13 builds of Qwen 3 4B measured on real machines — different packagers, precisions and engines, each a different file with its own numbers. The best each engine proved is first; every build has its own page with the machines and signatures behind it.
trained 40K · proved 32K context proved by unsloth IQ3_XXS
140.7 tok/s peak, by Qwen Q4_K_M
proved 64K context proved by qwen
proved by at least one build: Tool calling · Tool loop finished · Faithful to tool results · Chained tool calls · Structured output · Sustained output · Extended thinking · Vision · Counts in images · Spatial reasoning · Writes code · Reasons about code · Synthesises sources · Follows instructions
| build | publisher | precision | engine | whose work | context proved | best tok/s | machines |
|---|---|---|---|---|---|---|---|
| Qwen/Qwen3-4B | qwen | — | vllm | — | 65,536 | — | 1 |
| hf.co/unsloth/Qwen3-4B-GGUF:IQ3_XXS | unsloth | IQ3_XXS | ollama | quantizer | 32,768 | 62.9 | 4 |
| hf.co/Qwen/Qwen3-4B-GGUF:Q5_0 | Qwen | Q5_0 | ollama | vendor | 32,768 | 48.5 | 3 |
| hf.co/Qwen/Qwen3-4B-GGUF:Q5_K_M | Qwen | Q5_K_M | ollama | vendor | 32,768 | 47.0 | 2 |
| hf.co/Qwen/Qwen3-4B-GGUF:Q8_0 | Qwen | Q8_0 | ollama | vendor | 32,768 | 38.9 | 2 |
| hf.co/Qwen/Qwen3-4B-GGUF:Q4_K_M | Qwen | Q4_K_M | ollama | vendor | 16,384 | 140.7 | 2 |
| hf.co/unsloth/Qwen3-4B-GGUF:BF16 | unsloth | BF16 | ollama | quantizer | 16,384 | 80.8 | 1 |
| hf.co/unsloth/Qwen3-VL-4B-Instruct-GGUF:Q4_0 | unsloth | Q4_0 | ollama | — | 4,096 | 40.9 | 4 |
| hf.co/unsloth/Qwen3-VL-4B-Instruct-GGUF:Q4_1 | unsloth | Q4_1 | ollama | — | 4,096 | 37.0 | 2 |
| hf.co/unsloth/Qwen3-VL-4B-Instruct-GGUF:IQ4_XS | unsloth | IQ4_XS | ollama | — | 4,096 | 6.4 | 1 |
| hf.co/unsloth/Qwen3-4B-GGUF:IQ1_S | unsloth | IQ1_S | ollama | quantizer | — | 90.4 | 3 |
| hf.co/unsloth/Qwen3-4B-GGUF:Q4_1 | unsloth | Q4_1 | ollama | quantizer | — | 19.4 | 2 |
| hf.co/unsloth/Qwen3-VL-4B-Instruct-GGUF:Q2_K | unsloth | Q2_K | ollama | — | — | 11.9 | 1 |