Qwen 3.6 · Qwen 3.8 · one row per file

Every current Qwen, side by side

One row per model × quant for the 3.6 and 3.8 families we actually talk about: 35B-A3B, 27B, Flash, Coder390, and the educational Huihui 27B abliterated study row. Those older 2.5-class and 8B names stay off this site. Why 35B-A3B fits one 32 GB card while dense 27B and Flash do not is the MoE hub. Decode and prefill fill only from verified data/runs/*.jsonl cards. File size is a Hugging Face listing when we have one. VRAM is the measured peak when a run card recorded it; otherwise the calculator estimate.

Open the calculator Compare 35B cards Run cards MoE on Radeon House models 35B model page Flash model page Sampling settings Chat templates 35B GGUF Flash GGUF Coder390 GGUF How we test

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 · 35B llama.cpp · 262K

Minimum system RAM needed

42.9 GB

Measured · 42.9 GiB raw · 256K context

local-desktop-20261012-2314-35b-udq5kxl-1gpu-256k-llamacpp-f16

RAM needed at full context

42.9 GB

Measured · 42.9 GiB raw · 256K full context

local-desktop-20261012-2314-35b-udq5kxl-1gpu-256k-llamacpp-f16

Measured peak system RAM on the local desktop: 42.9 GB (Linux GiB labeled GB). Measured, pending re-verify after RAM fix.

Breakdown
Offloaded weights ·
0 GiB
Separable embed in RAM ·
0 GiB
KV in RAM at selected ·
0 GiB
KV in RAM at full context ·
0 GiB
Runtime / backend ·
0.85 GiB
OS / desktop baseline ·
8.00 GiB
Headroom ·
16.0 GiB

Fast agent · Flash · 262K

Minimum system RAM needed

192 GB kit

Estimate · 159.9 GiB raw · 256K context

RAM needed at full context

192 GB kit

Estimate · 159.9 GiB raw · 256K full context

Estimate. Offloaded weights + KV in RAM + runtime + OS/desktop + headroom, rounded up to a 192 GB DIMM kit. The local desktop ships 256 GB DDR5; measured host-RAM usage is not recorded until Local AI numbers arrive.

Breakdown
Offloaded weights ·
125.4 GiB
Separable embed in RAM ·
0 GiB
KV in RAM at selected ·
7.61 GiB
KV in RAM at full context ·
7.61 GiB
Runtime / backend ·
2.88 GiB
OS / desktop baseline ·
8.00 GiB
Headroom ·
16.0 GiB

Measured on our rig, 2026-10

35B decode, measured

Qwen3.6-35B-A3B on one R9700 vs both R9700s, llama.cpp, f16 KV. One R9700 versus a split across both cards, per context. Cells that we still do not have stay not measured. Units match the hardware labels (GB = Linux GiB). Full hub: benchmarks. Filterable cards: comparable results.

