Back to all models

Qwen3.5 27B

Deploy Qwen3.5 27B on a dedicated GPU from $0.45 per hour, billed per minute, on 1× RTX 4090. A dense 27B native vision-language model combining hybrid linear/full attention, 201-language coverage and 262K native context for reasoning, coding and agent workloads.

Multimodal27B paramsApache 2.0Deploys in ~5 minSecure Cloud
Qwen/Qwen3.5-27B

HexGrid Cloud price

$0.45/hr

1× RTX 4090

Context

262K

1.01M extended

Parameters

27B

dense

Languages

201

languages & dialects

Running Qwen3.5 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.

Qwen3.5 27B GPU sizing and cost

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

Qwen3.5 27B vRAM requirement and hourly cost by precision
PrecisionvRAM neededRuns onPrice
BF16
Recommended		65 GB5× RTX A4000$1.00/hr
FP8
33 GB1× RTX A6000$0.55/hr
INT4
17 GB1× RTX 4090$0.45/hr

Official Qwen sources do not provide one standardized tok/s value applicable across GPUs and context lengths. Use HexGrid's own reproducible serving benchmark.

About Qwen3.5 27B

What Qwen3.5 27B is built for, and where it falls short.

Qwen3.5-27B is a dense native vision-language model for text, images and video. Its 64-layer language backbone combines Gated DeltaNet linear attention with periodic gated full attention.

The model has a native 262,144-token context window and Qwen documents extension to 1,010,000 tokens for ultra-long workloads.

Qwen3.5 expands the family's multilingual coverage to 201 languages and dialects and was post-trained for reasoning, coding, instruction following and agentic tool use.

What people run it for

Multimodal assistants

General assistants that work across natural-language prompts, screenshots, documents and video.

Coding and software agents

Repository analysis, coding, tool calls and terminal-oriented workflows.

Long-context RAG

Reason over large collections of retrieved material without aggressively shrinking the context.

Global applications

Multilingual systems benefiting from Qwen's stated support for 201 languages and dialects.

Strengths

  • Strong knowledge and reasoning results including 86.1 MMLU-Pro and 85.5 GPQA Diamond
  • Native image and video understanding
  • 262K native context with approximately one-million-token extension
  • 201-language-and-dialect coverage
  • Strong instruction following with 95.0 IFEval
  • Native OpenAI-compatible serving recipes for vLLM and SGLang

Limitations

  • Full 27B dense computation is used on each generated token
  • Million-token context extension is memory intensive
  • Multimodal serving requires additional vision processing compared with language-model-only serving
  • Agent benchmark values depend strongly on the agent scaffold and Qwen has published later re-evaluations under changed setups
  • Very large maximum-output configurations are not appropriate for every production workload

Quickstart

Serve with vLLM or SGLang and access it with an OpenAI-compatible client.

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/Qwen3.5-27B",
    messages=[
        {"role": "user", "content": "Review this architecture and identify scalability risks."}
    ],
)

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

Qwen's model card documents text, image and video requests through an OpenAI-compatible chat API.

Qwen3.5 27B reported benchmarks

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

  • MMLU-Pro86.1
  • IFEval95
  • GPQA Diamond85.5
  • SWE-bench Verified72.4
  • LiveCodeBench v680.7
  • BFCL-V468.5
  • MMMU82.3

Qwen3.5 27B specifications

Architecture and serving details for Qwen3.5 27B.

PublisherQwenParameters27BLicenceApache 2.0ReleasedFebruary 2026

Architecture

Type
Dense causal LM + vision encoder
Parameters
27B
Layers
64
Hidden dimension
5120
Layout
16 × (3 Gated DeltaNet + 1 Gated Attention)
Full-attention heads
24 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 27B frequently asked questions

Common questions about deploying Qwen3.5 27B on HexGrid Cloud.

Is Qwen3.5-27B text-only?

No. The official model is a native multimodal causal language model with a vision encoder and supports text, image and video input.

How long is the context window?

262,144 tokens natively, with Qwen documenting extension up to 1,010,000 tokens.

How many languages does it support?

Qwen states that the Qwen3.5 family expanded coverage to 201 languages and dialects.

Why can its SWE-bench number appear different in Qwen3.6 material?

Qwen later evaluated Qwen3.5-27B with an updated internal agent scaffold. This schema keeps the 72.4 value published on the Qwen3.5-27B model card rather than silently mixing evaluation setups.

Deploy Qwen3.5 27B today

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

Deploy Qwen3.5 27BSee all models