HomeAI models
MoonshotAI: Kimi K2 0905
Moonshot AI
•
Kimi K2 0905
•
Released 
September 2025

Moonshot AI
Kimi K2 0905

Open-weight 1T-parameter MoE with 256K context, tuned for agentic coding, frontend generation and tool calling without reasoning overhead.

Modality:
Text
PDF
model ID
moonshotai/kimi-k2-0905
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

N/A
tok/s

Intelligence Index *

15.3
/ 100

Context Window *

262144
tokens

Input price

3.6
Anytoken

Output price

15
Anytoken
Kimi K2 0905: Agentic Coding at 256K Context, Without a Thinking Mode Kimi K2 0905 (Kimi-K2-Instruct-0905) is Moonshot AI's September 2025 update to Kimi K2, an open-weight mixture-of-experts model with 1 trillion total and 32 billion active parameters. It sits between the original 0711 release and the later Kimi K2 Thinking line. Moonshot positions it around agentic coding, better frontend output and a context window doubled to 256K tokens. It suits coding agents and tool-calling loops that need non-thinking responses. Call Kimi K2 0905 through AnyAPI.ai

Performance

Where the 0905 Update Moved the Needle: Terminal and Repository Tasks

Kimi K2 0905 is tuned for agentic coding. Moonshot reports 69.2 on SWE-bench Verified, averaged over five runs in its in-house harness, up from 65.8 for K2-Instruct-0711. On Terminal-Bench (Terminus-2) it reports 44.5 against 37.5 for 0711. These are vendor-reported figures, but the larger terminal gain suggests the update helps multi-step tool loops more than single-shot code completion. In production, it fits coding agents that run shell, editor and test tools inside a 256K window, with non-thinking responses and no reasoning overhead.

Benchmarks

Independent Data: Above-Median Intelligence, Below-Average Speed

Artificial Analysis scores Kimi K2 0905 at 15 on its Intelligence Index, above the median of 12 for comparable open-weight non-reasoning models. It measures about 52.5 output tokens per second against an average of 63. On Novita, the one provider still benchmarked, time to first token is 1.19 seconds on a 10k-token input. Artificial Analysis marks the model deprecated and only updates that 10k-input workload, so treat these as historical reference points rather than current service levels.

Output Speed

*
N/A
tok/s

Intelligence Index

*
15.3
/ 100

MMLU *

Broad world knowledge and problem-solving
82
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
77
%

HLE *

Adherence to multi-step structured instructions.
6
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
61
%

Technical Specifications

What the model supports

Kimi K2 0905 is a mixture-of-experts model with 1 trillion total parameters, 32 billion activated per token, a 256K context window, text input and text output. The window doubled from 128K in the 0711 release, which matters most for repository-scale agent loops where file contents, history and tool results accumulate. The official model card does not state a maximum output length and third-party listings disagree, so confirm the output cap with your serving provider before planning long generations.
Verified Specifications — 
Kimi K2 0905
*
Input modalities
Text
PDF
output modalities
Text
Context window
262144
 tokens
Maximum output tokens
262144
Reasoning
No
Knowledge cutoff
September 2025
Pricing (standard)
3.6
 AnyTokens in
 / 
15
 AnyTokens out

