README
Qwen3.6-35B-A3B — Q4_0-ROCmFP4_FAST (GGUF)
A ROCmFP4_FAST (4.25 bpw) quantization of Qwen3.6-35B-A3B, tuned for the AMD Strix Halo
iGPU (Radeon 8060S, gfx1151, RDNA3.5). Quantized directly from the BF16 source with the MTP
(nextn) head preserved.
⚠️ Requires the ROCmFPX fork — will NOT load in stock llama.cpp
ROCmFP4 is a custom weight format built on a Q4_0 base. This file only runs on the
ROCmFPX fork of llama.cpp: https://github.com/charlie12345/ROCmFPX. Loading it in
upstream llama.cpp, Ollama, LM Studio, etc. will fail. Build the fork
(scripts/build-strix-rocmfp4-mtp.sh) and use its llama-server.
Why ROCmFP4_FAST
Measured on a Framework Desktop (Ryzen AI Max+ 395, Radeon 8060S gfx1151, ROCm 7.2, Vulkan
backend), against Q4_K_M of the same base weights, plain decode:
| format | decode (short) | decode (~40k ctx) | prefill | file size |
|---|---|---|---|---|
| Q4_K_M | 67.6 t/s | 55.9 t/s | 211 / 746 t/s | 21 GB |
| ROCmFP4_FAST | 73.0 t/s | 59.5 t/s | 232 / 784 t/s | 18 GB |
| delta | +8% | +6.4% | +5–10% | −14% |
Single-user, -fa on, f16 KV, Vulkan. Your mileage varies by driver/build.
Quality holds vs higher precision: in blind-graded coding benchmarks this FP4_FAST quant scored within noise of the BF16 source (same band), so the format costs little to no accuracy — the big quality lever is the reasoning mode below, not the quant.
Files
| filename | quant | notes |
|---|---|---|
Qwen3.6-35B-A3B-Q4_0-ROCmFP4_FAST.gguf | Q4_0_ROCMFP4_FAST (4.25 bpw) | MTP/nextn head included |
Run
# built ROCmFPX fork
llama-server \
--model Qwen3.6-35B-A3B-Q4_0-ROCmFP4_FAST.gguf \
-dev Vulkan0 -ngl 999 -fa on -c 131072 \
-ctk f16 -ctv f16 --jinja --host 0.0.0.0 --port 8080
The GGUF carries an MTP (nextn) head, so speculative decoding works with
--spec-type draft-mtp. On this hardware, conservative draft settings work best
(e.g. --spec-draft-n-max 4 --spec-draft-p-min 0.55); high n_max tends to regress decode.
Sampling & reasoning (thinking on/off)
Recommended sampling (validated on this stack):
--temp 0.7 --top-p 0.8 --top-k 20 --min-p 0 --presence-penalty 0.
Reasoning is task-dependent — it is the single most important serving choice for this model. Blind-graded coding benchmarks on this quant (0–100, same-grader; unknown-cause debugging suite + fully-specified implementation suite):
| task type | thinking OFF | thinking ON |
|---|---|---|
| fully-specified implementation ("build X, constraints 1–N, output full file") | ~88 | ~81 |
| unknown-cause debugging / hard algorithms / judgment | ~38 | ~66 |
- Default: thinking OFF —
--reasoning off --reasoning-format deepseek --reasoning-budget 0. Best for agentic / well-specified coding: ~8× lower latency, cleaner output, and it scores higher than thinking-on on well-specified work. The right mode for delegated "implement X with these constraints" tasks. - Debugging / hard reasoning: thinking ON —
--reasoning-format deepseek --reasoning-budget 4096. For "find why this misbehaves", novel algorithms, review/critique, or planning, reasoning is worth roughly +28 points and the extra latency. Without it the model tends to loop and leak deliberation into the answer.
Keep --reasoning-format deepseek in both modes — it ensures <think> blocks are parsed and
closed correctly (Qwen3.6 can otherwise leak an unclosed <think> into the output).
Provenance & credits
- Base weights: unsloth/Qwen3.6-35B-A3B-MTP-GGUF (BF16 source).
- Quant format & tooling: ROCmFPX (
llama-quantize ... Q4_0_ROCMFP4_FAST). - License inherited from the base model.
Quantized 2026-07 for a Strix Halo inference stack.
