Instructions to use FINAL-Bench/Armoring-Qwen3.6-27B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use FINAL-Bench/Armoring-Qwen3.6-27B with Transformers:

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

pipe = pipeline("text-generation", model="FINAL-Bench/Armoring-Qwen3.6-27B")

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("FINAL-Bench/Armoring-Qwen3.6-27B", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use FINAL-Bench/Armoring-Qwen3.6-27B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "FINAL-Bench/Armoring-Qwen3.6-27B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "FINAL-Bench/Armoring-Qwen3.6-27B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/FINAL-Bench/Armoring-Qwen3.6-27B

SGLang

How to use FINAL-Bench/Armoring-Qwen3.6-27B 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 "FINAL-Bench/Armoring-Qwen3.6-27B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "FINAL-Bench/Armoring-Qwen3.6-27B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "FINAL-Bench/Armoring-Qwen3.6-27B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "FINAL-Bench/Armoring-Qwen3.6-27B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use FINAL-Bench/Armoring-Qwen3.6-27B with Docker Model Runner:
```
docker model run hf.co/FINAL-Bench/Armoring-Qwen3.6-27B
```

VIDRAFT Armoring

Armoring-Qwen3.6-27B

This repository contains a model-card report for a VIDRAFT Attention Armoring study on Qwen/Qwen3.6-27B.

The repository is intentionally card-first. No full model weights, implementation artifact, internal transformation procedure, training command, internal mapping, or infrastructure path is included.

One-Screen Result

This is the balanced long-context efficiency card in the current armoring series. The central result is that the candidate preserved local format behavior while showing useful long-context cache economics after the break-even region.

What changed	Why it matters
Final quality ratio measured 1.0118x with 15 / 15 format preservation	The candidate recovered close to the reference behavior in this local validation lane.
16k net result measured 1.20x	The method starts to become useful after the long-context break-even region.
32k net result measured 1.35x	Long-context serving is the main target.
32k KV per sequence moved from 2048 MiB to 1158 MiB	The measured cache footprint fell by 43.5% in this paired check.
Estimated 80 GiB cache capacity moved from 12 to 22 sessions	Same-memory capacity improved by about 1.83x in the measured setup.

These figures should be read as long-context architecture-efficiency evidence, not as a broad capability claim.

Before / After Architecture View

Before and after architecture diagram for VIDRAFT Armoring

Standard Attention vs. Attention Armoring

Standard LLM attention architecture versus VIDRAFT attention armoring architecture

In a conventional Transformer-style attention block, the model projects the current hidden state into query, key, and value streams. During long-context generation, key and value states are retained in the KV cache so later tokens can attend back to earlier context. This is useful and standard, but it creates a direct memory bottleneck: the longer the context and the more concurrent sessions, the more cache state the serving system must keep alive.

VIDRAFT Attention Armoring keeps the public model interface intact while changing the long-context attention-state path inside the eligible attention block. The public point is not the private transformation procedure; the public point is the measured operational effect. In this 27B study, the paired measurements show a clear long-context tradeoff: short contexts can lose, while 16k and 32k contexts show positive net economics.

Public Transformation Summary

Layer of the system	Before	After Armoring
Base model family	Qwen3.6-27B-class model	Same base-family interface
Public architecture view	Original attention-state path	Attention subsystem is armored at a high level
Runtime pressure point	KV cache grows heavily with long context	KV-cache footprint is reduced in the tested 16k and 32k settings
Operational value	Memory pressure limits long-context serving	More same-memory cache capacity after the break-even region
Release posture	Standard model-card summary	Public-safe report; implementation details retained privately

Measured Summary

Measurement	Result
Immediate quality ratio after transformation	1.0197x
Final quality ratio	1.0118x
Best observed ratio	1.0114x
Format probe	15 / 15
4k net result	0.74x
16k net result	1.20x
32k net result	1.35x
32k KV per sequence	2048 MiB -> 1158 MiB
32k KV reduction	43.5%
Estimated 80 GiB cache capacity	12 -> 22 sessions
Same-memory capacity multiplier	1.83x

Operating Notes

The practical break-even region is around 13k to 16k context in the paired measurements. Short-context service should not use this candidate by default. The 4k result was negative, while 16k and 32k were positive.

The measured decode path was slower per user, while cache capacity increased enough to create positive net long-context economics at 16k and 32k. This is a capacity-oriented result, not a per-user decode-speed result.

Current KV and serving figures are based on the evaluated cache path. Paged-cache serving-engine behavior should be checked separately before deployment claims.

Important Limitations

This repository does not claim:

broad model-capability improvement over the base model;
better per-user decode speed than the base model;
production readiness;
short-context throughput improvement;
full safety, reasoning, coding, multilingual, tool-use, or multimodal coverage;
disclosure of the proprietary armoring procedure.

Quality ratios from narrow local measurements should not be used to rank armoring variants at the 1 percentage-point level. A larger document-level evaluation should report per-document distributions and the number of documents exceeding the original model.

License

The base model Qwen/Qwen3.6-27B is listed on Hugging Face with the Apache-2.0 license. This card follows the same license metadata while retaining VIDRAFT-specific implementation details privately.

Citation

VIDRAFT / FINAL-Bench. "Armoring-Qwen3.6-27B." Hugging Face Model Repository, 2026.

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for FINAL-Bench/Armoring-Qwen3.6-27B

Base model

Qwen/Qwen3.6-27B

Finetuned

(349)

this model

Collection including FINAL-Bench/Armoring-Qwen3.6-27B

Armoring Models

Collection

4 items • Updated about 16 hours ago • 29