A coding-focused version of Qwen3.6 35B-A3B that fits entirely on a 12 GB GPU. It keeps 187 of the 256 experts in each layer: the ones the full model uses most for code, thinking and tool calls. It was made with TruceAI, without any training.
| Qwen3.6 35B-A3B (UD-IQ3_XXS) | This model | |
|---|---|---|
| Experts per layer | 256 | 187 |
| File size | 13.2 GB | 10.2 GB |
| Fits in 12 GB of VRAM | no | yes, with a 16K context |
| Writing speed on an RTX 3060 12 GB (TruceAI engine) | slower: part of it runs from system RAM | 52-58 tokens/s |
Router rows travel with their experts, so the model routes among the experts it kept. No weights were retrained or edited.
On a 20-problem Python test with hidden unit tests (3 tries each, thinking off, Qwen's recommended sampling), it solved 58 of 60:
| Version | Solved |
|---|---|
| 128 experts, cut from the Q3_K_M file | 53 / 60 |
| 160 experts, cut from the UD-IQ3_XXS file | 57 / 60 |
| This model: 187 experts | 58 / 60 |
This is a small test, so treat differences of a few points as noise. Expect it to be weaker than the full model outside coding: languages other than English, general knowledge and creative writing.
TruceAI-Qwen3.6-Code-26B-A3.5B-UD-IQ3_XXS.gguf: the model. Its experts keep the formats of Unsloth's UD-IQ3_XXS file (mostly IQ2_S and IQ3_XXS); everything else is unchanged.mmproj-F16.gguf: Qwen3.6's image encoder, unchanged. A simple picture test was answered correctly.Made for TruceAI, which applies Qwen's recommended sampling (temperature 0.6, top_p 0.95, top_k 20, presence penalty 1.5). The file uses the standard qwen35moe architecture with 187 experts; other GGUF runners have not been tested. On a 12 GB card, keep the context at about 16K tokens so everything stays in VRAM.
Apache 2.0, the same license as Qwen3.6.
Qwen/Qwen3.6-35B-A3B