Benchmark result
unsloth/Qwen3.5-4B-MTP-GGUF on NVIDIA GeForce RTX 4070 — 122.8 tok/s
Measured with Unsloth Studio on September 29, 2026.
What the model built
Open full screen →The coding scenario asks for a playable game in a single HTML file. This is exactly what the model returned, unedited; the frame only adds a line that tells this page how tall it is.
How this run compares
7th fastest of 20 runs of this model on NVIDIA GeForce RTX 4070 · median 117.4 tok/s.
sort
- 1149.8tok/s70.5Unsloth Studio · UD-Q4_K_XL · KV f16
- 2146.2tok/s70.1Unsloth Studio · UD-Q4_K_XL · KV f16
- 3141.6tok/s56.9Unsloth Studio · UD-Q4_K_XL · KV q8_0
- 4138.5tok/s70.5Unsloth Studio · UD-Q4_K_XL · KV q4_0
- 5135.8tok/s67.9Unsloth Studio · UD-Q4_K_XL · KV q8_0
- 6123.0tok/s53.8Unsloth Studio · KV f16
- 7122.8tok/s57.6Unsloth Studio · UD-Q6_K_XL · KV f16this run
- 8121.1tok/s57.6Unsloth Studio · UD-Q4_K_XL · KV f16
- 9119.6tok/s61.6Unsloth Studio · UD-Q6_K_XL · KV f16
- 10117.7tok/s60.4Unsloth Studio · UD-Q6_K_XL · KV f16
- 11117.0tok/s63.5Unsloth Studio · UD-Q6_K_XL · KV f16
- 12117.0tok/s66.9Unsloth Studio · UD-Q6_K_XL · KV f16
- 13116.6tok/s64.7Unsloth Studio · UD-Q6_K_XL · KV f16
- 14111.9tok/s65.5Unsloth Studio · UD-Q6_K_XL · KV q4_0
- 15111.0tok/s67.9Unsloth Studio · UD-Q6_K_XL · KV q8_0
- 16109.0tok/s72.4Unsloth Studio · UD-Q6_K_XL · KV q4_0
- 17107.5tok/s67.5Unsloth Studio · UD-Q6_K_XL · KV q8_0
- 18106.9tok/s66.3Unsloth Studio · UD-Q6_K_XL · KV q8_0
- 1925.5tok/s58.1Unsloth Studio · UD-Q6_K_XL · KV f16
- 2021.9tok/s69.6Unsloth Studio · UD-Q6_K_XL · KV iq4_nl
unsloth/Qwen3.5-4B-MTP-GGUF on other hardware
Reproduce this run
toolUnsloth Studio
modelunsloth/Qwen3.5-4B-MTP-GGUF (UD-Q6_K_XL)
context65,536
kv cachef16
thinkingon
think limitnone
samplingtemperature 0.7 · top_p 0.8 · top_k 20 · min_p 0 · presence_penalty 1.5
clientv0.4.85+101
Set those in Unsloth Studio, then:
llm-benchmark benchmark --model "unsloth/Qwen3.5-4B-MTP-GGUF" --tool "Unsloth Studio"The client prompts for context length and thinking mode, and for the KV cache dtype on Unsloth Studio. Sampling is left at the model default — the values above are what the tool reported using, not overrides the client sent.