Best Qwen Models in 2026

Alibaba ships a new Qwen generation every few months. Here’s what’s actually worth using right now, and how to pick.

Any guide to the best Qwen model published before 2026 is describing a different product. Alibaba shipped three generations between February and June: Qwen 3.5, 3.6, and 3.7. Each shifted which model is worth recommending, and one changed the rules entirely.

This is July 2026. Here’s what’s worth using now, and how to pick.

Qwen ships fast: here’s what’s current as of July 2026

The 2026 Qwen release calendar moved faster than most companies’ full-year roadmaps:

• February 2026: Qwen 3.5 launches with the open-weight flagship Qwen3.5-397B-A17B. Qwen 3.5 Flash, the hosted API version of Qwen3.5-35B-A3B, follows within a week. Smaller variants down to 0.8B release by early March.
• April 2026: Qwen 3.6 arrives in two forms. Qwen3.6-35B-A3B (MoE, Apache 2.0) ships April 16. Qwen3.6-27B, a dense 27B model, follows April 22 and outperforms the larger Qwen3.5-397B-A17B flagship on agentic coding, while fitting on a single consumer GPU.
• May/June 2026: Qwen3.7-Max-Preview surfaces on LM Arena May 14. The API goes live May 19. The formal announcement comes at the Alibaba Cloud Summit in Hangzhou on May 20. Qwen3.7-Plus, the multimodal, closed-weight sibling, follows on June 1.

Two facts stand out beyond the pace. First, the 3.6 generation showed a well-designed 27B dense model can outperform a 397B MoE on specific tasks, so bigger no longer automatically means better. Second, Qwen3.7-Max is Alibaba’s first fully closed-weight flagship: no weights on Hugging Face, no local deployment. This is a deliberate strategic choice rather than a gap, and it reshapes the decision framework for teams that need model weights. Qwen3.7-Plus follows the same closed-weight, API-only pattern. The pivot applies to the whole 3.7 flagship generation, a clear break from the fully open 3.5 and 3.6 lines.

Both facts belong in any honest answer to “which Qwen model should I use?” Two pre-2026 models, QwQ-32B (March 2025) and Qwen3-235B-A22B (April 2025), also appear in this guide: both are still useful for specific scenarios covered below.

The best Qwen models in 2026

Qwen3.7-Max and Qwen3.7-Plus parameter counts are undisclosed: Alibaba hasn’t published architecture details for either model. The 256K context figure for Qwen3-235B-A22B reflects the -2507 Instruct release specifically; earlier deployments ran 131K via YaRN extension.

Best Qwen model for coding: Qwen3-Coder-480B-A35B

No Coder-specific variant has shipped with Qwen 3.6 or 3.7. Qwen3-Coder-480B-A35B-Instruct remains the dedicated coding flagship by default, and it still earns that position on benchmarks.

When you need the ceiling: API-based production use

Qwen3-Coder-480B-A35B-Instruct achieved state-of-the-art performance among open models on SWE-Bench Verified at its July 2025 release, and Together AI benchmarks place it in range of Claude Sonnet 4 on agentic coding. The 256K context window, extendable to 1M with extrapolation, handles large repository contexts without truncation, under an Apache 2.0 license.

The total parameter count is 480B, but MoE architecture means you’re paying for 35B active parameters at inference. For production coding pipelines where quality is the constraint, this is the API call to make.

Local and lightweight coding: Qwen3.6-27B vs Qwen3-Coder-Next

For local or self-hosted coding, two options fit different roles.

Qwen3.6-27B is the better general local pick for most teams. It’s dense, with no MoE routing complexity, so it’s straightforward to quantize and run in Ollama or llama.cpp. Per Alibaba, it outperforms Qwen3.5-397B-A17B on agentic coding while fitting on a single GPU. If your local use case is general agentic development rather than deep specialized coding, start here.

Qwen3-Coder-Next (built on Qwen3-Next-80B-A3B-Base, ~3B active parameters) scores over 70% on SWE-Bench Verified, well above what its active parameter count would predict. It’s the pick when you need a coding-specific local model and want to maximize benchmark performance per hardware dollar.

These aren’t competing for the same role. Qwen3.6-27B is the general local workhorse that handles coding well; Qwen3-Coder-Next is the dedicated coding specialist for teams where coding is the only local inference use case.

Best Qwen model for reasoning and agentic work: Qwen3.7-Max

Qwen3.7-Max is the current performance ceiling in the Qwen lineup. It scores 60.6 on SWE-Bench Pro and 56.6 on the Artificial Analysis Intelligence Index, ranking as the highest-scoring Chinese model at time of writing, and sits at approximately 13th globally on LM Arena’s text leaderboard. For reasoning-intensive and agentic workloads compared against the frontier, these are the numbers that matter.

What changed: Alibaba’s first closed-weight flagship

Qwen3.7-Max has no public weights, worth naming directly: every major Qwen release before it shipped with weights on Hugging Face, and the 3.5 and 3.6 generations are still fully open under Apache 2.0. Qwen3.7-Max is API-only, available through Alibaba Cloud Model Studio or compatible gateway providers. Its multimodal sibling, Qwen3.7-Plus, follows the same closed-weight, API-only pattern across the entire 3.7 flagship generation.

Pricing reflects the positioning: $2.50 per million input tokens, $7.50 per million output, with a 90% cached-input discount at $0.25 per million. The 1M token context window and support for the Anthropic API protocol mean it integrates with tooling like Claude Code with minimal adapter work.

Alibaba demonstrated a 35-hour autonomous session with over 1,000 tool calls during a vendor showcase. That’s a vendor demonstration, not an independently verified result: take it as directional evidence of agentic stamina, not a confirmed specification.

