Fact-checked: August 4, 2026 · Model status: active, but no longer Kimi’s flagship · Evidence level: official documentation and model card; not independently benchmarked by Kimi AI Guide.
Kimi K2.6 is Moonshot AI’s open-weight, general-purpose multimodal model. Its Global Kimi API ID is kimi-k2.6, it supports a 262,144-token context window, and the hosted API accepts text, image and video. Unlike Kimi K3 and Kimi K2.7 Code, K2.6 can run with thinking enabled or disabled.
K2.6 remains useful for mixed workloads where cost, multimodal input and control over reasoning mode matter. It is not the best Kimi choice for every task: K3 is the current flagship, while K2.7 Code is purpose-built for long-horizon software engineering.
Kimi K2.6 at a glance
| Field | Verified value |
|---|---|
| Current status | Active in the official Kimi API model list |
| Release date | April 20, 2026 |
| Global Kimi API model ID | kimi-k2.6 |
| Official open-weight repository | moonshotai/Kimi-K2.6 |
| Architecture | Mixture of Experts with MLA attention and MoonViT vision encoder |
| Total / activated parameters | 1T / 32B |
| Context window | 262,144 tokens |
| Hosted API input | Text, image and video |
| Thinking | Enabled by default; can be disabled |
| Tool calling | Supported, with mode-specific constraints |
| License | Modified MIT |
Sources: Kimi’s K2.6 API guide, official model card and technical release post.
Is Kimi K2.6 deprecated?
No—not as of August 4, 2026. The official Kimi API model list places kimi-k2.6 in its active multimodal section with K3 and K2.7 Code.
The confusion comes from a separate notice saying the older “Kimi K2 series” IDs were discontinued on May 25, 2026. That deprecated list names:
kimi-k2-0905-previewkimi-k2-0711-previewkimi-k2-turbo-previewkimi-k2-thinkingkimi-k2-thinking-turbo
It does not name kimi-k2.6. Kimi’s API homepage also continues to show K2.6 with live pricing. Because model status can change, production applications should check Kimi’s live model list and maintain a migration plan.
What is Kimi K2.6 designed for?
Moonshot positions K2.6 as a general-purpose model for dialogue, reasoning, coding and agent tasks. Its release materials emphasize long-horizon coding, frontend generation, proactive agents and multi-agent orchestration, but its practical distinction in the current lineup is flexibility: K2.6 can handle general text and vision work without forcing thinking on every request.
Potential workloads include:
- document analysis and structured extraction;
- image or video understanding through the hosted API;
- writing, rewriting and conversation;
- coding tasks that do not require the latest coding specialist;
- tool-using agents with a 256K context budget;
- applications that switch between deliberate and non-thinking responses.
Those are capability categories, not independent performance results. Output quality, speed and reliability must be measured on the exact deployment and prompts you plan to use.
Architecture and context window
| Specification | Official value |
|---|---|
| Total parameters | 1 trillion |
| Activated parameters | 32 billion |
| Layers | 61 |
| Experts / selected per token | 384 / 8 |
| Context length | 256K (262,144 tokens) |
| Attention mechanism | Multi-head Latent Attention (MLA) |
| Vision encoder | MoonViT, 400M parameters |
K2.6 and K2.7 Code share the same listed architecture and scale. K2.7 is built on K2.6 and is post-trained for coding-focused agent work, so parameter count alone cannot tell you which one is better for a task.
A 256K context window is the total model capacity, not a promise of perfect recall. System instructions, tool schemas, previous messages, media tokens and output allowance all consume room. Agent clients may compact or discard earlier history before the advertised limit.
Thinking and non-thinking modes
K2.6 supports two API behaviors:
| Mode | Request value | Documented sampling behavior | Best considered for |
|---|---|---|---|
| Thinking | {"type":"enabled"} | Temperature fixed at 1.0; top-p 0.95 | Complex analysis, reasoning and multi-step tool work |
| Non-thinking | {"type":"disabled"} | Temperature fixed at 0.6; top-p 0.95 | Routine extraction, rewriting, classification or latency-sensitive responses |
Thinking is enabled by default. The “best considered for” column is editorial guidance, not a measured latency or accuracy claim.
When thinking is enabled, Kimi documents two important tool-use constraints: tool_choice can be auto or none, and the prior assistant reasoning_content must remain in context during multi-step tool calls. K2.6 also supports preserved thinking, but the official model card says it is disabled by default; this differs from K2.7 Code, where preservation is forced.
Kimi’s current built-in WebSearch guide documents the tool as active and still instructs developers to disable K2.6 thinking when using that specific $web_search path. Search-result content also counts toward prompt tokens. Verify the current tool declaration, model/mode compatibility and billing before implementation; first-party pages currently disagree on the per-call search fee.
Multimodal inputs
The official hosted API accepts text, image and video for K2.6. Kimi documents image support for PNG, JPEG, WebP and GIF and a range of common video formats.
Practical caveats:
- public media URLs are not supported in the documented hosted flow; use base64 or uploaded files;
- Kimi recommends images at or below 4K and video at or below 1080p;
- higher-resolution media can consume more tokens and processing time without improving understanding;
- the platform token-estimation endpoint can estimate media cost before inference;
- large or reusable videos should use file upload rather than a very large request body;
- the open-weight model card describes video chat as experimental and official-API-only at publication time.
Self-hosted vision behavior depends on the serving stack and preprocessing. Do not treat a third-party endpoint as equivalent to Moonshot’s hosted API without verification.
