Roboflow
Qwen

Qwen: Qwen3.6 35B A3B

Qwen3.6 35B A3B Overview

Qwen3.6-35B-A3B is a sparse Mixture-of-Experts (MoE) multimodal language model developed by the Qwen team at Alibaba Group. It carries 35 billion total parameters but activates only approximately 3 billion per forward pass via a learned routing mechanism, giving it the representational capacity of a large dense model at a fraction of the inference compute. The model is natively multimodal, processing images, documents, and video alongside text as a core architectural capability rather than an add-on. It supports a native context window of 262,144 tokens, extensible up to 1,010,000 tokens via YaRN. A key design feature is the unified thinking/non-thinking mode framework: users can switch between deliberate chain-of-thought reasoning and fast direct responses within a single model, and a "thinking preservation" option retains reasoning context across multi-turn agentic workflows to reduce redundant computation.

The model is specifically optimized for agentic coding tasks, including repository-level reasoning, frontend workflow generation, multi-step tool use, and MCP (Model Context Protocol) integration. On SWE-bench Verified it scores 73.4%, on Terminal-Bench 2.0 it scores 51.5%, and on MCPMark it scores 37.0%. For vision-language tasks it achieves 92.0 on RefCOCO, 89.9 on OmniDocBench 1.5, and 83.7 on VideoMMMU. The model also supports Multi-Token Prediction (MTP) for speculative decoding. All Qwen3.6 open-weight models are released under the Apache 2.0 license.

Qwen3.6 35B A3B Interactive Demo

Model settings

Thinking level

Max output tokens

Default 65,536 · max 65,536

Sign in to adjust thinking and output length per run.

Results appear here. Add an image or pick an example to run Qwen3.6 35B A3B.

Qwen3.6 35B A3B Details & Performance

Details

Resources

GitHub

Vision Tasks

CaptioningChart Question AnsweringClassificationDocument Question AnsweringImage TaggingMulti-Label ClassificationOCRObject DetectionPhrase GroundingVideo ClassificationVision LanguageVisual Question Answering

Features

Foundation VisionLLMs with Vision CapabilitiesMultimodal Vision

Usage

Past 30 Days

Performance

Avg. Latency

Qwen3.6 35B A3B Vision Evals

Vision Evals is Roboflow's ground-truth benchmark: every model runs the same real-world samples across six vision tasks, and answers are scored against ground truth.

Evals updated September 5, 2026Pricing updated September 21, 2026

Overall score#21 of 53
71.9%
Avg cost / sample#19 of 53
$0.0030
Avg speed / sample#43 of 53
27.10s
Avg tokens / sample
4.6K

Strengths and weaknesses

Qwen3.6 35B A3B averages 71.9% across the six Vision Evals tasks, ranking #21 of 53 models overall.

Its weakest relative showing is OCR, ranking #34 of 53 at 87.7%.

At $0.0030 per sample it is the 19th cheapest of the 53 benchmarked models, and its average inference time of 27.1s per sample makes it the 43rd fastest.

Performance profile

Field medianQwen3.6 35B A3B

Field medians: Object Detection 53.9%, Counting 56.8%, Identification 84.4%, OCR 88.7%, Data Extraction 84.5%, Reasoning 54.1%.

Results by task

TaskScoreField (0 to 100)RankCost / sampleSpeed
Object Detection
57.0%
±1.3, Mean of 3 runs, range 56.1 to 58.7
#21 of 53$0.001210.26s
Counting
65.3%
±2.7, Mean of 3 runs, range 62.2 to 67.6
#22 of 53$0.005650.42s
Identification
82.3%
±6.3, Mean of 3 runs, range 75.0 to 87.5
#32 of 53$0.003325.94s
OCR
87.7%
±0.0, Mean of 3 runs, range 87.6 to 87.7
#34 of 53$0.002119.73s
Data Extraction
84.5%
±1.0, Mean of 3 runs, range 83.5 to 85.6
#25 of 53$0.001714.71s
Reasoning
54.8%
±1.3, Mean of 3 runs, range 53.0 to 55.6
#26 of 53$0.005853.56s

Price vs. performance

Score vs. cost

Overall benchmark score against estimated cost per sample, on a log scale. Upper-left is the sweet spot: high quality at low cost.

