2× RTX 3090
no documented speed
Intelligenceno independent value
- Experimental
Local AI · Recipe · 1× Strix Halo
Qwen3 30B-A3B UD-Q4_K_XL with llama.cpp on 1× Strix Halo
Qwen3 30B-A3B UD-Q4_K_XL with llama.cpp on 1× Strix Halo: 85.1 tok/s according to strixhalo.wiki (dataset as of Nov 3, 2026).
by deseven/lhl (strixhalo.wiki)
Intelligence (original model) · Artificial Analysis8with thinkingno thinking 7
Engine
- Engine
- llama.cpp Build 2025 (nicht angegeben)
- Quantization
- GGUF UD-Q4_K_XL
- Model family
- Qwen3
- Context
- —
- Parameters
- 30B MoE / 3B active
- Creator
- deseven/lhl (strixhalo.wiki)
- GitHub stars
- 73
- Repo updated
- Aug 31, 2026
Measurements
Every sourced value of this recipe, each with its condition and source. Bars relative to the largest value in the group.
Decode · 1 request
Prefill by context
What you need
- Hardware
- 1 × AMD Strix Halo
- Weights
- Main weightsunsloth/Qwen3-30B-A3B-GGUF Apache-2.0
- Engine
- llama.cpp Build 2025 (nicht angegeben)
Related recipes
1× Strix Halo
28.4tok/sEveryday
1 request, prose prompt, no speculative decoding Source
Intelligenceno independent value
- License unclear
1× Strix Halo
41.3tok/sMixed
1 request, mixed prompt set, with ndt=3 Source
Intelligence40with thinking
- Custom kernel required
- License unclear
1× Strix Halo
70.2tok/sPeak
1 request, synthetic prompt Source
Intelligence18with thinkingno thinking 15
1× Strix Halo
16.3tok/sPeak
1 request, code prompt Source
Intelligence34with thinking
- ≤ 3 bit: quantization may cost quality
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.