What Is Qwen3? The Qwen3 Family and Qwen3.7-Max Explained
Here is the first thing to clear up: Qwen3 is not one model. It is a generation. The Qwen3 line opened in April 2025 and has shipped four sub-releases since, ending (for now) with Qwen3.7-Max in May 2026. So when someone asks what Qwen3 is, the answer depends on whether they mean the architecture all these models share, the open-weight checkpoints they can download today, or the proprietary flagship sitting at the top of the stack. This breakdown covers all three, with the numbers grounded against Alibaba documentation, independent leaderboards, and provider pricing pages.
Qwen3 in one sentence: Qwen3 is Alibaba Cloud's third model generation, built on a hybrid Gated DeltaNet attention design, spanning free Apache 2.0 open-weight models you can run on a laptop up to the proprietary, agent-first Qwen3.7-Max (an agent is a model that acts and calls tools over many steps, not just one that chats) that places in the global top 5 for raw intelligence and leads several agentic-coding benchmarks.
What "Qwen3" Actually Refers To
Qwen is Alibaba Cloud's family of large language models, built by the team behind Tongyi Qianwen. The number in the name marks the generation. Qwen2 arrived in June 2024. Qwen3 followed in April 2025 and introduced the architecture that still defines the line today. Everything after it (Qwen3.5, Qwen3.6, and Qwen3.7-Max) is a point release inside that third generation, not a clean break to a new one. For a plain-language primer on what Qwen is as a platform, start with our overview.
Put governance around how your team uses AI. The AI Acceptable Use Policy: a deploy-ready template that sets the rules for AI use.
Your purchase helps keep our hubs free to read.
This matters because the naming trips people up. "Qwen3" can mean the original April 2025 release, or it can be shorthand for the whole third-generation stack. In practice, almost nobody is deploying the original Qwen3 checkpoint in mid-2026. They are running one of its descendants. The two that matter most are Qwen3.7-Max for hosted API work and Qwen3.6-35B-A3B for local deployment. Our guide to run Qwen locally walks through the setup on consumer hardware.
Reading the model names helps. The format for Mixture-of-Experts models (a design that activates only a fraction of its total parameters for each chunk of text) is [total]B-A[active]B. So Qwen3.6-35B-A3B carries 35 billion total parameters but activates only 3 billion per token (tokens are roughly word-pieces; a model's context window is how much of this text it can consider at once). The Max tier drops the parameter suffix because Alibaba does not publish exact counts for its proprietary flagships; it confirms only the rough scale.
The through-line across the generation is consistency of design rather than a single product. Each release reused the hybrid attention backbone, expanded modality or context, and pushed harder on agentic behavior. That is why a breakdown of Qwen3 has to be a breakdown of a family. You can explore each model and guide from the Qwen tools hub.
The Qwen3 Family, Release by Release
Four sub-generations sit inside Qwen3, and they alternate between open-weight community releases and proprietary frontier launches. The pattern is deliberate: Alibaba tends to ship the closed Max flagship first, then follow weeks or months later with open weights at the tier below. Understanding the order tells you which model to reach for and whether you can self-host it.
One detail worth holding onto: the open-weight ceiling and the proprietary ceiling are different models. The best you can download and self-host today is Qwen3.5-397B-A17B. The best you can call through an API is Qwen3.7-Max. If a project depends on owning the weights, you are working a generation behind the absolute frontier, and that gap is a real trade-off rather than a rounding error.
Qwen3.7-Max: The Current Flagship
Qwen3.7-Max launched on May 18, 2026, and was presented two days later at the Alibaba Cloud Summit in Hangzhou. It is the first Chinese model to credibly sit in the top tier for agentic AI. It is proprietary and API-only, with roughly 1 trillion total parameters and around 24 billion active per token under a sparse Mixture-of-Experts design. It is also, deliberately, a text-only model: there is no vision or audio input.
