Gaps · 2026-10
Wanted recipes
Gaps only. A wanted row is not a speed claim. No invented tok/s or prices.
Most-requested Radeon gaps from the house stacks. A wanted row is not a claim that the combo fits. Submit a recipe.
-
Radeon RX 9070 XT
Qwen3.6-35B-A3B
Same gfx1201 as the R9700, half the VRAM. We have no one-card recipe.
llama.cpp Vulkan, Q5 or Q4 if it fits, short context first. Log plus screenshot.
-
Radeon RX 7900 XTX
Qwen3.6-35B-A3B
Best RDNA 3 consumer VRAM. Community runs exist; we have not reproduced them.
llama.cpp first. Q3 stays estimate/community unless you measured it here.
-
Radeon AI PRO R9700 ×2
Qwen3.6-35B-A3B
The box has two cards. Seeded recipes are one-card only.
Split-server vs two copies. Do not send a number we did not ask you to measure.
-
Radeon AI PRO R9700
Qwen 3.8 Flash
Seeded Flash proofs are UD-Q6_K_XL only (Strata 32.6 vs llama.cpp 16).
Q5 or Q4 offload run with host RAM measured. Q3 is allowed as estimate/community.
-
Radeon AI PRO R9700
ComfyUI SDXL / Flux-class
ComfyUI already runs on the local desktop. AMDai has no first-party amdai-bench/run@1 image-gen card, so speeds stay not yet measured by AMDai.
First-party image-gen card on one R9700. Log plus screenshot. Do not invent seconds-per-image or VRAM megabytes.
-
Radeon AI PRO R9700
whisper.cpp STT / Piper TTS
whisper.cpp Vulkan and Piper are the AMDai first speech path. AMDai has no first-party amdai-bench/run@1 speech card, so speeds stay not yet measured by AMDai.
First-party whisper.cpp Vulkan card on one R9700. Log plus screenshot. Do not invent real-time factor or tok/s.
-
Radeon AI PRO R9700
llama.cpp embeddings / local RAG
llama.cpp --embeddings is the AMDai first RAG embed path. AMDai has no first-party amdai-bench/run@1 retrieval card, so speeds and recall stay not yet measured by AMDai.
First-party llama.cpp Vulkan embedding card on one R9700 next to the daily-driver chat sidecar. Log plus screenshot. Do not invent retrieval scores or tok/s.