← Models

Qwen 3.6 Max Preview

Alibaba · Qwen3

Alibaba's flagship Qwen 3.6 Max Preview, a sparse MoE model for agentic coding, tool use, and long-context reasoning.

Part of Qwen3 family · Other versions: Qwen3.6-27B , Qwen3-Max
Type
language
Context
262K tokens
Max Output
66K tokens
Status
preview
Input
$1.3/1M tok
Output
$7.8/1M tok
API Access
Yes
License
proprietary
chinese multilingual reasoning coding agentic long-context tool-use
Released April 2026 · Updated May 1, 2026

Overview

Freshness note: Model capabilities, limits, and pricing can change quickly. This profile is a point-in-time snapshot last verified on May 1, 2026.

Qwen 3.6 Max Preview is Alibaba Cloud’s hosted flagship in the Qwen 3.6 generation. Alibaba positions it as a frontier reasoning and coding model with an integrated thinking mode that preserves reasoning traces across multi-turn conversations, sitting above the open-weight Qwen 3.6 27B and Qwen 3.6 35B-A3B variants released earlier in April 2026. The “Preview” label reflects that Alibaba is still rolling the model out and may adjust pricing or behavior before it reaches stable status.

This entry covers Max Preview as the hosted top-tier route. Open-weight Qwen 3.6 variants (27B and 35B-A3B) and the mid-tier hosted Plus variant are referenced in prose rather than getting separate entries at this snapshot.

Capabilities

Alibaba’s release materials and Hugging Face cards highlight a specific capability profile:

  • Strong agentic coding behavior, with Qwen 3.6 Max Preview reportedly leading a set of common coding benchmarks at launch.
  • Long-context reasoning across the full 262K context window, designed for retrieval-heavy analysis and multi-turn agent loops.
  • Integrated thinking mode that keeps reasoning traces available across turns, useful for tool-using agents that need to revisit earlier reasoning.
  • Multilingual coverage extending Qwen3’s existing 119-language footprint, with continued strength in Chinese-English bilingual workloads.
  • Tool calling and structured output suitable for production agent orchestration.

The open-weight Qwen 3.6 27B and 35B-A3B variants share the generation’s design philosophy but trade peak intelligence for self-hosted deployability and lower operating cost.

Technical Details

Public anchors at this snapshot:

  • Approximately 1 trillion total parameters in a sparse mixture-of-experts architecture.
  • 262K-token context window, with up to 65K max output.
  • Hosted only as a managed Alibaba Cloud Model Studio endpoint at this preview stage, not released as open weights.
  • OpenAI-compatible and Anthropic-compatible API surfaces through Model Studio.

Pricing & Access

Listed Alibaba Cloud Model Studio pricing (per 1M tokens):

  • Input: $1.30
  • Output: $7.80

OpenRouter and other gateways list slightly lower rates. Pricing may shift before the model exits preview, so production cost models should leave headroom.

Access options:

  • Alibaba Cloud Model Studio (DashScope) as the primary hosted route
  • OpenAI-compatible and Anthropic-compatible client SDKs against Model Studio
  • Third-party gateways such as OpenRouter

Best Use Cases

Choose Qwen 3.6 Max Preview for:

  • Frontier-tier agentic coding work where the long context window helps keep repo state in one prompt.
  • Tool-using agents that benefit from preserved reasoning traces across turns.
  • Bilingual Chinese-English assistants where Western frontier models still lag on Chinese-language nuance.
  • Long-context document analysis at hosted-API economics rather than self-hosted.

For self-hosted or air-gapped deployments, fall back to open-weight Qwen 3.6 27B or 35B-A3B variants. For lower-cost production traffic, the Qwen 3.6 Plus tier or Qwen 3.5 remains a reasonable starting point.

Comparisons

  • Claude Opus 4.7 (Anthropic): Western premium frontier alternative with broader enterprise distribution and stronger governance, at higher cost per token.
  • GPT-5.5 (OpenAI): Comparable frontier route with deeper product surfaces and tooling ecosystem; Qwen 3.6 Max competes more directly on long-context and Chinese-language behavior.
  • Kimi K2.6 (Moonshot AI): Chinese open-weight peer with explicit agent-swarm features; Qwen 3.6 Max stays proprietary at this tier but inherits the broader Qwen ecosystem.