Atlas

Models

← All models

Kimi K2 Instruct 0905

Moonshot AI · 2025-09-05 · 1T parameters

Kimi K2 Instruct 0905 is Moonshot AI's September 2025 open-weight update to Kimi K2 Instruct, retaining K2's trillion-parameter MoE architecture and base pretrained on a reported 15.5T tokens. The checkpoint improves agentic coding, frontend coding, and tool use, extends the supported context from 128K to 256K tokens, and has approximately 32.6B activated parameters according to the K2 technical report.

Benchmark scores

BenchmarkScore
A-Fantasia76.0
A-Fantasia (Valid Attempts with Retries)74.7
A-Fantasia Backwards Spelling89.0
A-Fantasia Chess58.0
A-Fantasia Chess (Valid Attempts with Retries)58.0
A-Fantasia Cube (Valid Attempts with Retries)77.1
A-Fantasia Cube Rotation81.0
A-Fantasia Spelling (Valid Attempts with Retries)89.0
AA-LCR52.3
AA-LCR v1.153.0
AA-Omniscience Index-26.6
AIME 2025 (Artificial Analysis avg@10)57.3
ALE-Bench776
Arena Text — Business, Management & Financial Ops (No Style Control)1354
Arena Text — Chinese (No Style Control)1415
Arena Text — Creative Writing (No Style Control)1349
Arena Text — English (No Style Control)1379
Arena Text — Entertainment, Sports & Media (No Style Control)1342
Arena Text — Exclude Ties (No Style Control)1352
Arena Text — Exclude Ties (Style Controlled)1407
Arena Text — Expert (No Style Control)1365
Arena Text — French (No Style Control)1382
Arena Text — French (Style Controlled)1426
Arena Text — German (No Style Control)1387
Arena Text — Hard Prompts (No Style Control)1384
Arena Text — Hard Prompts English (No Style Control)1379
Arena Text — Instruction Following (No Style Control)1348
Arena Text — Japanese (No Style Control)1349
Arena Text — Korean (No Style Control)1323
Arena Text — Legal & Government (No Style Control)1403
Loading Atlas data…