture matches Kimi K2.5. Kimi K3 is the next generation, and Kimi K3.1 Pro, a post-trained version of K3, has since followed it.
This page covers K3 and K2.6 only. For K3.1 Pro and its Highspeed variant, see the Kimi K3 and K3.1 profile; for the full line-up, see the Moonshot AI Kimi Models hub.
Every model at a glance
As of 2026-11-15. Sources: Kimi K3 and K2.6 Hugging Face cards; Kimi API pricing page; Kimi K3 launch blog.
| Model | API ID | Released | Status |
|---|---|---|---|
| Kimi K3 | kimi-k3 | July 2026 (blog data current as of 2026-07-16; full weights due by 2026-07-27) | Current |
| Kimi K2.6 | Not stated by Moonshot AI on the pages checked (2026-11-15); the card's examples use the repository name moonshotai/Kimi-K2.6 | Not stated by Moonshot AI on the pages checked (2026-11-15) | Closed to new users from 2026-10-31; retirement planned by end of December 2026 |
| Field | Kimi K3 | Kimi K2.6 |
|---|---|---|
| Parameters | 2.8T total, 104B active | 1T total, 32B active |
| Vision encoder | MoonViT-V2, 401M parameters | MoonViT, 400M parameters |
| Open weights | Yes (moonshotai/Kimi-K3) | Yes (moonshotai/Kimi-K2.6) |
| Licence | Kimi K3 License (code and weights) | Modified MIT License (code and weights) |
| Weight format | Native MXFP4 weights, MXFP8 activations | Native INT4 |
| Knowledge cutoff | Not stated by Moonshot AI on the pages checked (2026-11-15) | Not stated by Moonshot AI on the pages checked (2026-11-15) |
When was each model released?
As of 2026-11-15. Moonshot AI's pages do not state exact announcement days for either model, so the table gives only the dates the official pages state.
| Date | Event | Model |
|---|---|---|
| 2 February 2026 | Kimi K2.5 paper (arXiv 2602.02276), linked from the K2.6 card as the architecture reference | K2.5 (K2.6 predecessor) |
| Not stated | Kimi K2.6 released on Kimi.com, the Kimi app, the API and Kimi Code, per its blog | K2.6 |
| 2026-07-16 | "Current as of" date of the K3 launch blog's benchmark data | K3 |
| By 2026-07-27 | Full K3 weights release date stated in the launch blog | K3 |
| 2026-10-31 | K2 series closed to newly registered users | K2.6 |
| End of December 2026 (planned) | Retirement of all K2 series models | K2.6 |
The K2.6 blog's image files carry 2026-04-20 in their paths, which is not a stated release date, so this page does not use it. See LLM Naming and Versioning Explained.
How much does each model cost?
As of 2026-11-15. Per million tokens. The pricing page's tables did not render in the text checked, so only figures stated in prose are given.
| Model | Input (cache miss) | Input (cache hit) | Output | Batch | Source |
|---|---|---|---|---|---|
| Kimi K3 (kimi-k3) | $3.00 | $0.30 | $15.00 | 60% of standard price | K3 launch blog (USD); batch rule from pricing page |
| Kimi K2.6 | Not stated by Moonshot AI on the pages checked (2026-11-15) | Not stated by Moonshot AI on the pages checked (2026-11-15) | Not stated by Moonshot AI on the pages checked (2026-11-15) | Not stated by Moonshot AI on the pages checked (2026-11-15) | K2.6 blog links to the pricing page without figures |
- Caching: the pricing page states that context caching applies automatically to repeated request prefixes with a default TTL of 5 minutes. For K3 series models, cache writes are billed separately by TTL (5 minutes or 1 hour).
- Billing model: pay-as-you-go with no subscription plan; prices exclude applicable taxes, calculated at checkout by jurisdiction.
- Batch: the pricing page names kimi-k3 and kimi-k3.1-pro for batch inference at 60% of the standard price; no K2 model is named for batch.