Qwen3.6-35B-A3B llama.cpp decode, prefill, VRAM, and system RAM from local desktop run cards — one R9700 versus a split.
Quant Context One R9700 Both R9700s (split)
Prefill Decode VRAM / card VRAM total System RAM Prefill Decode VRAM / card VRAM total System RAM
UD-Q5_K_XL short 2761.6 tok/s 102.39 tok/s 25.45 GB 25.97 GB 37.32 GB 2637.3 tok/s 82.78 tok/s 13.28 / 13.11 GB 26.39 GB 37.6 GB
UD-Q5_K_XL 32K 1937.9 tok/s 90.52 tok/s 26.16 GB 26.7 GB 37.83 GB 1916 tok/s 74.8 tok/s 13.77 / 13.64 GB 27.41 GB 38.16 GB
UD-Q5_K_XL 128K 942.8 tok/s 66.43 tok/s 28.22 GB 28.82 GB 40.18 GB 1071.1 tok/s 57.94 tok/s 15.47 / 15.48 GB 30.94 GB 41.02 GB
UD-Q5_K_XL 256K 519.2 tok/s 44.4 tok/s 28.73 GB 29.34 GB 42.9 GB 671.7 tok/s 42.03 tok/s 17.8 / 17.99 GB 35.74 GB 44.71 GB
Q5_K_M short 2710.8 tok/s 109.73 tok/s 24.16 GB 24.68 GB 35.83 GB 2647.2 tok/s 88.19 tok/s 12.71 / 12.41 GB 25.13 GB 37.17 GB
Q5_K_M 32K 1902.6 tok/s 96.21 tok/s 24.87 GB 25.39 GB 36.46 GB 1920.9 tok/s 78.91 tok/s 13.2 / 12.95 GB 26.13 GB 37.8 GB
Q5_K_M 128K 940.3 tok/s 69.64 tok/s 26.93 GB 27.43 GB 38.65 GB 1067.3 tok/s 59.86 tok/s 14.98 / 14.79 GB 29.76 GB 40.73 GB
Q5_K_M 256K 541.7 tok/s 40.54 tok/s 28.56 GB 29.15 GB 42.79 GB 664.6 tok/s 42.33 tok/s 17.45 / 17.3 GB 34.54 GB 46.21 GB
Q4_K_M short 2733.9 tok/s 117.69 tok/s 20.84 GB 21.43 GB 33.74 GB 2660.4 tok/s 93.28 tok/s 10.83 / 10.85 GB 21.69 GB 33.74 GB
Q4_K_M 32K 1902.1 tok/s 101.88 tok/s 21.55 GB 22.17 GB 34.58 GB 1933.1 tok/s 82.57 tok/s 11.34 / 11.39 GB 22.72 GB 34.39 GB
Q4_K_M 128K 942.1 tok/s 72.48 tok/s 23.61 GB 24.27 GB 36.5 GB 1078.4 tok/s 62.12 tok/s 13.03 / 13.23 GB 26.26 GB 38.04 GB
Q4_K_M 256K 558.5 tok/s 47.77 tok/s 26.36 GB 27.03 GB 39.05 GB 673 tok/s 42.88 tok/s 15.28 / 15.74 GB 31.02 GB 40.84 GB
Q6_K short 2623.6 tok/s 103.23 tok/s 28.78 GB 29.26 GB 40.15 GB 2572.1 tok/s 83.5 tok/s 15.01 / 14.63 GB 29.64 GB 38.22 GB
Q6_K Partly offloaded to DDR5 on 1 GPU 32K 1683 tok/s 71.19 tok/s 28.68 GB 29.17 GB 44.01 GB 1890.7 tok/s 75.43 tok/s 15.5 / 15.16 GB 30.66 GB 38.98 GB
Q6_K Partly offloaded to DDR5 on 1 GPU 128K 800.3 tok/s 54.51 tok/s 28.78 GB 29.25 GB 45.43 GB 1059.7 tok/s 58.04 tok/s 17.21 / 17 GB 34.21 GB 41.41 GB
Q6_K Partly offloaded to DDR5 on 1 GPU 256K 471.6 tok/s 37.78 tok/s 28.81 GB 29.3 GB 43.69 GB 666.2 tok/s 42.2 tok/s 19.46 / 19.51 GB 38.96 GB 45.15 GB
  • Cells come from verified amdai-bench/run@1 cards on the local desktop (overnight sweep 2026-10-12, llama.cpp c060ca974 Vulkan). A blank stays not measured. GB is the Linux GiB value.
  • Side by side: prefill, decode, VRAM per card, VRAM total, and peak system RAM. At short context one R9700 decodes faster than splitting the same 35B across both cards (UD-Q5_K_XL 102.39 vs 82.78 tok/s at 4K).
  • Measured, pending re-verify after RAM fix. DDR5-5200 later proved unstable; a kernel oops at 00:14 CT aborted the sweep. All 33 cards predate the oops. Each card already notes one transient bad read in earlier sha256 passes.
  • Q6_K single-GPU 4K table cell is 103.23 tok/s from the 23:55 card (the 4K point in the complete Q6_K single-GPU series). The earlier 23:26 card at 82.62 tok/s stays on /results and is not averaged into the cell.
  • Q6_K at 32K and up on one GPU is partly offloaded to DDR5.
  • Flash, Strata, and Q3 runs are pending a rerun after AMDai stabilizes the DDR5.

Q3 sizes on the same card (estimate / community)

Not measured on the local desktop. File sizes from unsloth/Qwen3.6-35B-A3B-GGUF. Quality deltas are not first-party.

Q3 35B file sizes from the Unsloth listing. Decode and VRAM are not measured on the local desktop.
Quant Publisher listing Decode / VRAM
UD-IQ3_S 13.7 GB not measured · estimate / community
UD-Q3_K_S 15.4 GB not measured · estimate / community
UD-Q3_K_M 16.6 GB not measured · estimate / community
UD-Q3_K_XL 16.8 GB not measured · estimate / community

Comparison

Filter the grid

GPU count, RAM kit, and backend write the My Rig profile (same engine as the calculator). Fits-on-my-rig hides rows that need more VRAM + DDR5 than that profile. The URL is shareable. Without JavaScript the full table stays visible.

