Instructions to use FINAL-Bench/Armoring-Qwen3.5-2B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use FINAL-Bench/Armoring-Qwen3.5-2B with Transformers:

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

pipe = pipeline("text-generation", model="FINAL-Bench/Armoring-Qwen3.5-2B")

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("FINAL-Bench/Armoring-Qwen3.5-2B", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use FINAL-Bench/Armoring-Qwen3.5-2B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "FINAL-Bench/Armoring-Qwen3.5-2B"
# 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-Qwen3.5-2B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/FINAL-Bench/Armoring-Qwen3.5-2B

SGLang

How to use FINAL-Bench/Armoring-Qwen3.5-2B 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-Qwen3.5-2B" \
    --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-Qwen3.5-2B",
		"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-Qwen3.5-2B" \
        --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-Qwen3.5-2B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use FINAL-Bench/Armoring-Qwen3.5-2B with Docker Model Runner:
```
docker model run hf.co/FINAL-Bench/Armoring-Qwen3.5-2B
```

VIDRAFT Armoring

Armoring-Qwen3.5-2B

This repository contains a model-card report for a VIDRAFT Attention Armoring study on Qwen/Qwen3.5-2B.

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 small-model armoring prototype in the current series. Its role is to show that the method can be applied across all eligible attention blocks in a compact Qwen3.5-class model while preserving basic output behavior.

What changed	Why it matters
All eligible attention blocks were armored	A compact model gives a fast, low-cost validation target for architecture behavior.
Document-level perplexity ratio measured 1.0222x over 24 held-out documents	Language loss is small but clearly non-zero at this model size.
EOS probe passed 5 / 5	Basic termination behavior was preserved in the local probe.
Format probe passed 15 / 15	The candidate preserved the expected local response-format behavior.
8k KV per sequence moved from 96.0 MiB to 73.4 MiB	The measured cache footprint fell by 23.5% in this serving-style check.

These figures should be read as an engineering validation result, not as a general intelligence claim.

Measurement Protocol

Perplexity is reported over 24 held-out documents of 4,096 tokens each, and the per-document ratio distribution is published alongside the aggregate. A narrow probe over a handful of sentences is not sufficient to characterize language loss at this scale: on this same candidate, a two-document sample reports 1.0357x where the 24-document measurement reports 1.0222x. Aggregate-only figures from small samples should not be compared across armoring variants.

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 small-model check, the armored candidate preserved local format and termination probes while reducing the measured KV footprint in the tested serving-style setup.

Public Transformation Summary

Layer of the system	Before	After Armoring
Base model family	Qwen3.5-2B-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 with context and sessions	KV-cache footprint is reduced in the tested 8k setting
Operational value	Small model used mainly for fast validation	Low-cost prototype for behavior and cache checks
Release posture	Standard model-card summary	Public-safe report; implementation details retained privately

Measured Summary

Measurement	Result
Document PPL ratio vs. reference (24 documents)	1.0222x
Per-document ratio, minimum	0.9501x
Per-document ratio, median	1.0302x
Per-document ratio, maximum	1.0455x
Documents scoring better than reference	3 / 24
EOS probe	5 / 5
Format probe	15 / 15
8k KV per sequence	96.0 MiB -> 73.4 MiB
8k KV reduction	23.5%
Estimated 80 GiB cache capacity	773 -> 1,010 sessions
Same-memory capacity multiplier	1.31x
8k decode throughput, batch 8 to 64	0.71x - 0.73x

KV-cache accounting includes the armored attention path's own resident buffers, not only the standard per-layer cache tensors.

Operating Notes

Attention Armoring is a long-context efficiency intervention. In the measured family-level economics, short-context service can lose net throughput, while benefits emerge around long-context workloads. Treat this 2B model as a prototype and validation card, not as proof that every short-context deployment benefits.

The current 8k serving-style check showed higher capacity but lower per-user speed. Net GPU-level throughput was measured below the original in this short-context setup. This is why the recommended application area is long-context serving, not small-context latency-sensitive use.

Behavior Notes

The candidate passed local format and EOS probes. In difficult technical multiple-choice prompts, the model family may prefer explanation over forced single-letter output. That should be documented as a response-style characteristic rather than hidden as a failure.

Small benchmark probes are treated as directional only. They are not used here to claim a model-ranking win.

Important Limitations

This repository does not claim:

broad model-capability improvement over the base model;
production readiness;
full safety, reasoning, coding, multilingual, tool-use, or multimodal coverage;
short-context throughput improvement;
long-context retrieval parity with the base model, which is measured separately and is not reported here;
disclosure of the proprietary armoring procedure.

License

The base model Qwen/Qwen3.5-2B 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-Qwen3.5-2B." Hugging Face Model Repository, 2026.

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

Model tree for FINAL-Bench/Armoring-Qwen3.5-2B

Base model

Qwen/Qwen3.5-2B-Base

Finetuned

Qwen/Qwen3.5-2B

Finetuned

(314)

this model

Collection including FINAL-Bench/Armoring-Qwen3.5-2B

Armoring Models

Collection

4 items • Updated about 17 hours ago • 29