When QwQ-32B is still the better call

QwQ-32B predates Qwen3. It’s built on Qwen2.5-32B, released March 2025, roughly a month before Qwen3 launched with native thinking mode across the family. That overlap is real: a team starting fresh today often covers the same reasoning ground with a current Qwen3 model in thinking mode, and gets more general capability in the same call.

QwQ-32B is the open-weight reasoning alternative. At 32B parameters with a 131K context window, it’s self-hostable on hardware with 24-48GB VRAM in quantized form. For teams that can’t route reasoning calls through a closed API due to compliance, data-residency, or cost-control constraints, QwQ-32B delivers genuine chain-of-thought reasoning without an external dependency.

It won’t match Qwen3.7-Max’s benchmark scores, but for many reasoning workloads it doesn’t need to. The gap shows up mainly at the hardest end of the distribution: multi-hop research, complex proofs, long agentic sessions. For everything else, QwQ-32B covers the ground.

Best Qwen model for general production use: Qwen3-235B-A22B

Qwen3-235B-A22B launched April 29, 2025 as the original Qwen3 flagship. Over a year and three generations later, it’s no longer the frontier model in the lineup, but it still offers a known quantity: tested at scale, a predictable cost profile, Apache 2.0 licensing, and 256K context on the -2507 Instruct release.

The MoE architecture keeps per-call costs close to a 22B dense model rather than a 235B one. For RAG pipelines, multilingual workloads (100+ languages), and general production use where you’re not pushing capability ceilings, it remains a solid mid-tier option.

Teams already running integrations built around it have little reason to migrate unless evals show gaps a newer model would close. If you’re starting fresh and cost isn’t the deciding factor, evaluate the 3.6 or 3.7 generation before defaulting here.

Best Qwen model for speed and cost: Qwen 3.5 Flash

Qwen 3.5 Flash, the hosted API version of Qwen3.5-35B-A3B, is purpose-built for throughput. The 35B/3B active MoE architecture and pricing around $0.10/M input tokens make it the right model for high-volume, lower-complexity workloads: classification, intent detection, lightweight summarization, and customer-facing bots where per-call cost accumulates fast and latency is visible to end users.

It supports a 1M token context window and handles vision inputs, giving it more range than the price point suggests, with reasoning depth configurable at call time via chain-of-thought settings.

Qwen3.6 Flash exists as a newer sibling. If your provider offers it, benchmark it against your workload; Qwen 3.5 Flash remains the reliable default where 3.6 isn’t yet available. One underused feature: the 1M context window makes it viable for long-document summarization, not just short, high-frequency requests.

How to choose: a quick reference

The right Qwen model is a function of your workload:

• Coding at API scale: Qwen3-Coder-480B-A35B
• Local coding: Qwen3.6-27B for general agentic work; Qwen3-Coder-Next for coding-specific local inference
• Deep reasoning or agentic pipelines: Qwen3.7-Max (closed API)
• Self-hosted reasoning: QwQ-32B
• General production: Qwen3-235B-A22B
• High-volume, low-latency: Qwen 3.5 Flash

How Infron Handles Qwen Model Access

Three Qwen generations in under five months, spanning open-weight, closed-weight, dense, and MoE architectures. Tracking which model is available on which provider at what latency is overhead before your application has done any real work.

One endpoint across every Qwen generation. Infron routes requests to Qwen3-Coder, Qwen3.7-Max, QwQ-32B, Qwen 3.5 Flash, and the rest of the current Qwen lineup through a single OpenAI-compatible endpoint. Switch models by changing a parameter, not by rewriting your provider integration.

Automatic routing to the fastest available host. Infron evaluates provider availability and latency on each request and routes to the best current option. When a provider serving Qwen3.7-Max is slow or unavailable, requests shift automatically, and your application doesn’t handle this.

99.99% uptime SLA across generations. Each Qwen release adds new providers and API endpoints. Infron AI’s routing layer abstracts the churn, giving you one consistent, SLA-backed endpoint that stays reliable as the Qwen lineup keeps evolving.

Access any Qwen model, and 400+ AI models in total, through a single API at Infron.

FAQ

Is Qwen open source?

Most Qwen models are open-weight under permissive licenses: the 3.5 and 3.6 lines are Apache 2.0. Qwen3.7-Max and Qwen3.7-Plus are the exceptions, marking Alibaba’s first fully closed-weight flagship generation, with no public weights and API-only access. If open-weight access is a hard requirement, stay within the 3.5 or 3.6 generation.

What’s the difference between Qwen 3.5, 3.6, and 3.7?

Qwen 3.5 (February 2026) set the baseline, with the 397B-A17B as flagship and Flash as the speed tier. Qwen 3.6 (April 2026) focused on efficiency: the 27B dense model outperforms the larger 3.5 flagship on agentic coding while fitting on a single GPU. Qwen 3.7 (May-June 2026) introduced the first closed-weight models at the top of the stack, Qwen3.7-Max and multimodal sibling Qwen3.7-Plus. Each generation improved on the last; if you have no strong reason to stay on 3.5, evaluating 3.6 or 3.7 is worth the time.

Which Qwen model is best for coding?

Qwen3-Coder-480B-A35B for API-based production coding. For local use, Qwen3.6-27B handles general agentic coding on a single GPU, and Qwen3-Coder-Next scores over 70% on SWE-Bench Verified with only ~3B active parameters. No dedicated Coder variant has shipped for 3.6 or 3.7 yet, so the 480B remains the coding flagship.

Can I run Qwen models locally?

Yes, for open-weight models. Qwen3.6-27B and QwQ-32B both run on consumer hardware with 24-48GB VRAM in quantized form. Smaller 3.5 variants (9B, 4B, 2B) work on lighter hardware. Qwen3.7-Max and Qwen3.7-Plus are exceptions: neither has locally runnable weights, both are API-only.