36 rows

GPU count

RAM size

Backend

My Rig default: 1× R9700 · 256 GB DDR5 · llama.cpp. Saved locally as amdai-my-rig, or pass a calculator share query.

Column picker

Hide dense columns on a small screen. Prefill stays off until you turn it on.

Qwen 3.6 · MoE · original

Qwen3.6-35B-A3B

UD-IQ3_S

File
13.7 GB
VRAM @ full
18.5 GiB est.
1× R9700
yes · est.
DDR5 offload
none · est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.6 · MoE · original

Qwen3.6-35B-A3B

UD-Q3_K_S

File
15.4 GB
VRAM @ full
20.1 GiB est.
1× R9700
yes · est.
DDR5 offload
none · est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.6 · MoE · original

Qwen3.6-35B-A3B

UD-Q3_K_M

File
16.6 GB
VRAM @ full
21.2 GiB est.
1× R9700
yes · est.
DDR5 offload
none · est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.6 · MoE · original

Qwen3.6-35B-A3B

UD-Q3_K_XL

File
16.8 GB
VRAM @ full
21.4 GiB est.
1× R9700
yes · est.
DDR5 offload
none · est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.6 · MoE · original

Qwen3.6-35B-A3B

Q4_K_M

File
19.8 GiB est.
VRAM @ full
27.03 GB
1× R9700
yes · est.
DDR5 offload
none · est.
Decode llama.cpp
117.69 @ short
Decode Strata
—

Hugging Face Run cards

Qwen 3.6 · MoE · original

Qwen3.6-35B-A3B

Q5_K_M

File
23.2 GiB est.
VRAM @ full
29.15 GB
1× R9700
yes · est.
DDR5 offload
none · est.
Decode llama.cpp
109.73 @ short
Decode Strata
—

Hugging Face Run cards

Qwen 3.6 · MoE · original

Qwen3.6-35B-A3B

UD-Q5_K_XL

File
26.6 GB
VRAM @ full
29.34 GB
1× R9700
tight · est.
DDR5 offload
none · est.
Decode llama.cpp
102.39 @ short
Decode Strata
—

Hugging Face Run cards

Qwen 3.6 · MoE · original

Qwen3.6-35B-A3B

Q6_K

File
26.7 GiB est.
VRAM @ full
29.3 GB
1× R9700
no · est.
DDR5 offload
0.67 GiB est.
Decode llama.cpp
103.23 @ 4K
Decode Strata
—

Hugging Face Run cards

Qwen 3.6 · MoE · original

Qwen3.6-35B-A3B

Q8_0

File
34.6 GiB est.
VRAM @ full
40.7 GiB est.
1× R9700
no · est.
DDR5 offload
8.69 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.6 · dense · original

Qwen3.6-27B

Q3_K_S

File
11.3 GiB est.
VRAM @ full
28.3 GiB est.
1× R9700
yes · est.
DDR5 offload
none · est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.6 · dense · original

Qwen3.6-27B

Q3_K_M

File
13.59 GB
VRAM @ full
29.6 GiB est.
1× R9700
yes · est.
DDR5 offload
none · est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.6 · dense · original

Qwen3.6-27B

UD-Q3_K_XL

File
14.47 GB
VRAM @ full
30.5 GiB est.
1× R9700
yes · est.
DDR5 offload
none · est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.6 · dense · original

Qwen3.6-27B

Q4_K_M

File
16.82 GB
VRAM @ full
32.7 GiB est.
1× R9700
no · est.
DDR5 offload
0.69 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.6 · dense · original

Qwen3.6-27B

Q5_K_M

File
19.51 GB
VRAM @ full
35.2 GiB est.
1× R9700
no · est.
DDR5 offload
3.23 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.6 · dense · original

Qwen3.6-27B

Q6_K

File
22.52 GB
VRAM @ full
38.1 GiB est.
1× R9700
no · est.
DDR5 offload
6.08 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.6 · dense · original

Qwen3.6-27B

Q8_0

File
27.5 GiB est.
VRAM @ full
44.7 GiB est.
1× R9700
no · est.
DDR5 offload
12.7 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · MoE · original