Because the K2.6 price is not stated in readable text, confirm both prices in the Kimi console on the day you compare them. See Model Pricing Explained.
What changed between models?
Only changes documented on the model cards, blogs and API pages are listed.
- Scale: 1T total and 32B active (K2.6) to 2.8T total and 104B active (K3). The K3 blog states about 2.5 times better overall scaling efficiency than Kimi K2.
- Attention: K2.6 uses Multi-head Latent Attention (MLA) in 61 layers. K3 has 93 layers, 69 using Kimi Delta Attention (KDA) and 24 using Gated MLA, and adds Attention Residuals.
- Experts: 384 experts with 8 selected and 1 shared (K2.6) to 896 experts with 16 selected and 2 shared, in what the K3 blog calls Stable LatentMoE.
- Quantisation: native INT4 (K2.6, same method as Kimi-K2-Thinking) to MXFP4 weights and MXFP8 activations through quantisation-aware training from the SFT stage (K3).
- Thinking: K2.6 has Thinking (default) and Instant modes, with preserve_thinking off by default. K3 always thinks, returns reasoning_content, and uses preserved thinking history: the full assistant message, including reasoning_content and tool_calls, must be passed back in multi-turn and tool-call conversations. The K3 launch blog said low and high effort would come later; the K3 card now lists low, high and max.
- Vision and video: K2.6 supports image input, with video described as experimental and official-API only. The K3 blog describes native vision with video understanding in the same model, while the K3 card's summary table lists only text and image.
- Agents: the K2.6 blog states agent swarm scale of 300 sub-agents across 4,000 coordinated steps, up from 100 sub-agents and 1,500 steps on K2.5. The K3 pages checked do not state a swarm figure.
For background on always-on reasoning, see Reasoning Models Explained.
What workloads does Moonshot AI position each model for?
This section records the provider's own positioning; it is not an assessment.
| Model | Positioning stated by provider | Related notes stated by provider |
|---|---|---|
| Kimi K3 | Native multimodal model with vision and video understanding; 2.8T parameters | Benchmarks reported at max effort, temperature 1.0 and top_p 1.0; Kimi Code CLI recommended as the agent framework |
| Kimi K2.6 | Long-horizon coding across languages such as Rust, Go and Python, and agent swarms of self-created sub-agents | "Claw Groups" research preview for mixed human and agent teams |
Worked cost example
Using the K3 launch prices, 1,000 requests a day with a 50,000-token prompt (half served from cache) and a 5,000-token response use 25 million cache-miss input tokens ($75), 25 million cache-hit tokens ($7.50) and 5 million output tokens ($75), about $157.50 a day before cache-write charges and tax. Because K3 always reasons, reasoning tokens add to output, so measure real output length on your prompts. Through the Batch API at 60% of standard price, the same work would be about $94.50. A K2.6 comparison is not possible from the readable pages.
What is the context window of each model?
As of 2026-11-15. Sources: Kimi API pricing page and model cards.
| Model | Context window | Max output | Input | Output | Reasoning controls |
|---|---|---|---|---|---|
| Kimi K3 | 1,048,576 tokens | Not stated by Moonshot AI on the pages checked (2026-11-15) | Text, image (video per blog) | Text | Always thinking; reasoning_effort low, high, max (default max) |
| Kimi K2.6 | 262,144 tokens (256K on the card) | Not stated by Moonshot AI on the pages checked (2026-11-15) | Text, image; video experimental, official API only | Text | Thinking (default) or Instant mode; optional preserve_thinking |
The move from 262,144 to 1,048,576 tokens is four times the window. The K3.1 Pro quickstart documents max_completion_tokens up to 1,048,576 for that model; the equivalent K3 limit is not stated on the pages checked. See Context Windows Explained.
