# Qwen3.5

> Alibaba's open-weight 397B-A17B multimodal agent model.

- Maker: Alibaba (Qwen)
- Line: Qwen (open-weight)
- Canonical page: https://ai-tldr.dev/models/qwen3-5/
- Updated: 2026-06-18

## In short

Qwen3.5 is the flagship open-weight model family released by Alibaba's Qwen team on 16 February 2026. It has a 256K-token context window and costs $0.54 per million input tokens and $3.40 per million output tokens. You can use it through DeepInfra and Alibaba Cloud Model Studio (DashScope).

## Facts

| Fact | Value | Source |
| --- | --- | --- |
| Context window | 256K | [1] |
| Input price | $0.54 / 1M tokens | [2] |
| Output price | $3.40 / 1M tokens | [2] |
| Licence | Apache-2.0 · open weights | [1] |
| Released | 16 February 2026 | [3] |
| Parameters | 397B total / 17B active (MoE, 512 experts) | [1] |

## Sources

1. [Qwen — Wikipedia](https://en.wikipedia.org/wiki/Qwen)
2. [Qwen3.5-397B-A17B API benchmarks & pricing (DeepInfra)](https://deepinfra.com/blog/qwen3-5-397b-a17b-api-benchmarks)
3. [Qwen3.5 — official announcement](https://qwen.ai/blog?id=qwen3.5)

## Overview

Qwen3.5 is the flagship open-weight model family released by Alibaba's Qwen team on 16 February 2026. The headline model, Qwen3.5-397B-A17B, is a sparse mixture-of-experts network with 397 billion total parameters that activates only about 17 billion per token, giving it the capability of a very large model at the inference cost of a much smaller one. It ships under the Apache 2.0 license with weights on Hugging Face, GitHub, and ModelScope.

Architecturally, Qwen3.5 moves away from a pure softmax-attention transformer. Roughly three quarters of its layers use Gated DeltaNet linear attention, with the remaining quarter using conventional grouped-query softmax attention and RoPE. This hybrid design is what lets Alibaba claim large throughput gains over the dense Qwen3-Max at long context lengths while matching its quality. It is also the first open-weight Qwen model with native vision and video input, folding the previous text-only Qwen3 and vision-focused Qwen3-VL lines into one early-fusion backbone.

The flagship was followed by smaller open-weight variants — Qwen3.5-122B-A10B, 35B-A3B, and 27B on 24 February 2026, then 9B, 4B, 2B, and 0.8B on 2 March 2026 — covering everything from data-center MoE to edge-sized dense models. A hosted, proprietary Qwen3.5-Plus endpoint extends the context window to 1M tokens. The series supports 201 languages and dialects (up from 119 in Qwen3) and runs in both reasoning and non-reasoning modes.

## Benchmarks

- GPQA Diamond: 88.4% ([source](https://deepinfra.com/blog/qwen3-5-397b-a17b-api-benchmarks))
- AIME 2026: 91.3% ([source](https://codersera.com/blog/qwen-3-5-complete-guide-2026/))
- LiveCodeBench v6: 83.6% ([source](https://codersera.com/blog/qwen-3-5-complete-guide-2026/))
- SWE-bench Verified: 76.4% ([source](https://deepinfra.com/blog/qwen3-5-397b-a17b-api-benchmarks))
- MMLU-Pro: 87.8% ([source](https://deepinfra.com/blog/qwen3-5-397b-a17b-api-benchmarks))
- IFBench: 76.5% ([source](https://www.datacamp.com/blog/qwen3-5))
- Tau2-Bench: 86.7% ([source](https://www.analyticsvidhya.com/blog/2026/02/qwen3-5-open-weight-qwen3-5-plus/))
- MMMU: 85% ([source](https://www.datacamp.com/blog/qwen3-5))
- MMMU-Pro: 79% ([source](https://www.datacamp.com/blog/qwen3-5))
- OmniDocBench v1.5: 90.8% ([source](https://www.datacamp.com/blog/qwen3-5))
- Video-MME: 87.5% ([source](https://www.datacamp.com/blog/qwen3-5))
- BrowseComp: 78.6% ([source](https://www.datacamp.com/blog/qwen3-5))
- Terminal-Bench 2.0: 52.5% ([source](https://www.datacamp.com/blog/qwen3-5))

## Pricing

- Input: $0.54 / 1M tokens per 1M tokens
- Output: $3.40 / 1M tokens per 1M tokens

Third-party hosted pricing for the open-weight Qwen3.5-397B-A17B (FP8) on DeepInfra; the weights themselves are free to self-host under Apache 2.0. Alibaba's hosted Qwen3.5-Plus endpoint is priced separately.

