Qwen3.5 9B

Deploy Qwen3.5 9B on a dedicated GPU from $0.20 per hour, billed per minute, on 1× RTX A4000. A compact 9B native vision-language model with 262K context, 201-language coverage and Qwen's hybrid linear/full-attention architecture.

Multimodal9B paramsApache 2.0Deploys in ~5 minSecure Cloud

Qwen/Qwen3.5-9B

HexGrid Cloud price

$0.20/hr

1× RTX A4000

Context

262K

1.01M extended

Parameters

dense

Languages

201

languages & dialects

Running Qwen3.5 9B on HexGrid Cloud

What you get when you deploy with us, beyond the hourly rate.

Single tenant by default

The model runs on a GPU that is yours for the life of the instance. No shared endpoint, no queue behind other tenants.

Per-minute billing

You pay for the GPU, not the token. Stop the instance and billing stops with it.

OpenAI-compatible endpoint

Point an existing SDK at your instance by changing the base URL. No rewrite required.

Your weights, your data

Prompts and outputs stay inside your instance. Nothing is logged, sampled or used for training.

Qwen3.5 9B GPU sizing and cost

What the model needs at each precision, and the cheapest HexGrid Cloud configuration that holds it.

Qwen3.5 9B vRAM requirement and hourly cost by precision
Precision	vRAM needed	Runs on	Price
BF16 Recommended	22 GB	1× RTX 4090	$0.45/hr
FP8	11 GB	1× RTX A4000	$0.20/hr
INT4	6 GB	1× RTX A4000	$0.20/hr

Populate from a reproducible HexGrid benchmark. Qwen's own documentation does not supply one GPU-independent throughput figure.

About Qwen3.5 9B

What Qwen3.5 9B is built for, and where it falls short.

Qwen3.5-9B brings the Qwen3.5 native vision-language architecture to a substantially smaller dense model. It supports text, image and video input with text generation.

Its 32-layer backbone repeats three Gated DeltaNet linear-attention blocks followed by one gated full-attention block, giving it the same broad architectural approach as the larger 27B checkpoint.

The model retains a native 262K context window, documented extension to around 1.01M tokens and Qwen3.5's support for 201 languages and dialects.

What people run it for

Private multimodal assistants

Deploy an image-aware and multilingual assistant without moving to a 20B-70B model.

Visual document workflows

Analyze images, forms, screenshots and other visual content alongside long text.

Tool-using applications

Use structured calls for retrieval, automation and application actions.

Multilingual chat

Serve users across Qwen's stated 201-language-and-dialect coverage.

Strengths

9B dense size makes private deployment much more practical than 27B-class models
Native text, image and video understanding
262K native context at a relatively compact parameter scale
Strong published GPQA Diamond result of 81.7
201-language-and-dialect coverage
Apache 2.0 licence

Limitations

Lower coding performance than the larger Qwen3.5 27B model on LiveCodeBench
Million-token extended context still requires substantial memory regardless of weight size
Image and video inputs add preprocessing cost
Dense 9B weights remain substantially heavier than 3B-4B edge models
Some agent benchmarks depend heavily on external scaffolding and evaluation settings

Quickstart

Serve with vLLM or SGLang and call the OpenAI-compatible chat endpoint.

pythonPOST /v1/chat/completions

from openai import OpenAI

client = OpenAI(
    base_url="https://<your-instance>.hexgrid.cloud/v1",
    api_key="<your-key>",
)

response = client.chat.completions.create(
    model="Qwen/Qwen3.5-9B",
    messages=[
        {"role": "user", "content": "Explain semantic caching in a production RAG system."}
    ],
)

print(response.choices[0].message.content)

The official card demonstrates OpenAI-compatible calls for text, image and video inputs.

Qwen3.5 9B reported benchmarks

Scores published by the model authors. A starting point, not a guarantee on your own data.

MMLU-Pro82.5
GPQA Diamond81.7
IFEval91.5
LiveCodeBench v665.6
BFCL-V466.1
MMMU78.4

Qwen3.5 9B specifications

Architecture and serving details for Qwen3.5 9B.

PublisherQwenParameters9BLicenceApache 2.0ReleasedMarch 2026

Architecture

Type: Dense causal LM + vision encoder
Parameters: 9B
Layers: 32
Hidden dimension: 4096
Layout: 8 × (3 Gated DeltaNet + 1 Gated Attention)
Full-attention heads: 16 Q / 4 KV
MTP: Trained with multi-step prediction

Context

Native context: 262,144 tokens
Extended context: Up to 1,010,000 tokens
Input: Text, image, video
Output: Text
Languages: 201 languages and dialects

Serving

Engines: Transformers, vLLM, SGLang, KTransformers
API: OpenAI-compatible
Reasoning parser: qwen3
Tool parser: qwen3_coder

Qwen3.5 9B frequently asked questions

Common questions about deploying Qwen3.5 9B on HexGrid Cloud.

Does Qwen3.5-9B support vision?

Yes. The official checkpoint is a causal language model with a vision encoder and supports image and video inputs.

Does the 9B model keep the large context of Qwen3.5-27B?

Yes. Qwen lists 262,144 tokens natively and extension to 1,010,000 tokens.

How does it differ architecturally from Qwen3.5-27B?

The 9B model has 32 layers and a 4096-wide hidden state instead of 64 layers and 5120 hidden dimensions, while retaining the same repeating three-linear-attention-to-one-full-attention layout.

Can I use it commercially?

Yes. The checkpoint is released under Apache 2.0.

Deploy Qwen3.5 9B today

$0.20 per hour on 1× RTX A4000, billed per minute, never shared.

Deploy Qwen3.5 9B See all models