Capabilities and tools compared
As of 2026-11-15. Source: model cards.
| Model | Architecture | Serving frameworks stated by provider | Recommended sampling | API compatibility |
|---|---|---|---|---|
| Kimi K3 | 93 layers (1 dense; 69 KDA, 24 Gated MLA); attention hidden 7168, 96 heads; 896 experts, 16 selected, 2 shared; vocabulary about 160K | vLLM, SGLang, TokenSpeed; Transformers, Docker Model Runner | Temperature 1.0 in benchmarks; top_p 0.95 single-step, 1.0 agentic | OpenAI- and Anthropic-compatible endpoints on platform.kimi.ai |
| Kimi K2.6 | 61 layers (1 dense); MLA, 64 heads; attention hidden 7168; 384 experts, 8 selected, 1 shared; vocabulary about 160K | vLLM, SGLang, KTransformers; Transformers 4.57.1 or later and below 5.0.0; Docker Model Runner | Temperature 1.0 Thinking, 0.6 Instant; top_p 0.95 | OpenAI- and Anthropic-compatible API at platform.moonshot.ai |
The K3 preserved-history rule is the main code change for migrations: clients that drop reasoning_content between turns do not meet K3's documented requirement, whereas K2.6 keeps preserve_thinking off by default. Sampling defaults also differ. Fine-tuning services: not stated by Moonshot AI on the pages checked (2026-11-15). See Mixture of Experts Explained.
Where is each model available?
| Model | Availability stated by provider |
|---|---|
| Kimi K3 | Kimi app (iOS, Android, HarmonyOS) and kimi.com; Kimi Work desktop 3.1.0 or later (Windows and Apple silicon Macs); Kimi Code via /model; Kimi API as kimi-k3; Kimi Enterprise; Hugging Face weights; DeepInfra listed as an inference provider |
| Kimi K2.6 | Kimi.com, the Kimi app, the API and Kimi Code (blog); Hugging Face weights; API access for existing users only from 2026-10-31 |
Cloud marketplaces and regions: not stated by Moonshot AI on the pages checked (2026-11-15) for either model. Self-hosted weights move patching and access control to the organisation; see the Self-Hosted LLM Security Guide.
Model cards and system cards
| Model | Model card or page | System card | Technical report |
|---|---|---|---|
| Kimi K3 | Hugging Face card; launch blog | Not stated by Moonshot AI on the pages checked (2026-11-15) | k3_tech_report.pdf on GitHub, linked from the card |
| Kimi K2.6 | Hugging Face card; blog | Not stated by Moonshot AI on the pages checked (2026-11-15) | Kimi K2.5 paper (arXiv 2602.02276), linked from the card |
See Model Cards and System Cards Explained.
Benchmarks
As of 2026-11-15. All figures are provider-reported on the two Hugging Face cards. Neither card has a column for the other model, so figures come from separate tables with separate settings (K3 at max effort and temperature 1.0; K2.6 official evaluations at temperature 1.0 and top_p 1.0). Only benchmark names that appear on both cards are paired. Independent leaderboard results (LMArena, Artificial Analysis, Epoch AI) were not checked for this page.
| Benchmark | Kimi K3 | Kimi K2.6 |
|---|---|---|
| GPQA Diamond | 93.5 | 90.5 |
| HLE-Full (without / with tools) | 43.5 / 56.0 | 34.7 / 54.0 |
| BrowseComp | 91.2 | 83.2 |
| DeepSearchQA (F1) | 95.0 | 92.5 |
| APEX-Agents | 41.0 | 27.9 |
| OSWorld-Verified | 84.8 | 73.1 |
| SciCode | 58.7 | 52.2 |
| MMMU-Pro (without / with python) | 81.6 / 83.4 | 79.4 / 80.1 |
| CharXiv RQ (without / with python) | 84.8 / 91.3 | 80.4 / 86.7 |
| MathVision (without / with python) | 94.3 / 97.8 | 87.4 / 93.2 |
Not paired because the suite versions differ: Terminal-Bench (K3 reports 2.1 at 88.3; K2.6 reports 2.0 at 66.7), Toolathlon-Verified against Toolathlon, and MCPMark-Verified against MCPMark. The K3 card labels its paired figures with a slash without restating the condition for each; the reading above follows the K2.6 card's separate rows. See How to Read LLM Benchmarks.
