Can I run it?

Can I run Kwaipilot KAT Coder V2.5 Dev on GeForce RTX 3080 10GB?

Expert offload

Expert offload on your GeForce RTX 3080 10GB: Q4_K_M keeps shared weights on the GPU and about 12.7 GB of experts in system RAM. About 26–36 tok/s. Active-parameter speed is approximate.

Kwaipilot KAT Coder V2.5 Dev details · GeForce RTX 3080 10GB page · How this estimate is made

Run commands

# Expert offload is approximate. llama.cpp can pin experts to CPU:
#   -ot "\\.ffn_.*_exps\\.=CPU"
llama-server -hf bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF:Q4_K_M -c 8192 -ngl 99

lms get bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF
# Then load quant Q4_K_M

Assumptions

• Estimate from memory bandwidth, not a benchmark. Decode speed uses 55–75% of peak bandwidth.
• KV cache is FP16. Engine overhead assumes llama.cpp (0.3 GB + 1% of the weights). Ollama is closer to 0.9 GB + 3%, LM Studio closer to 1 GB + 3%.
• Q4_K_M size is the real GGUF file size from Hugging Face.
• 0.5 GB is reserved because this GPU also drives a display, then 5% of the rest is left free.
• Speed is quoted at 8K context, not at the maximum context that still fits.
• Mixture-of-experts split is approximate: shared layers on the GPU, expert tensors in system RAM. Speed reads more than the active-parameter slice, because shared layers stay on the bus.

Models that fit this machine more comfortably

• Qwen3 VL 32B Instruct — Heavily degraded, UD-IQ1_S, 47–65 tok/s (estimate)
• Can I run Qwen3 Coder 30B A3B Instruct on GeForce RTX 3080 10GB? — Heavily degraded, UD-IQ1_S, 112–153 tok/s (estimate)
• Can I run Qwen3.8 27B on GeForce RTX 3080 10GB? — Heavily degraded, UD-IQ2_S, 47–64 tok/s (estimate)

Updated 2026-10-24. Profile on this page is the device default plus 32 GB of DDR5-5600 system RAM.