All models › moonshot

Kimi K2.6

moonshot · Modified MIT license · on Arena since 22 Apr 2026

Lab chat Download
Arena score1455#44 of 416 in Chat, paid models included
Size1026.9 B33.3 B active per token (mixture of experts)
4-bit download622 GBweights only, Q4_K_M
Max context256Ktokens, from the model config

Where to use it for free

Arena scores

ArenaScoreRankVotes
Chat1455#44 of 41640K
Coding1510#56 of 14410K
Vision1283#33 of 16516K

Rank among all models, paid ones included. Source: Arena leaderboard dataset.

Speed on popular hardware

HardwareGPU memoryFits?Speed
RTX 3060 12 GB12 GBDoes not fit—
RTX 4060 Ti 16 GB16 GBDoes not fit—
RTX 309024 GBDoes not fit—
RTX 409024 GBDoes not fit—
RTX 509032 GBDoes not fit—
RX 7900 XTX24 GBDoes not fit—
2 × RTX 309048 GBDoes not fit—
4 × RTX 309096 GBDoes not fit—
Mac M4 Max 128 GB96 GBDoes not fit—
Ryzen AI Max+ 395 128 GB96 GBDoes not fit—
8 × H100 SXM640 GBFits in GPU86 tok/s

Estimate at 4-bit with 32K context, 64 GB of DDR5 for offload. Speed ≈ memory bandwidth × 70% ÷ data read per token; real numbers vary by about ±30%.

GPU memory needed

Level8K context32K context128K contextQuality vs full
Q3_K_M515 GB516 GB523 GB+4.7% noticeable loss
Q4_K_M623 GB625 GB632 GB+1.2% very small loss
Q5_K_M731 GB733 GB740 GB+0.4% near-lossless
Q6_K843 GB845 GB852 GB+0.2% near-lossless
Q8_01092 GB1094 GB1101 GB≈ 0% same as full
F162055 GB2057 GB2064 GB≈ 0% same as full

Weights + context cache (from the model's architecture) + runtime. Quality: increase in perplexity against the full model, from llama.cpp tests, scaled by model size. Simulate your own setup.

Chat score over time

Kimi K2.6 1455Best paid model

Arena score from 17 Apr 2026 to 15 Oct 2026.

Vision score over time

Kimi K2.6 1283Best paid model

Arena score from 10 Apr 2026 to 9 Oct 2026.

Similar free models in Chat