Deprecations and retirements
As of 2026-11-15. Source: Kimi API pricing page.
| Model | API ID | Notice or deprecation date | Retirement or shutdown date | Stated replacement |
|---|---|---|---|---|
| Kimi K3 | kimi-k3 | None listed | None listed | None stated |
| Kimi K2.6 (K2 series) | K2 IDs not named on the pricing page | 2026-10-31: no longer available to newly registered users | Planned by end of December 2026 | The notice links a migration guide for kimi-k3-1-pro |
Moonshot AI describes the K2 retirement as planned, not dated to the day, so watch the pricing page for an exact date. The open K2.6 weights on Hugging Face are separate from the hosted API retirement. See the AI Model Retirement Guide and the LLM Deprecation and Retirement Tracker.
Data handling and security
Moonshot AI's API data security page states that inputs and outputs submitted through the API are not used to train or improve Kimi's models, that data is used for the current request and is not persistently stored for training purposes, that all API communication uses HTTPS/TLS, that data from different users is strictly isolated, and that uploaded files can be deleted through the file API or console. It does not state a general retention period, encryption at rest or where data is processed. These terms are platform-wide, so they do not change between K2.6 and K3.
A Kimi API key reaches every model the account can call, so the K2 retirement is a good moment to rotate keys and remove unused ones. See the LLMjacking Guide, the Open-Weight Model Security Guide and AI Data Retention and Deletion.
How to choose a model: factual criteria
| Criterion | What the documentation shows |
|---|---|
| Hosted lifecycle | K2 series retiring by end of December 2026; no notice for K3 |
| New accounts | K2 series unavailable to accounts registered from 2026-10-31 |
| Context above 262,144 tokens | K3 only (1,048,576) |
| Non-reasoning responses | K2.6 Instant mode; K3 always thinks |
| Licence for self-hosting | Modified MIT (K2.6) or Kimi K3 License (K3); read each LICENSE file |
| Self-hosting footprint | 1T total, 32B active, INT4 (K2.6) against 2.8T total, 104B active, MXFP4 (K3) |
| Published API price | Stated in prose for K3 only |
Run your own evaluation on representative prompts before migrating. For a cross-provider view, see DeepSeek vs Qwen vs Kimi vs GLM and Open-Weight Frontier Models Compared.
Practitioner checklist
- Inventory every service, agent and script calling a K2 series model and plan migration before the end of December 2026.
- Update clients to pass back the full assistant message, including reasoning_content and tool_calls, before switching to K3.
- Replace any reliance on K2.6 Instant mode with an explicit reasoning_effort setting on K3 and measure the latency and output-token change.
- Re-test prompts that approach 262,144 tokens; on K3 they no longer need truncation, but cost rises with length.
- Confirm current K3 and K2.6 prices in the Kimi console and record the date checked.
- Read the Kimi K3 License before self-hosting or redistributing K3 weights.
- Rotate Kimi API keys during the migration and set spend limits; see How to set spend limits on LLM API keys.
Standards and references
- Moonshot AI: Kimi API pricing (accessed 2026-11-15)
- Moonshot AI: Kimi K3 model card (accessed 2026-11-15)
- Moonshot AI: Kimi K2.6 model card (accessed 2026-11-15)
- Moonshot AI: Introducing Kimi K3 (accessed 2026-11-15)
- Moonshot AI: Kimi K2.6 blog (accessed 2026-11-15)
- Moonshot AI: Kimi K3.1 Pro quickstart (accessed 2026-11-15)
- Moonshot AI: Kimi API data security (accessed 2026-11-15)
Related NHI Mgmt Group resources: Moonshot AI Kimi Models, Kimi K3 and K3.1, DeepSeek vs Qwen vs Kimi vs GLM, Frontier AI Models Compared, Chinese vs US Frontier Models Compared.