Contributors
7 commits
Similar repos
README
Qwen3.6-35B-A3B — Q4_0-ROCmFP4_FAST (GGUF)
A ROCmFP4_FAST (4.25 bpw) quantization of Qwen3.6-35B-A3B, tuned for the AMD Strix Halo
iGPU (Radeon 8060S, gfx1151, RDNA3.5). Quantized directly from the BF16 source with the MTP
(nextn) head preserved.
⚠️ Requires the ROCmFPX fork — will NOT load in stock llama.cpp
ROCmFP4 is a custom weight format built on a Q4_0 base. This file only runs on the
ROCmFPX fork of llama.cpp: https://github.com/charlie12345/ROCmFPX. Loading it in
upstream llama.cpp, Ollama, LM Studio, etc. will fail. Build the fork
(scripts/build-strix-rocmfp4-mtp.sh) and use its llama-server.
Why ROCmFP4_FAST
Measured on a Framework Desktop (Ryzen AI Max+ 395, Radeon 8060S gfx1151, ROCm 7.2, Vulkan
backend), against Q4_K_M of the same base weights, plain decode:
| format | decode (short) | decode (~40k ctx) | prefill | file size |
|---|---|---|---|---|
| Q4_K_M | 67.6 t/s | 55.9 t/s | 211 / 746 t/s | 21 GB |
| ROCmFP4_FAST | 73.0 t/s | 59.5 t/s | 232 / 784 t/s | 18 GB |
| delta | +8% | +6.4% | +5–10% | −14% |
Single-user, -fa on, f16 KV, Vulkan. Your mileage varies by driver/build.
Quality holds vs higher precision: in blind-graded coding benchmarks this FP4_FAST quant scored within noise of the BF16 source (same band), so the format costs little to no accuracy — the big quality lever is the reasoning mode below, not the quant.
Files
| filename | quant | notes |
|---|---|---|
Qwen3.6-35B-A3B-Q4_0-ROCmFP4_FAST.gguf | Q4_0_ROCMFP4_FAST (4.25 bpw) | MTP/nextn head included |
Run
# built ROCmFPX fork
llama-server \
--model Qwen3.6-35B-A3B-Q4_0-ROCmFP4_FAST.gguf \
-dev Vulkan0 -ngl 999 -fa on -c 131072 \
-ctk f16 -ctv f16 --jinja --host 0.0.0.0 --port 8080
The GGUF carries an MTP (nextn) head, so speculative decoding works with
--spec-type draft-mtp. On this hardware, conservative draft settings work best
(e.g. --spec-draft-n-max 4 --spec-draft-p-min 0.55); high n_max tends to regress decode.
Sampling & reasoning (thinking on/off)
Recommended sampling (validated on this stack):
--temp 0.7 --top-p 0.8 --top-k 20 --min-p 0 --presence-penalty 0.
Reasoning is task-dependent — it is the single most important serving choice for this model. Blind-graded coding benchmarks on this quant (0–100, same-grader; unknown-cause debugging suite + fully-specified implementation suite):
| task type | thinking OFF | thinking ON |
|---|---|---|
| fully-specified implementation ("build X, constraints 1–N, output full file") | ~88 | ~81 |
| unknown-cause debugging / hard algorithms / judgment | ~38 | ~66 |
- Default: thinking OFF —
--reasoning off --reasoning-format deepseek --reasoning-budget 0. Best for agentic / well-specified coding: ~8× lower latency, cleaner output, and it scores higher than thinking-on on well-specified work. The right mode for delegated "implement X with these constraints" tasks. - Debugging / hard reasoning: thinking ON —
--reasoning-format deepseek --reasoning-budget 4096. For "find why this misbehaves", novel algorithms, review/critique, or planning, reasoning is worth roughly +28 points and the extra latency. Without it the model tends to loop and leak deliberation into the answer.
Keep --reasoning-format deepseek in both modes — it ensures <think> blocks are parsed and
closed correctly (Qwen3.6 can otherwise leak an unclosed <think> into the output).
Provenance & credits
- Base weights: unsloth/Qwen3.6-35B-A3B-MTP-GGUF (BF16 source).
- Quant format & tooling: ROCmFPX (
llama-quantize ... Q4_0_ROCMFP4_FAST). - License inherited from the base model.
Quantized 2026-07 for a Strix Halo inference stack.
Similar repos
Linked in READMEs
Contributors
7 commits