Source: https://deepinfra.com/blog/qwen3-5-397b-a17b-api-benchmarks

## Strengths

- Strong agentic and tool-use results: 86.7 on Tau2-Bench, 78.6 on BrowseComp, and 52.5 on Terminal-Bench 2.0
- Top-tier open-weight reasoning and coding — 88.4 GPQA Diamond, 91.3 AIME 2026, 83.6 LiveCodeBench v6, 76.4 SWE-bench Verified
- Native multimodality in an open model: text, images, video, and document understanding from one backbone (90.8 on OmniDocBench v1.5)
- Hybrid linear + softmax attention gives much higher long-context throughput than dense models at comparable quality
- Permissive Apache 2.0 license with a full ladder of sizes from 397B-A17B down to 0.8B for self-hosting
- Broad multilingual coverage across 201 languages and dialects

## Best for

- Self-hosted agentic systems that need open weights for tool use, browsing, and terminal/coding workflows
- Multimodal document, chart, and video understanding pipelines on private infrastructure
- Long-context retrieval and analysis up to 256K tokens (1M via the hosted Plus endpoint)
- On-device and edge deployments using the smaller 27B/9B/4B/2B/0.8B dense variants
- Cost-sensitive high-throughput inference where the MoE's 17B active footprint cuts serving cost
- Multilingual assistants and translation across 201 languages

## How to access

- [DeepInfra](https://deepinfra.com/blog/qwen3-5-397b-a17b-api-benchmarks) — model id `Qwen/Qwen3.5-397B-A17B`
- [Alibaba Cloud Model Studio (DashScope)](https://qwen.ai/blog?id=qwen3.5) — model id `qwen3.5-plus`

## Qwen (open-weight) versions

- [Qwen3.6](https://ai-tldr.dev/models/qwen3-6/) (April 2026)
- Qwen3.5 (16 February 2026) — this page
- [Qwen3 (2507 update)](https://ai-tldr.dev/models/qwen3-2507/) (July 2025)
- [Qwen3](https://ai-tldr.dev/models/qwen3/) (28 April 2025)
- [Qwen2.5](https://ai-tldr.dev/models/qwen2-5/) (September 2024)
- [Qwen2](https://ai-tldr.dev/models/qwen2/) (June 2024)

## FAQ

### Is Qwen3.5 open source?

The weights are open under the Apache 2.0 license and downloadable from Hugging Face, GitHub, and ModelScope, so you can self-host and use it commercially. The hosted Qwen3.5-Plus endpoint is a separate proprietary service.

### How big is Qwen3.5 and how is it different from earlier Qwen models?

The flagship Qwen3.5-397B-A17B has 397 billion total parameters but activates only about 17 billion per token via a 512-expert MoE. Unlike earlier Qwen models it uses a hybrid architecture — mostly Gated DeltaNet linear attention with some full softmax attention — and is the first open-weight Qwen with native vision and video, merging the old Qwen3 and Qwen3-VL lines.

### What context length does Qwen3.5 support?

The open-weight Qwen3.5-397B-A17B handles a native context window of roughly 256K tokens (around 262K). The hosted Qwen3.5-Plus endpoint extends this to 1 million tokens.

### What modalities does Qwen3.5 handle?

It is a native multimodal model that accepts text, images, and video, and performs strong document understanding (90.8 on OmniDocBench v1.5). Audio is handled by the separate Qwen-Omni line, not Qwen3.5.

## Links

- [Qwen3.5 — official announcement](https://qwen.ai/blog?id=qwen3.5)
- [Qwen3.5: Features, Access, and Benchmarks (DataCamp)](https://www.datacamp.com/blog/qwen3-5)
- [Qwen3.5-397B-A17B API benchmarks & pricing (DeepInfra)](https://deepinfra.com/blog/qwen3-5-397b-a17b-api-benchmarks)
- [Qwen 3.5 complete guide 2026 (Codersera)](https://codersera.com/blog/qwen-3-5-complete-guide-2026/)
- [Hands-on with Qwen3.5 open-weight & Plus (Analytics Vidhya)](https://www.analyticsvidhya.com/blog/2026/02/qwen3-5-open-weight-qwen3-5-plus/)
- [Qwen — Wikipedia](https://en.wikipedia.org/wiki/Qwen)

Source: AI/TLDR LLM registry, https://ai-tldr.dev/models/qwen3-5/