Kimi K2.6 API pricing
Official list prices checked August 4, 2026:
| Billing component | Price per 1M tokens |
|---|---|
| Input with cache hit | $0.16 |
| Input with cache miss | $0.95 |
| Output | $4.00 |
Prices exclude applicable taxes. Input and output are billed by actual token use; media is converted into tokens. Context caching is automatic, but only eligible reused context is charged at the cache-hit rate.
Cost example
For 500,000 cache-miss input tokens and 50,000 output tokens, the list-price estimate is $0.675 before tax: $0.475 input + $0.20 output.
The example excludes retries, extra agent turns and paid tools. See Kimi API pricing for the complete product-level explanation.
Where can you use K2.6?
Global Kimi API Platform
Use kimi-k2.6 on the OpenAI-compatible Global API. This is the surface covered by the model ID, context and token rates above.
Kimi consumer products
Moonshot’s April 20 release post said K2.6 was available through Kimi.com and the Kimi app. Consumer products may not expose the exact backend model ID for every feature, and the default model can change without matching the API lifecycle. A result from the consumer interface should be labelled by the visible product and date unless the model version is explicitly confirmed.
Kimi Code
Current Kimi Code documentation lists K3 and K2.7 Code as its selectable model families. It says requests for K3 or K2.7 with thinking disabled route to K2.6. That fallback is useful, but it is not the same as selecting kimi-k2.6 on the Global API.
Open weights
Moonshot publishes the model at Hugging Face under the Modified MIT License. The official card recommends vLLM, SGLang and KTransformers and lists native INT4 quantization.
K2.6 is still a one-trillion-parameter model. Quantization reduces resource needs but does not make high-quality deployment trivial. Hardware, inference engine, chat template, reasoning-field handling and multimodal preprocessing all matter.
Vendor-reported benchmark snapshot
The following figures are from Moonshot’s official K2.6 model card and are not independent Kimi AI Guide results.
| Benchmark | K2.6 reported score | Setup note |
|---|---|---|
| HLE-Full with tools | 54.0 | Search, code interpreter and browsing tools; vendor setup |
| BrowseComp | 83.2 | Vendor context-management policy |
| Terminal-Bench 2.0 | 66.7 | Terminus-2, preserve-thinking mode |
| SWE-Bench Pro | 58.6 | In-house framework adapted from SWE-agent; 10-run average for coding tasks |
| MMMU-Pro | 79.4 | Three-run visual evaluation protocol |
Moonshot says K2.6 experiments generally used thinking mode, temperature 1.0 and a 262,144-token context. Some competitor scores in the source table were re-evaluated by Moonshot and marked with an asterisk. Some evaluations use tools or internal harnesses. Read the full methodology before comparing any single number.
K2.6 versus K2.7 Code
| Decision factor | K2.6 | K2.7 Code |
|---|---|---|
| Primary role | General-purpose | Coding specialist |
| Context | 256K | 256K |
| Thinking | Optional | Required |
| Standard API cache-miss input | $0.95 / 1M | $0.95 / 1M |
| Standard API output | $4.00 / 1M | $4.00 / 1M |
| Best initial candidate | Mixed content, conversation, optional fast mode | Long-horizon coding and coding agents |
Moonshot reports stronger coding and agent scores for K2.7 Code than K2.6 under its evaluation setup, along with lower average thinking-token use. Choose K2.6 when general-purpose balance or an actual non-thinking mode matters more.
Important limitations
- K2.6 is active but no longer the flagship, so its lifecycle should be monitored.
- A 256K limit does not guarantee accurate recall across the full context.
- Non-thinking mode can reduce deliberation but is not guaranteed to be faster or better for every simple prompt without measurement.
- Tool use with thinking has preserved-history and
tool_choiceconstraints. - Kimi currently documents built-in WebSearch as active, but the specific
$web_searchpath requires K2.6 thinking to be disabled; verify current tool declarations, model/mode compatibility and billing. - Hosted video behavior may not transfer to self-hosted or third-party endpoints.
- Vendor benchmark results depend on tools, harnesses and context management.
- Generated text, code and citations can be wrong and require validation.
Our independent testing status
We have not yet run or published an independent K2.6 test. We do not claim measured response speed, factual accuracy, coding success or cost beyond the official list rates and arithmetic examples on this page.
A future thinking-versus-non-thinking comparison must preserve the exact prompt, attachments, model ID, thinking value, system message, tool definitions, trial count, response, token usage, cost and UTC timestamp. We will not invent a result if authenticated access is unavailable. See our testing methodology.
Frequently asked questions
Is Kimi K2.6 still available?
Yes. kimi-k2.6 appears in the active section of the official Kimi API model list as of August 4, 2026. Older K2 preview and thinking IDs are separately marked deprecated.
What is the Kimi K2.6 API model ID?
Use kimi-k2.6 on the Global Kimi API. Consumer Kimi and Kimi Code are separate product surfaces and may not expose that ID directly.
Can K2.6 turn thinking off?
Yes. Set thinking to {"type":"disabled"} in the documented API request. Thinking is enabled by default.
Does K2.6 support images and video?
Yes on the official hosted API. The open-weight model card calls video chat experimental and official-API-only, so verify self-hosted and third-party deployments separately.
Should I use K2.6 or K2.7 Code?
Choose K2.6 for balanced general-purpose work and optional non-thinking responses. Choose K2.7 Code for coding-focused, long-horizon agent tasks. Their standard Global API cache-miss input and output rates are the same as of this page’s fact-check date.