52 models on the current benchmark · scores and efficiency pooled across all six tasks at low effort · Qwen3.6 35B-A3B highlighted

Qwen3.6 35B A3B scores are the mean of 3 runs per task at low effort; this model has no separate high-effort tier because its chat template only switches thinking on or off · Methodology

View all Vision Evals →

Self-hosted benchmarks

2 quantizations of Qwen3.6 35B-A3B, served with vLLM from the published weights and scored on the same tasks as the hosted models. Expand a row for the GPUs it was measured on.

Self-hosted rows run with thinking on, the same setting as the hosted frontier models. Rows marked with a run count are the mean of three runs per task under the benchmark protocol; the rest are single runs awaiting their re-run. Some hosted open-weight rows still run without thinking, so a self-hosted quant can score above its own hosted API. Quantization still costs a little precision, and scores vary between runs.

QuantizationWeightsOverall scoreDetection mAP50Single-stream tok/s
FP83 runsRanked
37.5 GB
71.9%
57.0%199.2
AWQ INT4
25.0 GB
71.3%
56.4%212.2

Qwen3.6 35B A3B Pricing

Qwen3.6 35B A3B costs $0.150 per 1M input tokens and $1.00 per 1M output tokens.

Input$0.150 / 1M tokens
Output$1.00 / 1M tokens
Cached input$0.050 / 1M tokens

Pricing updated Sep 21, 2026

Alternatives to Qwen3.6 35B A3B

Other models worth comparing for similar use cases.

Qwen
Qwen3.8 27B
Qwen3.8-27B is a dense vision-language model of roughly 27.78 billion parameters from Alibaba's Qwen team, built on the architectural foundation established by the Qwen3.5 generation. It takes interleaved text, image, and video input through an integrated vision encoder and produces text, covering document and chart understanding, visual question answering, front-end code generation from visual references, and agentic computer-use workflows. The decoder is a hybrid stack of 64 layers that repeats a pattern of three Gated DeltaNet linear-attention blocks, each followed by a feed-forward block, then one grouped-query full-attention block, so only a quarter of the layers maintain a conventional key-value cache. Native context length is 262,144 tokens.Post-training targets long-horizon agentic work, and the model exposes per-request thinking control that lets callers switch reasoning traces on or off and choose a reasoning effort level. Multi-token prediction weights are trained alongside the main network for speculative decoding. Qwen reports improvements over Qwen3.6-27B on agentic and multimodal evaluations including Terminal-Bench 2.1, OSWorld-Verified, and SWE-bench Multimodal, with additional results on MathVision, CharXiv, and an internal Vision2Web suite.
Qwen
Qwen3.6 27B
Qwen3.6-27B is a dense 27-billion-parameter multimodal language model developed by Alibaba's Qwen team and released on April 22, 2026. It combines a causal language model with an integrated vision encoder, supporting text, image, and video inputs natively. The architecture employs a hybrid attention design that interleaves Gated DeltaNet linear attention blocks with standard Gated Attention layers across 64 transformer layers with a hidden dimension of 5,120. Unlike Mixture-of-Experts variants in the Qwen3.6 family, all 27 billion parameters are active on every inference pass, simplifying deployment and quantization. The model supports a native context window of 262,144 tokens, extensible to approximately 1,010,000 tokens via YaRN scaling. It is released under the Apache 2.0 license with open weights available on Hugging Face and ModelScope.The model introduces two notable capabilities relative to prior Qwen releases: enhanced agentic coding support covering frontend workflows and repository-level reasoning, and a Thinking Preservation mechanism that retains chain-of-thought reasoning context across multi-turn conversation history to reduce redundant token generation in iterative agent sessions. It supports both a thinking mode for multi-step reasoning and a non-thinking mode for faster responses within a single model. On coding benchmarks, Qwen reports scores of 77.2 on SWE-bench Verified, 59.3 on Terminal-Bench 2.0, and 48.2 on SkillsBench. Vision capabilities include chart understanding (CharXiv RQ: 78.4), OCR (CC-OCR: 81.2), and video understanding (VideoMME with subtitles: 87.7).
Google
Gemma 4 26B A4B
Gemma 4 26B A4B is the Mixture-of-Experts variant in Google's Gemma 4 family, with 25.2B total parameters but only 3.8B active per token. Built from the same Gemini 3 research as the 31B dense sibling and released as open weights under the Apache 2.0 license, it supports a 256K token context window with text and image input and configurable thinking mode. The "A4B" in the name refers to its approximately 4B active parameters. The MoE design makes it significantly faster at inference than the dense 31B, running nearly as fast as a 4B-parameter model while delivering roughly 97% of the dense model's quality.For vision tasks, the 26B A4B shares the same multimodal capabilities as the 31B image understanding with variable aspect ratios and resolutions, and structured bounding box output for UI element detection. The tradeoff versus the 31B dense model is a small quality reduction in exchange for much faster inference and lower hardware requirements, fitting in 18GB of VRAM at 4-bit quantization. It ranked #6 among open models on the Arena AI text leaderboard at launch.
Google
Gemma 4 31B
Gemma 4 31B is the largest dense model in Google's Gemma 4 family, built from the same research as Gemini 3 and released as open weights under the Apache 2.0 license. It supports a 256K token context window with text and image input, configurable thinking mode for step-by-step reasoning, and multilingual support across 140+ languages. The unquantized model fits on a single 80GB GPU.For vision tasks, Gemma 4 31B supports image understanding with variable aspect ratios and resolutions, and can output structured bounding boxes for UI element detection, making it useful for document parsing and UI understanding. Compared to Gemma 3, it delivers stronger reasoning and multimodal performance. It is part of a four-size family alongside the 26B A4B MoE variant and two on-device models (E2B, E4B), with the 31B dense variant optimized for output quality and fine-tuning over inference speed.