Its 1 million token context window is large enough to load most enterprise codebases into a single request without retrieval scaffolding. It is built agent-first, tuned for long-horizon coding, MCP (Model Context Protocol, a standard for letting models call external tools) tool orchestration, and tasks measured in hours rather than seconds. The price, which we cover further down, comes to a fraction of comparable frontier models.
The demo Alibaba leads with is a 35-hour autonomous coding session. The model fired 1,158 tool calls across 432 kernel evaluations to speed up a GPU kernel on hardware it had never seen in training, and reached a roughly 10x speedup over the reference implementation without a human in the loop. Treat the specifics as vendor-reported, because they are, but the underlying claim is about sustained focus: most models drift after a few dozen interactions, and holding a coherent plan across a thousand-plus tool calls is the hard part of agentic work.
There is a practical hook for teams already on Anthropic tooling: Qwen3.7-Max implements the Anthropic API protocol natively (documented in Alibaba DashScope developer reference, June 2026). Point Claude Code at the DashScope endpoint with three environment variables and the terminal agent runs on Qwen instead, with tool use and file edits intact. That makes it cheap to evaluate as a drop-in rather than a migration.
The Architecture: Gated DeltaNet and MoE
The defining technical choice of Qwen3 is how it handles attention. Standard transformer attention scales quadratically with sequence length, which is why long context windows get expensive fast. Qwen3 sidesteps that with a hybrid layout: three linear-complexity Gated DeltaNet blocks for every one full Gated Attention block, a 3:1 ratio repeated throughout the network.
Gated DeltaNet is a linear-attention mechanism that behaves more like a recurrent network than a classic transformer block. Instead of building a full attention matrix, it compresses history into a fixed-size memory state and updates that state token by token using a delta rule with decay and update gates. That is what scales linearly. The trade-off is that a fixed memory state is a bottleneck, so the interleaved full-attention block is retained to perform global attention and prevent precision loss over very long contexts.
The second pillar is sparsity. Qwen3 models use high-sparsity Mixture-of-Experts layers so total capacity can be enormous while the compute cost per token stays small. The open flagship Qwen3.5-397B-A17B carries 512 experts and routes each token to just 10 plus one shared expert, a roughly 1-in-50 activation ratio. The smaller Qwen3.6-35B-A3B uses 256 experts, routing to 8 plus a shared one. This is why a 397-billion-parameter model can run on a single high-memory machine: you pay for the active parameters, not the total.
A third efficiency trick is Multi-Token Prediction. During training the model learns to predict several future tokens at once, which natively supports speculative decoding at inference time and lets it emit more than one token per forward pass without a separate draft model.
Why this is the headline feature. The hybrid Gated DeltaNet design is the reason Qwen3 models reach 1 million token context windows with far flatter memory growth than a dense transformer. It is the architectural bet that separates Qwen3 from a straightforward copy of Western model recipes.
Thinking Modes and Thinking Preservation
Qwen3 ships with a dual-mode design that you control per request. In thinking mode, the model writes an internal reasoning trace inside
Alibaba publishes different sampling settings for each mode, which is a small but telling detail: temperature 0.6 with top-p 0.95 for thinking, temperature 0.7 with top-p 0.8 for non-thinking. The reasoning mode is tuned to be more deterministic, the fast mode slightly looser.
The newer capability, introduced with Qwen3.6, is thinking preservation. Most models throw away their chain-of-thought between conversation turns to save tokens, then rebuild it from scratch when the next hard step arrives. Qwen3.6 and later can retain reasoning traces across turns. In a long-running agent workflow that keeps decision-making consistent, reduces redundant re-reasoning, and improves how the model uses its key-value cache. It is enabled with a single API flag.
The takeaway for builders is that Qwen3 treats reasoning as a budget you allocate rather than a fixed behavior. You decide when the model should think out loud, when it should answer instantly, and whether it should carry its reasoning forward. For agent pipelines running hundreds of turns, those switches change both cost and reliability.
