Kolibri-1 78B (3.46B active) A3.5B
AMD Radeon 890M · llama.cpp · 262,000 ctx
- generation:
- 27.3 tokens/s
- prompt processing (prefill):
- 430.0 tokens/s
- quant:
- IQ2_M (GGUF)
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.
User benchmarks Kolibri-1 78B at 27.28 t/s decode on a Minisforum AI X1 Pro with a Radeon 890M and 96GB RAM (48GB VRAM share). Setup is llama.cpp over Vulkan with IQ2_M at 262k context; prefill is 430 t/s and the model scored 5/7 on the user's gauntlet. User ranks it the fastest smart agent on the box, ahead of Qwen3.6-35B at 26.8-27.0 t/s and Qwen3-Next-80B at 26 t/s, and notes nothing bigger clears 10 t/s except small-active MoEs.