Deploy Qwen3.6 35B A3B with an API

Qwen3.6 35B A3B runs as a hosted REST endpoint through Roboflow Workflows. Pick a task, then hand the prompt to your coding agent or copy the code. Deploying the workflow into a free Roboflow workspace replaces the your-workspace and YOUR_API_KEY placeholders with your own.

Connect your agent to Roboflow (once)

Add the Roboflow MCP server

claude mcp add --transport http roboflow https://mcp.roboflow.com/mcp

Run /mcp and authorize Roboflow in your browser when the OAuth flow opens.

Start a new Claude Code session so the MCP loads, then paste the prompt below (it works the same in any agent).

Deploy this workflow to your Roboflow workspace to use it.

Integrate the Roboflow "Qwen3.6 35B A3B" workflow into my app.

- Endpoint: POST https://serverless.roboflow.com/<your-workspace>/workflows/qwen3-6-35b-a3b-captioning
- Auth: send my Roboflow API key as `api_key` in the request body, read from the ROBOFLOW_API_KEY env var (never hardcode).
- Body: { "api_key": ..., "inputs": { `image`: { type: "url" | "base64", value } } }.
- Billing: this workflow needs no provider API key — inference runs on my Roboflow credits. A BYO provider key can be added to the model step in the Roboflow workflow editor later.

With the Roboflow MCP connected, call `workflows_get` on "qwen3-6-35b-a3b-captioning" to read the exact input schema (the source of truth), then `workflows_run` on a sample image to confirm the output shape before writing code (the MCP is authenticated, so this needs no key). Without the MCP, use the contract above.

Before running the app, set up these keys so it does not error at runtime:
- `ROBOFLOW_API_KEY` (sent as `api_key`) from https://app.roboflow.com/settings/api
Create a .gitignore'd .env with these variables, using placeholder values for any I haven't given you. Then pause and tell me directly, in your reply: the full path to the .env file, exactly which keys I need to paste in, and the link to get each one. Wait for me to confirm I've added them before you run anything. Do not run the app until I confirm.

Then add the integration to my codebase: match my project's language, framework, and conventions; read every key from environment variables (never hardcode); add basic error handling; and include a small runnable example. If you can't tell what language my project uses, ask me.
Installpip install inference-sdk

Deploy this workflow to your Roboflow workspace to use it.

# Inference runs on your Roboflow credits — no provider API key needed. To bill your own provider account instead, add an api_key to the model step in the Roboflow workflow editor.
# 1. Import the library
from inference_sdk import InferenceHTTPClient

# 2. Connect to your workflow
client = InferenceHTTPClient(
  api_url="https://serverless.roboflow.com",
  api_key="YOUR_API_KEY"
)

# 3. Run your workflow on an image
result = client.run_workflow(
  workspace_name="your-workspace",
  workflow_id="qwen3-6-35b-a3b-captioning",
  images={
    "image": "YOUR_IMAGE.jpg"  # Path to your image file
  },
  use_cache=True  # cache workflow definition for 15 minutes
)

# 4. Get your results
print(result)

Deploy this workflow to your Roboflow workspace to use it.

