Compare the listed model data for Kimi-K2 Instruct by Moonshot AI and Kimi-k2.5 by Moonshot AI. Benchmark and pricing records are not a controlled test and may change.
Data last updated: 2026-10-16. Verify current model details with each provider.
What the listed data says
Kimi-K2 Instruct: 128K tokens; Kimi-k2.5: 256K tokens. Kimi-k2.5 has the larger listed context window. Kimi-K2 Instruct is listed for coding, reasoning, agentic workflows, long-context tasks; Kimi-k2.5 is listed for no specific capability flags. Shared catalog benchmark entries: SWE-bench Verified: Kimi-K2 Instruct 65.8%; Kimi-k2.5 58.3%; LiveCodeBench: Kimi-K2 Instruct 75.2%; Kimi-k2.5 58.7%; AIME 2025: Kimi-K2 Instruct 78.4%; Kimi-k2.5 71.2%. These records are reference data, not a controlled head-to-head test; verify current pricing and specifications with each provider.
Catalog source links
Use these model and provider links to verify current specifications, pricing, and availability.
The first listed parameter quantities are 1T total (32B active via MoE) and 1T total (32B active); the larger quantity is 1× the smaller. Total and active MoE parameters are not directly equivalent.
The catalog release dates are July 2025 and July 2025, about 0 months apart at the precision supplied.
Cost example: 1M input and 1M output tokens
A cost example for 1M input plus 1M output tokens cannot be calculated reliably because separate numeric input and output rates are not listed for both models.
Choose Kimi-K2 Instruct if…
Choose Kimi-K2 Instruct when your task aligns with catalog-listed uses or strengths such as Elite software engineering AI: autonomous code review, bug fixing in production repos, Enterprise agentic systems with multi-tool orchestration, Frontier mathematical reasoning, 65.8% SWE-bench Verified — highest open-model software engineering score ever recorded.
Choose Kimi-k2.5 if…
Choose Kimi-k2.5 when your task aligns with catalog-listed uses or strengths such as Long agentic workflows requiring sustained context, Complex software refactoring across large codebases, Extended 256K context vs K2's 128K, Enhanced tool use for multi-step autonomous tasks.
Feature Comparison — Kimi-K2 Instruct vs Kimi-k2.5
Catalog specifications for Kimi-K2 Instruct and Kimi-k2.5
Specification
Kimi-K2 Instruct
Kimi-k2.5
Provider
Moonshot AI
Moonshot AI
Parameters
1T total (32B active via MoE)
1T total (32B active)
Context window
128K tokens
256K tokens
Maximum output
Not listed
Not listed
Catalog pricing
Moonshot API: beta. Research weights released. Third-party: emerging
Moonshot API: beta
Listed input/output cost per 1M tokens
Separate numeric input/output rates not listed
Separate numeric input/output rates not listed
License
Modified Apache 2.0 (Kimi K2 Model License)
Kimi K2 Model License
Coding support
Listed
Not listed
Reasoning support
Listed
Not listed
Vision input
Not listed
Not listed
Multimodal
Not listed
Not listed
Open weights
Yes
Yes
Benchmark: SWE-bench Verified
65.8%
58.3%
Benchmark: LiveCodeBench
75.2%
58.7%
Benchmark: AIME 2025
78.4%
71.2%
Figures are transcribed from LITHOVEX model records. Benchmark versions and evaluation setups may differ; pricing and specifications can change. Check each provider's current documentation before making a decision.
Frequently Asked Questions
What is different between Kimi-K2 Instruct and Kimi-k2.5?
Kimi-K2 Instruct: 128K tokens; Kimi-k2.5: 256K tokens. Kimi-k2.5 has the larger listed context window. Kimi-K2 Instruct is listed for coding, reasoning, agentic workflows, long-context tasks; Kimi-k2.5 is listed for no specific capability flags. Shared catalog benchmark entries: SWE-bench Verified: Kimi-K2 Instruct 65.8%; Kimi-k2.5 58.3%; LiveCodeBench: Kimi-K2 Instruct 75.2%; Kimi-k2.5 58.7%; AIME 2025: Kimi-K2 Instruct 78.4%; Kimi-k2.5 71.2%. These records are reference data, not a controlled head-to-head test; verify current pricing and specifications with each provider.
What benchmark values are listed for Kimi-K2 Instruct and Kimi-k2.5?
SWE-bench Verified: Kimi-K2 Instruct 65.8%; Kimi-k2.5 58.3%; LiveCodeBench: Kimi-K2 Instruct 75.2%; Kimi-k2.5 58.7%; AIME 2025: Kimi-K2 Instruct 78.4%; Kimi-k2.5 71.2%. The catalog values may use different versions and evaluation setups, so they are reference points rather than a controlled head-to-head result.
Which model has the larger listed context window?
Kimi-K2 Instruct: 128K tokens; Kimi-k2.5: 256K tokens. Kimi-k2.5 has the larger listed context window.
What are the listed costs per million tokens for Kimi-K2 Instruct and Kimi-k2.5?
Kimi-K2 Instruct: Separate numeric input/output rates not listed (Moonshot API: beta. Research weights released. Third-party: emerging). Kimi-k2.5: Separate numeric input/output rates not listed (Moonshot API: beta). Where separate numeric input and output rates are available, the worked example uses 1M input plus 1M output tokens. Confirm current terms with each provider.
How far apart are the listed release dates for Kimi-K2 Instruct and Kimi-k2.5?
The catalog lists Kimi-K2 Instruct as July 2025 and Kimi-k2.5 as July 2025, about 0 months apart at the precision supplied.
When was the comparison data last updated?
The model catalog records 2026-10-16 as its data update date. Provider specifications and prices may have changed since then.
How should I choose between Kimi-K2 Instruct and Kimi-k2.5?
Compare the listed specifications and capabilities, review the benchmark notes, then test both with the same representative prompt for your task. This page does not claim a universal winner.
Create your free LITHOVEX account and compare Kimi-K2 Instruct and Kimi-k2.5 on your actual tasks. No credit card, no API key, no subscriptions needed.