Qwen3.6 27B

Deploy Qwen3.6 27B on a dedicated GPU from $0.45 per hour, billed per minute, on 1× RTX 4090. A dense 27B multimodal Qwen model focused on agentic coding, reasoning and long-context work, with 262K native context and optional extension to roughly one million tokens.

Multimodal27B paramsApache 2.0Deploys in ~5 minSecure Cloud

Qwen/Qwen3.6-27B

HexGrid Cloud price

$0.45/hr

1× RTX 4090

Context

262K

1.01M extended

Parameters

27B

dense

Input

Text + Vision

image / video

Running Qwen3.6 27B 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.

Qwen3.6 27B GPU sizing and cost

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

Qwen3.6 27B vRAM requirement and hourly cost by precision
Precision	vRAM needed	Runs on	Price
BF16 Recommended	65 GB	5× RTX A4000	$1.00/hr
FP8	33 GB	1× RTX A6000	$0.55/hr
INT4	17 GB	1× RTX 4090	$0.45/hr

Qwen does not provide a single hardware-independent tok/s value. Measure on the precise HexGrid accelerator, precision, tensor-parallel layout and prompt/output distribution.

About Qwen3.6 27B

What Qwen3.6 27B is built for, and where it falls short.

Qwen3.6-27B is a dense multimodal causal language model with a vision encoder, released as the 27B open-weight member of the Qwen3.6 family.

The language model uses 64 layers arranged as repeating blocks of three Gated DeltaNet linear-attention layers followed by one gated full-attention layer. This hybrid layout is designed to make long-context computation more efficient while periodically restoring full global attention.

Qwen3.6 retains multimodal thinking and non-thinking operation and adds thinking preservation, allowing reasoning context from previous turns to be retained for iterative agent workflows.

What people run it for

Coding agents

Repository-scale software tasks, shell workflows, code editing and iterative agent development.

Multimodal agents

Agents that combine screenshots, visual material and text with structured tool calls.

Long-context analysis

Large documents, codebases and accumulated multi-turn context extending beyond conventional 128K models.

Visual reasoning

Reasoning over diagrams, screenshots, documents and sampled video frames.

Strengths

Strong agentic coding results, including 77.2 on Qwen's SWE-bench Verified evaluation
Dense architecture avoids MoE routing complexity
262K native context with documented extension to roughly 1.01M tokens
Image and video understanding alongside text
Thinking preservation for multi-turn agentic tasks
MTP training for speculative multi-token generation
Apache 2.0 licensing

Limitations

27B dense inference activates the full language model on every token
Very long 262K-1M contexts require large KV-cache allocations
Qwen's SWE-bench scores use an internal agent scaffold and should be compared only with models evaluated under the documented setup
Video processing adds media preprocessing and context overhead
Extended million-token operation may require special serving configuration and additional memory

Quickstart

Serve through vLLM or SGLang and access the model using the OpenAI-compatible chat-completions API.

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/Qwen3.6-27B",
    messages=[
        {"role": "user", "content": "Inspect this service design and suggest a fault-tolerant architecture."}
    ],
    max_tokens=4096,
)

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

Qwen documents OpenAI-compatible serving with both vLLM and SGLang. Tool use uses the qwen3_coder tool-call parser.

Qwen3.6 27B reported benchmarks

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

SWE-bench Verified77.2
SWE-bench Pro53.5
Terminal-Bench 2.059.3
MMLU-Pro86.2
GPQA Diamond87.8
LiveCodeBench v683.9
MMMU82.9

Qwen3.6 27B specifications

Architecture and serving details for Qwen3.6 27B.

PublisherQwenParameters27BLicenceApache 2.0ReleasedApril 2026

Architecture

Type: Dense causal LM + vision encoder
Parameters: 27B
Layers: 64
Hidden size: 5120
Layout: 16 × (3 Gated DeltaNet + 1 Gated Attention)
Full-attention heads: 24 Q / 4 KV
MTP: Trained with multi-step prediction

Context and modalities

Native context: 262,144 tokens
Extended context: Up to 1,010,000 tokens
Input: Text, image, video
Output: Text

Serving

Engines: Transformers, vLLM, SGLang, KTransformers
API: OpenAI-compatible chat completions
Reasoning parser: qwen3
Tool-call parser: qwen3_coder
Thinking preservation: Supported

Qwen3.6 27B frequently asked questions

Common questions about deploying Qwen3.6 27B on HexGrid Cloud.

Is Qwen3.6-27B a mixture-of-experts model?

No. Qwen explicitly describes Qwen3.6-27B as a dense 27-billion-parameter model.

Does it support images and video?

Yes. It is a multimodal causal language model with a vision encoder and Qwen documents both image and video inputs.

What is its maximum context?

The native context is 262,144 tokens. Qwen documents extension up to approximately 1,010,000 tokens.

What is thinking preservation?

It lets an application retain reasoning content from previous turns rather than discarding it, which Qwen recommends for iterative agentic workflows.

Deploy Qwen3.6 27B today

$0.45 per hour on 1× RTX 4090, billed per minute, never shared.

Deploy Qwen3.6 27B See all models