Qwen: every model, side by side

Qwen3.8 27B is an open-weight dense vision-language model from Qwen. It is suited for coding, professional workflows, research, multimodal interaction, and long-running agent tasks, with flexible thinking that can be enabled or disabled.

Qwen3.8 2.4T A95B is an open-weight sparse mixture-of-experts model from Qwen and the open-weight variant of Qwen3.8 Max, with 95 billion active parameters out of 2.4 trillion total. It is suited for coding, research, complex reasoning, and agentic workflows.

The top tier of Alibaba's Qwen3.8 series and the general-availability successor to Qwen3.8 Max Preview. A multimodal reasoning model for complex reasoning, visual understanding, coding and agent workflows.

The largest model in Alibaba's Qwen3.7 series, text in and text out. Built for agent-centric work: coding, office and productivity tasks, and long-horizon autonomous execution. Clear gains on coding and agent behaviour over earlier Qwen generations, with explicit prompt caching for repeated context.

The cost-effective tier of Alibaba's Qwen3.7 series, taking text and image. Carries the series' coding, tool use and productivity work, plus a large vision upgrade. The distinguishing trait is interactive multimodal agency: reading screens, driving GUIs, coding from a visual reference, navigating mobile apps.

Qwen3.5 Plus (April 2026) is a large-scale multimodal language model from Alibaba. It accepts text, image, and video input and produces text output, with a 1M token context window. This is an updated version of Qwen3.5 Plus with tiered pricing above 256K tokens.

Qwen3.6 Flash is a fast, efficient language model from Alibaba's Qwen 3.6 series. It supports text, image, and video input with a 1M token context window. Tiered pricing kicks in above 256K tokens. Prompt caching is supported, with both explicit cache read and cache creation pricing.

An open-weight multimodal model from Alibaba Cloud, 35B total with 3B active per token. Hybrid sparse mixture-of-experts mixing Gated DeltaNet linear attention with standard gated attention. 262K native context, extensible to 1M with YaRN. Text, image and video. Thinking mode built in. Apache 2.0.

A dense 27B from Alibaba's Qwen team, released April 2026. Text, image and video across a 262,144 token context. Built for agentic coding and reasoning, clearest on repository-level code comprehension and front-end work. Thinking mode built in and persistent. 201 languages, Apache 2.0.

Alibaba Cloud's proprietary Qwen3.6 Max, a sparse mixture of experts at roughly 1 trillion total parameters. For agentic coding, tool use and long-context reasoning over a 262K window, with a thinking mode that keeps traces across turns. Model Studio and Qwen Studio only, no open weights.

The preview of Qwen 3.6 Plus, on a hybrid architecture that improves efficiency and scaling over the 3.5 series, with more reliable agent behaviour. Strongest on agentic coding and front-end development. Free tier on OpenRouter, which collects prompt and completion data.

A multimodal foundation model in the Qwen3.5 family at 9B parameters. Unified vision-language design with early fusion of multimodal tokens: text, image and video in, text out, with reasoning built in.

The Qwen3.5 Series 35B-A3B is a native vision-language model designed with a hybrid architecture that integrates linear attention mechanisms and a sparse mixture-of-experts model, achieving higher inference efficiency. Its overall performance is comparable to that of the Qwen3.5-27B.

The Qwen3.5 27B native vision-language Dense model incorporates a linear attention mechanism, delivering fast response times while balancing inference speed and performance. Its overall capabilities are comparable to those of the Qwen3.5-122B-A10B.

The 122B-A10B vision-language model in the Qwen3.5 series, on a hybrid of linear attention and a sparse mixture of experts. Second only to Qwen3.5-397B-A17B overall, with text well ahead of Qwen3-235B-2507 and vision ahead of Qwen3-VL-235B.

The Flash tier of Qwen3.5, a native vision-language model on a hybrid of linear attention and a sparse mixture of experts. Built for fast responses, and a clear step up from the Qwen3 series on both text and multimodal tasks.

The Plus tier of Qwen3.5, a native vision-language model on a hybrid of linear attention and sparse mixture of experts for cheaper inference. Alibaba reports it level with leading models, and a clear step up from Qwen3 on text and multimodal work. Text, image and video in, with reasoning and tool use.