How Qwen3.7-Max Compares to Claude and GPT
For mid-2026, the framing is that Qwen3.7-Max, Claude Opus 4.7, and GPT-5.5 are roughly co-leaders, and the right pick depends on the workload. Qwen wins on agentic coding, tool orchestration, and cost. It loses on raw intelligence and knowledge-heavy question answering. The benchmark profile is unusual: it trails on chat quality but leads on the metrics that matter for production agents.
On SWE-Bench Pro, the harder version of SWE-Bench built from 2,294 real GitHub issues, Qwen3.7-Max scores 60.6%, ahead of Claude Opus 4.6 at 57.3%. On Terminal-Bench 2.0-Terminus, which simulates a software engineer in a sandboxed terminal with a multi-hour timeout, it scores 69.7%, the top public result at launch. These are vendor-reported figures run on Alibaba's own agent scaffold, so weight them accordingly, but SWE-Bench Pro is verified against held-out issues and is hard to game.
One number deserves a caution flag. On factual recall, measured by Artificial Analysis's AA-Omniscience benchmark, Qwen3.7-Max scores only about 30% accuracy. That is by design: the model was tuned to abstain rather than guess, dropping its attempt rate from 67% to 48% versus the prior generation. Hallucinations go down, but so does raw recall. For a search or knowledge-QA product that needs the model to commit to an answer, test on your own data before committing. For a fuller head-to-head with another Chinese lab, see our Qwen vs DeepSeek comparison.
Open Weights vs Proprietary: Access and Pricing
The Qwen3 family splits cleanly into two access models. The Max tiers are proprietary and API-only, so you rent intelligence by the token and route data to Alibaba's infrastructure. The mid-tier and open flagships ship under Apache 2.0, so you download the weights and run them anywhere, with no per-token cost after hardware. Knowing which tier a model belongs to is the first decision, not the last.
| Model | Params (total / active) | Context | Modality | License | Input price /1M |
|---|---|---|---|---|---|
| Qwen3.7-Max | ~1T / ~24B (MoE) | 1M | Text only | Proprietary, API | $2.50 |
| Qwen3.7-Plus | Undisclosed | Undisclosed | Multimodal | Proprietary, API | Not listed |
| Qwen3.5-397B-A17B | 397B / 17B (MoE) | 262K (to 1M) | Text, image, video | Apache 2.0 | ~$1.25-1.35 |
| Qwen3.6-35B-A3B | 35B / 3B (MoE) | 262K (to 1M) | Text, image, video | Apache 2.0 | $0.15 |
| Qwen3.6-27B | 27B (dense) | 262K (to 1M) | Text, image, video | Apache 2.0 | Self-host |
The pricing gap against Western frontier models is the headline. Qwen3.7-Max runs $2.50 per million input tokens and $7.50 per million output, with a 90% cached-input discount that drops reads to $0.25. For comparison, Claude Opus 4.7 lists at $15 input and $75 output, and GPT-5.5 at $10 and $30 (as of June 2026 list pricing; both vendors update pricing regularly). That makes Qwen3.7-Max roughly six times cheaper than Opus on input and four times cheaper than GPT-5.5, before the cache discount that meaningfully changes the math for agent workflows reusing system prompts across hundreds of turns.
Picking inside the family comes down to whether you need the absolute frontier, control over the weights, the lowest possible cost, or multimodal input. The four common profiles map cleanly to four models.
For the full per-tier cost breakdown including the cached-input math and the free-access options, see our Qwen pricing guide, and for hands-on API setup the Qwen API guide.
Limitations and Trade-offs
Qwen3's pricing and open-weight availability make a strong case, but the family carries real constraints that depend on which model and tier you choose. Four are worth weighing before you commit a production workload.
Frequently Asked Questions
Go Deeper
Resources from across Tech Jacks Solutions
Qwen, Qwen3, Tongyi Qianwen, and Alibaba Cloud are trademarks of Alibaba Group Holding Limited. Claude is a trademark of Anthropic. GPT is a trademark of OpenAI. DeepSeek is a trademark of its respective owner. All product names, logos, and brand identifiers are the property of their respective owners. Tech Jacks Solutions has no commercial relationship with Alibaba Cloud. This article is editorially independent.