Qwen3.6 35B-A3B

Qwen/Qwen3.6-35B-A3B

3B active / 35B totalMoEText generationImage → textText inputImage input

Qwen · 9 recorded configurations

Measured setup

Hardware

1× RTX 3090 · GPU-resident

Runtime and tool

live llama-server

Quantisation

MTP draft depth

Every option opens a recorded setup. Where the corpus recorded nothing at this page's other settings, the option names the fields it switches with it. Allocation and runtime detail stay in the technical record below.

PP · left axisTG · right axis

View exact measurements

Exact context depth and measured fresh prefill, bulk prefix, and decode rates for Qwen3.6 35B-A3B
Context depth (tokens)	Fresh prefill PP (tok/s)	TG (tok/s)
512	3,245.74	136.13
2,048	3,171.72	134.36
4,096	3,111.54	131.02
8,192	3,027.03	126.87
16,384	2,989.40	118.86
32,768	2,847.98	103.58
65,536	2,529.12	—

Technical record and provenance

Run details

Source run ID: rtx3090x1-qwen3.6-35b-a3b-iq4xs-2026-07
Hardware: 1× RTX 3090
Runtime: llamaCpp
Weight placement: gpu-resident
KV-cache precision: Q8_0
Prompt tokens: 512
Generated tokens: 128
Speculative decoding: None (standard decoding)
Notes: Single RTX 3090, llama.cpp, Unsloth IQ4_XS build (4.25 bpw) of the MTP-capable GGUF, Q8_0 KV cache. The 64K row has no TG: the generation phase was not measured at that depth (published as “—”). Build, driver, power limit, and repetitions not recorded in the article.
Source article: /local-ai/rtx-3090-for-local-ai-inference/

Compare measured entries