The 397B-A17B vision-language model in the Qwen3.5 series, on a hybrid of linear attention and sparse mixture of experts. Alibaba reports it level with leading models across language, logic, code, agent tasks, image and video understanding and GUI interaction. 201 languages and dialects.

The reasoning tier of Qwen3-Max, scaled on both model capacity and reinforcement learning compute. Heavy Mode iterates on an answer at test time, tools include search and a code interpreter, and it can switch between normal and compute-heavy reasoning mid-conversation.

An open-weight model for coding agents and local development. Sparse MoE with 80B total parameters and 3B activated per token, performing like models with 10 to 20x the active compute. Non-thinking mode only.

An image generation foundation model in the Qwen series that achieves significant advances in complex text rendering.

A fast Qwen text-to-image model optimized by PrunaAI for speed on Replicate.

The larger of Alibaba's two hosted Qwen3 Coder tiers, sold through Model Studio rather than released as open weights. 128K context, with OpenAI-compatible and Anthropic-compatible endpoints.

The cheap tier of Alibaba's hosted Qwen3 Coder line, aimed at autocomplete and quick edits where latency matters more than depth. 128K context, Model Studio only, no open weights.

The instruct member of Qwen3-Next, tuned for fast, stable answers with no visible chain of thought. It prioritises throughput and consistency on very long inputs and multi-turn dialogue, which is what makes it usable for retrieval-augmented generation and tool-calling agents.

The reasoning member of Qwen3-Next, emitting thinking traces by default in thinking-only mode. Built for maths proofs, code synthesis and debugging, logic and agent planning, and tuned to stay stable across long chains without drifting off task.

Qwen Plus 0728 with reasoning switched on. The same Qwen3-based 1 million token hybrid model, spending extra tokens to think before it answers.

Qwen Plus 0728, built on Qwen3: a hybrid reasoning model with a 1 million token context, balancing quality, speed and cost.

Alibaba's largest Qwen3 model, an update on the January 2025 release with better mathematics, coding, logic and science, and fewer hallucinations on open-ended questions. Over 100 languages, tuned for retrieval-augmented generation and tool calling. No dedicated thinking mode.

The thinking variant of Qwen3-30B-A3B, a 30B mixture-of-experts model that keeps its reasoning trace separate from the final answer. Longer output budgets than earlier 30B releases, with gains across logic, mathematics, science, coding and multilingual benchmarks.

A 30.5B mixture-of-experts model from Qwen with 3.3B active parameters per inference, running in non-thinking mode only. It answers directly rather than reasoning aloud, and beats the non-instruct variant on open-ended and subjective tasks while holding its coding and factual scores.

The thinking-only variant of Qwen3-235B-A22B, activating 22B of 235B parameters and holding 262,144 tokens of context. It always emits a reasoning trace and is built for long outputs, up to 81,920 tokens, on maths, science and long-form generation.

Qwen's 480B mixture-of-experts coding model, activating 35B parameters per pass from 8 of 160 experts. Built for agentic coding: function calling, tool use, and reasoning over a whole repository rather than a single file.

The July 2025 instruct refresh of Qwen3-235B-A22B, activating 22B of 235B parameters per pass. Native 262K context and no thinking mode: it answers directly. Gains over the base variant are largest in knowledge coverage, long-context reasoning, coding and multilingual maths.

A 0.6B dense model from the Qwen3 family, small enough to run on a laptop. Switches between thinking mode for hard tasks and non-thinking mode for chat. Trained on 36 trillion tokens across 119 languages, with tool use and multilingual support.

Qwen3 at 30.5B parameters with 3.3B activated. Reasoning, multilingual work and agent tasks, with a thinking/non-thinking mode switch. Up to 131K context with YaRN. Free tier on OpenRouter.

A 235B mixture-of-experts model from Alibaba's Qwen team, activating 22B parameters per forward pass. Switches between a thinking mode for hard tasks and a non-thinking mode for chat. Strong on reasoning, tool calling and over 100 languages. 32K context, extendable to 131K.

Qwen's 32B reasoning model. It thinks before answering, which buys it a wide margin over conventional instruction-tuned models on hard problems and puts it in range of much larger reasoning models like DeepSeek-R1 and o1-mini.