// Inference runs on your Roboflow credits — no provider API key needed. To bill your own provider account instead, add an api_key to the model step in the Roboflow workflow editor.
const response = await fetch('https://serverless.roboflow.com/your-workspace/workflows/qwen3-6-35b-a3b-captioning', {
  method: 'POST',
  headers: {
    'Content-Type': 'application/json'
  },
  body: JSON.stringify({
    api_key: 'YOUR_API_KEY',
    inputs: {
      "image": {"type": "url", "value": "IMAGE_URL"}
    }
  })
});

const result = await response.json();
console.log(result);

Deploy this workflow to your Roboflow workspace to use it.

# Inference runs on your Roboflow credits — no provider API key needed. To bill your own provider account instead, add an api_key to the model step in the Roboflow workflow editor.
curl --location 'https://serverless.roboflow.com/your-workspace/workflows/qwen3-6-35b-a3b-captioning' \
--header 'Content-Type: application/json' \
--data '{
  "api_key": "YOUR_API_KEY",
  "inputs": {
    "image": {"type": "url", "value": "IMAGE_URL"}
  }
}'

Qwen3.6 35B A3B License

Apache-2.0 · Permissive license

Qwen3.6 35B A3B is released under Apache-2.0, a permissive license. The Qwen3.6 35B A3B license lets you run, fine-tune, and redistribute the model in commercial products with no obligation to open-source related code changes, so no separate commercial license is required.

Commercial use
Permitted. Because Apache-2.0 is permissive, Qwen3.6 35B A3B can ship inside paid products and internal systems with no commercial license and no revenue threshold.
Modification
Permitted. Fine-tuning, quantizing, and distilling are all allowed, and your code changes can stay closed. Files you change must be marked as changed.
Redistribution
Permitted with attribution. Ship the Apache-2.0 license text and any NOTICE file alongside the weights or derived code.

Apache-2.0 grants an express patent license that terminates if you bring a patent claim over the work, and it disclaims warranties. Validate Qwen3.6 35B A3B on your own data before you depend on it in production.

Read the full Apache 2.0 license ↗

Do I need a commercial license for Qwen3.6 35B A3B?

This is the straightforward case: a permissive license is the best technical solution and you are free to deploy Qwen3.6 35B A3B commercially without open-sourcing your own code.

Do not hesitate to reach out with questions for your commercial project — our team will help you start solving business problems on the first call. See Roboflow commercial licensing for the models included in each plan.

Talk to sales

This model is released under the Apache License 2.0, a permissive open-source license that allows commercial use, modification, distribution, and patent use.

Yes. Under the terms of the Apache 2.0 license, you can freely use this model for commercial purposes, including in proprietary products. You must retain the copyright notice and disclaimers when redistributing.

License information is provided as a guide and is not legal advice.

Frequently Asked Questions About Qwen3.6 35B A3B Vision

Yes. Qwen3.6 35B A3B accepts image input and handles OCR, data extraction, object counting, identification, visual reasoning, and object detection. On Roboflow's Vision Evals its strongest task is Object Detection at 57% (#21 of 53 at low effort). You can test it on your own image in the demo above.

Yes. its transcriptions match the ground truth 87.7% on average (#34 of 53 at low effort) on Vision Evals OCR. Pulling specific fields out of documents (data extraction) scores 84.5%.

It's serviceable. On Vision Evals, Qwen3.6 35B A3B scores 57% mAP@50 on object detection (#21 of 53 at low effort) and 65.3% judge-graded accuracy on object counting.

On our benchmark's task mix, Qwen3.6 35B A3B averages $0.0030 per sample at $0.15 per 1M input and $1.00 per 1M output tokens (#19 of 53 on cost), with an average speed of 27.1s per sample across the benchmark. Actual cost depends on your images and prompts.

On the overall Vision Evals ranking, Qwen3.6 35B A3B sits #21 of 53 at 71.9%, just behind Qwen3.6 27B (73.6%) and just ahead of Qwen3.5 27B (70.8%). See the full side-by-side: Qwen3.6 35B A3B vs Qwen3.6 27B.