Instructions to use Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4")
model = AutoModelForMultimodalLM.from_pretrained("Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4

SGLang

How to use Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4 with Docker Model Runner:
```
docker model run hf.co/Capicua25x/Qwen3.8-27B-MXFP4-Quark-RDNA4
```

Capicua25x commited on Aug 17

Commit

40ecee5

verified ·

1 Parent(s): f5e378e

tau2: add airline (0.840 vs ref 0.760) and correct the telecom error count (1 of 114, not zero); drop the single-run quality comparison against AMD's build and state the structural difference instead

Browse files

Files changed (1) hide show

README.md +19 -9

README.md CHANGED Viewed

@@ -42,10 +42,15 @@ norm, embeddings, `lm_head`, routers/gates and the **entire vision path** stay b
 Verified by tensor inspection: a module counts as quantised only if it carries a real artifact
 (`weight_scale`, `weight_packed`, `qweight`, `weight_zero_point`).
-Keeping attention in bf16 is deliberate — it is what preserves the strict-format quality below
-while still getting the memory win, since the MLP stack is where the parameters actually are.
-For contrast, `amd/Qwen3.8-27B-Quark-AWQ-MXFP4` quantises the whole decoder including attention
-(W4A4) and measures 0.96/0.78 on GSM8K at 26.8 tok/s on this same hardware.
 - Format: MXFP4 (E2M1 + E8M0 scale per 32 weights), `pack_method: reorder`, `weight_format: real_quantized`
 - Size: **22.3 GB** across 18 shards (bf16 source ≈ 54 GB)
@@ -142,6 +147,7 @@ on this same box, differing only in KV cache dtype.
 | AIME 2025 | 30 | 0.9333 | 1.0000 | 0.9667 | **0.9333** |
 | AA-LCR (~107k-token prompts, judge-scored) | 100 | 0.780 | 0.800 ᵃ | 0.810 | **0.780** |
 | τ²-bench telecom (Pass^1) | 114 | 0.991 | 0.982 ᵇ | 0.939 | **0.868** |
 | HLE | 120 | 0.3083 | — | — | *running* |
 | SWE-bench Verified | 100 | — | — | — | *pending* |
 | Terminal-Bench Hard | 44 | — | — | — | *pending* |
@@ -150,11 +156,15 @@ on this same box, differing only in KV cache dtype.
 configuration's 131k window. Blended over the full 100 it reads 0.720.
 ᵇ Ran with thinking off by server default, so it is not strictly paired with the other columns.
-**The τ² result is the honest weak spot.** 0.868 against the bf16 reference's 0.991 is roughly
-14 simulations, with **zero agent errors and zero user-simulator errors** — so it is a capability
-gap on multi-turn tool use, not harness noise. Airline and retail domains are still running and
-will show whether telecom is representative. If you are picking a quant for agentic tool-calling
-specifically, weigh this cell above the others.
 Everything else is at or above bf16: GSM8K strict-match **+6 items**, GPQA **+8 items**, and
 long-context retrieval **identical** to bf16 at ~107k-token prompts.

 Verified by tensor inspection: a module counts as quantised only if it carries a real artifact
 (`weight_scale`, `weight_packed`, `qweight`, `weight_zero_point`).
+Keeping attention in bf16 is deliberate: the MLP stack is where the parameters are, so excluding
+attention costs little size and keeps those layers on the fast bf16 path.
+For structural comparison, `amd/Qwen3.8-27B-Quark-AWQ-MXFP4` quantises the decoder's attention as
+well — 496 quantised modules against 432 here, the difference being exactly the 16 full-attention
+layers' q/k/v/o — and is **AWQ-calibrated** (`algo_config.name = awq`) where this build is data-free
+RTN. Those are the two real differences. We have a single unrepeated n=50 GSM8K run against that
+build, which is not enough to publish a quality comparison from: strict-match moves by about
+±0.06 across seeds on this hardware, which is wider than any gap it showed.
 - Format: MXFP4 (E2M1 + E8M0 scale per 32 weights), `pack_method: reorder`, `weight_format: real_quantized`
 - Size: **22.3 GB** across 18 shards (bf16 source ≈ 54 GB)
 | AIME 2025 | 30 | 0.9333 | 1.0000 | 0.9667 | **0.9333** |
 | AA-LCR (~107k-token prompts, judge-scored) | 100 | 0.780 | 0.800 ᵃ | 0.810 | **0.780** |
 | τ²-bench telecom (Pass^1) | 114 | 0.991 | 0.982 ᵇ | 0.939 | **0.868** |
+| τ²-bench airline (Pass^1) | 50 | 0.760 | — | — | **0.840** |
 | HLE | 120 | 0.3083 | — | — | *running* |
 | SWE-bench Verified | 100 | — | — | — | *pending* |
 | Terminal-Bench Hard | 44 | — | — | — | *pending* |
 configuration's 131k window. Blended over the full 100 it reads 0.720.
 ᵇ Ran with thinking off by server default, so it is not strictly paired with the other columns.
+**τ² is domain-split, and the split is the finding.** On telecom this build scores 0.868 against
+the bf16 reference's 0.991 — about 14 simulations — but on airline it scores **0.840 against the
+reference's 0.760**, i.e. four items *ahead*. So multi-turn tool use is not uniformly degraded;
+telecom specifically is where it loses. Retail is still running and will add a third point.
+On telecom, 113 of 114 simulations ended normally and one hit the harness's error ceiling
+(`too_many_errors`, scored 0). That single run cannot account for a 14-item gap, so the telecom
+deficit is a real capability gap rather than harness noise — but it is one domain, not a blanket
+weakness. If you are choosing specifically for agentic tool-calling, weigh both cells.
 Everything else is at or above bf16: GSM8K strict-match **+6 items**, GPQA **+8 items**, and
 long-context retrieval **identical** to bf16 at ~107k-token prompts.