House recipes and cited cards
Models
The reverse of the GPU pages: pick a model, see the cited card and which Radeon setups fit. the house two daily drivers first, then cited open-weight pages for 16–64 GB Radeon boxes and the 320B / 18B GLM offload case. Those older 2.5-class and 8B names stay off this site. Q3 appears only as estimate/community unless we measured it. A name with no official card and GGUF listing stays contribute — we do not invent a shopper page. The Qwen compare table lists every current 3.6 / 3.8 file. MoE vs dense — why 35B-A3B fits one 32 GB R9700 — lives on /moe.
How we test — Measured cells cite official amdai-bench/run@1 cards. Estimates stay labeled. Q3 is estimate-only. Q1/Q2 stay off.
Daily driver on one R9700
Qwen3.6-35B-A3B
Official Qwen/Qwen3.6-35B-A3B · GGUF unsloth/Qwen3.6-35B-A3B-GGUF · 35B total / 3B active
the house daily driver: Unsloth UD-Q5_K_XL on llama.cpp Vulkan, one R9700, 262K, f16 KV. Every tok/s and measured GB comes from official amdai-bench/run@1 cards. Empty cells stay contribute. Q3 is a labeled estimate.
| Model | Setup | Backend | Quant | Ctx | Decode | Evidence |
|---|---|---|---|---|---|---|
| Qwen3.6-35B-A3B Qwen3.6-35B-A3B-Q4_K_M.gguf | 1× r9700 | llama.cpp vulkan | Q4_K_M | short | 117.69 | Tested on our rig |
| Qwen3.6-35B-A3B Qwen3.6-35B-A3B-Q5_K_M.gguf | 1× r9700 | llama.cpp vulkan | Q5_K_M | short | 109.73 | Tested on our rig |
| Qwen3.6-35B-A3B Qwen3.6-35B-A3B-UD-Q5_K_XL.gguf | 1× r9700 | llama.cpp vulkan | UD-Q5_K_XL | 256K | 44.4 | Tested on our rig |
| Qwen3.6-35B-A3B Qwen3.6-35B-A3B-UD-Q5_K_XL.gguf | 1× r9700 | llama.cpp vulkan | UD-Q5_K_XL | 32K | 90.52 | Tested on our rig |
| Qwen3.6-35B-A3B Qwen3.6-35B-A3B-UD-Q5_K_XL.gguf | 1× r9700 | llama.cpp vulkan | UD-Q5_K_XL | short | 102.39 | Tested on our rig |
Fast agent when Strata is up
Qwen 3.8 Flash
Official Qwen/Qwen3.8-Flash-Next · GGUF unsloth/Qwen3.8-Flash-Next-GGUF · 125B total / 6B active
the house fast agent: Unsloth UD-Q6_K_XL. Two 32 GB cards plus 96 GB DDR5 run Q4 at 256K — see Flash requirements. Strata is the cited Hermes path (~32.6 tok/s); llama.cpp is first on this page (~16 tok/s). Two-card Q6 lives on Flash daily. Untested cells say contribute. Q3 is a labeled estimate, not a shopper pick.
| Model | Setup | Backend | Quant | Ctx | Decode | Evidence |
|---|---|---|---|---|---|---|
| Qwen 3.8 Flash UD-Q6_K_XL (folder) | 2× r9700 | llama.cpp vulkan | UD-Q6_K_XL | 262K | 16 | Tested on our rig |
| Qwen 3.8 Flash UD-Q6_K_XL (folder) | 2× r9700 | Strata hip | UD-Q6_K_XL | 262K | 32.6 | Tested on our rig |
16 GB MoE — cited, not measured
gpt-oss 20B
Official openai/gpt-oss-20b · GGUF unsloth/gpt-oss-20b-GGUF · 21.5B total / 3.6B active
OpenAI’s 16 GB-class open-weight MoE. Official card: 21B parameters with 3.6B active; config.json lists 131,072 tokens. Q6 / Q5 / Q4 file sizes come from unsloth/gpt-oss-20b-GGUF. Q3 is a labeled estimate. Decode is not yet measured on our R9700s. llama.cpp first.
Not yet measured on our R9700s — open the cited page.
16–24 GB dense — cited, not measured
Gemma 3 27B
Official google/gemma-3-27b-it · GGUF unsloth/gemma-3-27b-it-GGUF · 27.4B total / 27.4B active
Google’s 27B instruct file for 16–24 GB Radeon boxes. Official card: 27B parameters, 128K input context. Q6 / Q5 / Q4 file sizes come from unsloth/gemma-3-27b-it-GGUF. Q3 is a labeled estimate. Decode is not yet measured on our R9700s. llama.cpp first.
Not yet measured on our R9700s — open the cited page.
32–64 GB dense — cited, not measured
Llama 3.3 70B
Official meta-llama/Llama-3.3-70B-Instruct · GGUF bartowski/Llama-3.3-70B-Instruct-GGUF · 70.6B total / 70.6B active
Meta’s 70B instruct file for 32–64 GB cards, or RAM offload on 16–24 GB. Official card context is 128K tokens. Q6 / Q5 / Q4 file sizes come from bartowski/Llama-3.3-70B-Instruct-GGUF. Q3 is a labeled estimate. Decode is not yet measured on our R9700s. llama.cpp first.
Not yet measured on our R9700s — open the cited page.
320B / 18B MoE — cited, not measured
GLM 5.3 Flash
Official zai-org/GLM-5.3-Flash · GGUF unsloth/GLM-5.3-Flash-GGUF · 320B total / 18B active
Z.ai’s 320B / 18B active multimodal MoE. Official card and config.json list 1,048,576 tokens. Q6 / Q5 / Q4 folder sizes come from unsloth/GLM-5.3-Flash-GGUF. Q3 is a labeled estimate. Consumer cards offload. Decode is not yet measured on our R9700s. llama.cpp first.
Not yet measured on our R9700s — open the cited page.
Dense 27B — cited, not measured
Qwen3.6 27B
Official Qwen/Qwen3.6-27B · GGUF unsloth/Qwen3.6-27B-GGUF · 27.78B total / 27.78B active
Qwen’s dense 27B file for 16–24 GB Radeon boxes. Official card lists 262,144 tokens. Q6 / Q5 / Q4 file sizes come from unsloth/Qwen3.6-27B-GGUF. Q3 is a labeled estimate. Decode is not yet measured on our R9700s. llama.cpp first.
Not yet measured on our R9700s — open the cited page.
Dense 27B — cited, not measured
Qwen3.8 27B
Official Qwen/Qwen3.8-27B · GGUF unsloth/Qwen3.8-27B-GGUF · 27.8B total / 27.8B active
Qwen’s dense 27B file for 16–24 GB Radeon boxes. Official card lists 262,144 tokens. Q6 / Q5 / Q4 file sizes come from unsloth/Qwen3.8-27B-GGUF. Q3 is a labeled estimate. Decode is not yet measured on our R9700s. llama.cpp first.
Not yet measured on our R9700s — open the cited page.
Contribute — no cited row yet
AMDai publishes a cited estimate page only when the official card and a GGUF listing already exist.
Names without that stay contribute. Abliterated files stay on /abliterated.
Send a measured amdai-bench/run@1 card if you have one of these on a Radeon.