Qwen 3.8 Flash

UD-IQ3_XXS

File
81.96 GB
VRAM @ full
85.6 GiB est.
1× R9700
no · est.
DDR5 offload
53.6 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · MoE · original

Qwen 3.8 Flash

UD-Q3_K_XL

File
90 GB
VRAM @ full
93.2 GiB est.
1× R9700
no · est.
DDR5 offload
61.2 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · MoE · original

Qwen 3.8 Flash

UD-Q4_K_XL

File
111 GB
VRAM @ full
113.1 GiB est.
1× R9700
no · est.
DDR5 offload
81.1 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · MoE · original

Qwen 3.8 Flash

UD-Q5_K_XL

File
158.3 GB
VRAM @ full
157.8 GiB est.
1× R9700
no · est.
DDR5 offload
125.8 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · MoE · original

Qwen 3.8 Flash

UD-Q6_K_XL

File
169 GB
VRAM @ full
167.9 GiB est.
1× R9700
no · est.
DDR5 offload
135.9 GiB est.
Decode llama.cpp
18.5 @ 8K
Decode Strata
32.8 @ 8K

Hugging Face Run cards

Qwen 3.8 · dense · original

Qwen3.8-27B

Q3_K_S

File
11.3 GiB est.
VRAM @ full
28.3 GiB est.
1× R9700
yes · est.
DDR5 offload
none · est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · dense · original

Qwen3.8-27B

Q3_K_M

File
13.15 GB
VRAM @ full
29.2 GiB est.
1× R9700
yes · est.
DDR5 offload
none · est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · dense · original

Qwen3.8-27B

UD-Q3_K_XL

File
13.15 GB
VRAM @ full
29.2 GiB est.
1× R9700
yes · est.
DDR5 offload
none · est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · dense · original

Qwen3.8-27B

Q4_K_M

File
16.46 GB
VRAM @ full
32.3 GiB est.
1× R9700
no · est.
DDR5 offload
0.35 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · dense · original

Qwen3.8-27B

Q5_K_M

File
19.77 GB
VRAM @ full
35.5 GiB est.
1× R9700
no · est.
DDR5 offload
3.48 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · dense · original

Qwen3.8-27B

Q6_K

File
21.98 GB
VRAM @ full
37.6 GiB est.
1× R9700
no · est.
DDR5 offload
5.57 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · dense · original

Qwen3.8-27B

Q8_0

File
27.5 GiB est.
VRAM @ full
44.7 GiB est.
1× R9700
no · est.
DDR5 offload
12.7 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · dense · original

Qwen3.8-27B Coder390

Q3 LynnStyle

File
17.3 GB
VRAM @ full
33.1 GiB est.
1× R9700
no · est.
DDR5 offload
1.14 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · dense · original

Qwen3.8-27B Coder390

Q4 LynnStyle

File
19.6 GB
VRAM @ full
35.3 GiB est.
1× R9700
no · est.
DDR5 offload
3.32 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · dense · original

Qwen3.8-27B Coder390

Q6_K-MTP

File
23.2 GB
VRAM @ full
38.7 GiB est.
1× R9700
no · est.
DDR5 offload
6.72 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · dense · original

Qwen3.8-27B Coder390

Q8_0-MTP

File
29.1 GB
VRAM @ full
44.3 GiB est.
1× R9700
no · est.
DDR5 offload
12.3 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · dense · abliterated

Huihui-Qwen3.8-27B-abliterated

Q3_K_M

File
12.7 GiB est.
VRAM @ full
29.6 GiB est.
1× R9700
yes · est.
DDR5 offload
none · est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · dense · abliterated

Huihui-Qwen3.8-27B-abliterated

Q4_K_M

File
15.7 GiB est.
VRAM @ full
32.7 GiB est.
1× R9700
no · est.
DDR5 offload
0.72 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · dense · abliterated

Huihui-Qwen3.8-27B-abliterated

Q5_K_M

File
18.4 GiB est.
VRAM @ full
35.5 GiB est.
1× R9700
no · est.
DDR5 offload
3.48 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

Qwen 3.8 · dense · abliterated

Huihui-Qwen3.8-27B-abliterated

Q6_K

File
21.2 GiB est.
VRAM @ full
38.3 GiB est.
1× R9700
no · est.
DDR5 offload
6.34 GiB est.
Decode llama.cpp
—
Decode Strata
—

Hugging Face Run cards

esc

Type to filter. Enter opens the first match. Full-text search lives on /search.