Limitations & Trade-offs

Where Kimi K2 0905 falls short

1
Superseded on Moonshot's own API. Moonshot's docs list kimi-k2-0905-preview as deprecated and say the kimi-k2 series was discontinued on May 25, 2026. Artificial Analysis also marks it deprecated and recommends a newer Kimi release. Teams starting new builds should weigh Kimi K2.6 or K2.7 Code; this model fits pinned baselines and existing integrations best.
2
No reasoning mode and no image input. Artificial Analysis lists text input only and no reasoning, and it scores K2 0905 at 15 against 22 for Kimi K2 Thinking. Hard multi-step reasoning, math-heavy analysis and screenshot or document-image workloads belong on a thinking or multimodal model such as Kimi K2 Thinking or K2.6.
3
Moderate speed. Artificial Analysis measures about 52.5 output tokens per second, below its 63 average for comparable models. Time to first token was 1.19 seconds on the single provider still benchmarked. Latency-sensitive chat, autocomplete and high-volume streaming may be better served by a faster or smaller model. Cost-sensitive volume work should also account for Artificial Analysis rating it relatively expensive among similar open-weight non-reasoning models.
4
Vendor-reported coding scores and unclear output cap. The SWE-bench and Terminal-Bench figures come from Moonshot's model card, mostly from its in-house harness, and neutral-harness data is limited. The maximum output length is also not in the official card, and third-party listings differ. Validate on your own repositories and confirm output limits with your provider before relying on long generations.

Best-Fit Workloads

Where this model earns its place

01

Terminal and Repository Coding Agents

‍
Moonshot reports 44.5 on Terminal-Bench and 69.2 on SWE-bench Verified, both improvements over the 0711 release. With tool calling and a 256K window, the model can hold file contents, command output and edit history in one session. It fits CLI coding agents and automated issue-fix pipelines. Treat the scores as vendor-reported, and test on your own codebase.

02

Frontend Code Generation

‍
Moonshot lists better frontend programming, in both aesthetics and practicality, as a key feature of the 0905 update. That makes it a candidate for generating components, layouts and UI scaffolding from text specifications. It takes text input only, so design-to-code from screenshots needs a vision-capable model.

03

Long-Horizon Tool-Calling Agents

‍
The model card describes strong tool-calling behavior, and the 256K window, double the 0711 release, lets agents keep longer histories before truncation. Non-thinking operation avoids reasoning-token overhead in each loop step. This suits multi-step automation across internal APIs and developer tools. For tasks needing deep planning, Kimi K2 Thinking is the stronger option.

04

Coding agents in tool-call loops


MoonshotAI positions this update around agentic coding and better generalization across agent scaffolds. It reports 69.2% on SWE-Bench Verified, and its vendor verifier recorded 100% schema accuracy on successful tool calls for the preview endpoint. Without a reasoning phase, each step returns directly, which suits terminal and IDE agents. For hard planning steps, pair it with a reasoning model.

‍

Pricing in anytokens via AnyAPI
Input
3.6
₳
Output
15
₳
Cache write
—
₳
Cache read
0
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access Kimi K2 0905 via AnyAPI.ai

Access Kimi K2 0905 through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Kimi K2 0905 without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access Kimi K2 0905 and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test Kimi K2 0905 against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use Kimi K2 0905 from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use Kimi K2 0905 for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

Kimi K2 0905 has a 256K-token context window, up from 128K in Kimi K2 0711, according to Moonshot's model card. Some third-party listings show it as 262,144 tokens. The context window covers the whole request and response, so it is not a guaranteed maximum output length, which the official card does not state.

No. Kimi K2 0905 is a non-thinking instruct model with text input and text output, and Artificial Analysis lists it as non-reasoning. For extended reasoning, use Kimi K2 Thinking. For visual input, Moonshot's docs describe Kimi K2.6 as accepting visual and text input.

Moonshot reports 69.2 on SWE-bench Verified and 44.5 on Terminal-Bench for Kimi K2 0905, up from 65.8 and 37.5 for the 0711 release. These are vendor-reported figures, mostly from an in-house harness. Run your own repository tests before choosing it for production coding agents.

No. Moonshot's documentation lists kimi-k2-0905-preview as deprecated and says the kimi-k2 series was discontinued on May 25, 2026. The open weights remain published on Hugging Face under a modified MIT license. You can still access the model through AnyAPI.ai.

Artificial Analysis rates Kimi K2 0905 as relatively expensive compared with other open-weight non-reasoning models of similar size. Costs on AnyAPI.ai are shown in the AnyTokens pricing component on this page. Long agent sessions with large contexts and long outputs will cost more than short chat turns.

* Benchmark data source: Artificial Analysis artificialanalysis.ai