4× RTX 3090
145tok/sEveryday
1 request, prose prompt, with DFlash (per recipe) Source
Intelligence34with thinkingno thinking 20
- Experimental
Local AI · Recipe · 1× RTX 5090
Qwen3.8 27.32B UD-IQ2_XXS with llama.cpp on 1× RTX 5090
by notwitcheer
Intelligence (original model) · Artificial Analysis34with thinkingno thinking 20Original model's value. Heavily compressed (~2.1 bit), intelligence is likely much lower.
≤ 3 bit: quantization may cost quality
Engine
- Engine
- llama.cpp b9653
- Quantization
- GGUF UD-IQ2_XXS … Q8_0
- Model family
- Qwen3.8
- Context
- —
- Parameters
- 27.32B dense (hybrid linear attn)
- Creator
- notwitcheer
- GitHub stars
- 40
- Repo updated
- Sep 20, 2026
Measurements
Every sourced value of this recipe, each with its condition and source. Bars relative to the largest value in the group.
No sourced measurements for this recipe.
What you need
- Hardware
- 1 × NVIDIA RTX 5090
- Weights
- Main weightsunsloth/Qwen3.8-27B-GGUF Apache-2.0
- Engine
- llama.cpp b9653
Notes
What matters before you rebuild it.
- ≤ 3 bit: quantization may cost qualityAt 3 bit and below the model may answer noticeably worse than the original. The intelligence number refers to the original.
Sources
Related recipes
2× RTX 3090
123tok/sEveryday
1 request, prose prompt, with Draft Source
Intelligence34with thinkingno thinking 20
4× RTX 3090
85.3tok/sEveryday
1 request, prose prompt, with DFlash2 (per recipe) Source
Intelligence34with thinkingno thinking 20
2× RTX 3090
76.4tok/sEveryday
1 request, prose prompt, with MTP Source
Intelligence34with thinkingno thinking 20
1× RTX 3090
42.8tok/sEveryday
1 request, prose prompt, no speculative decoding Source
Intelligence34with thinkingno thinking 20
1× RTX 3090
39.7tok/sEveryday
1 request, prose prompt, no speculative decoding Source
Intelligence34with thinkingno thinking 20
Local AI in your company?
In a workshop we work out which models and which hardware fit your tasks, and build the first agent on your infrastructure.