Qwen 3.6

Qwen 3.6 is the April 2026 generation of Alibaba Qwen's flagship open-weights language and multimodal model family, comprising the 27B dense variant and the 35B-A3B mixture-of-experts variant with 3 billion active parameters per token. The family operates as a unified image-text-to-text architecture with native support for hour-scale video understanding, a 256K context window extensible to approximately one million tokens via YaRN, and a hybrid thinking-or-instruct mode that generates <think> blocks before responding to complex tasks. As of May 2026, Qwen 3.6 sits in the leading tier of open-weights frontier models alongside DeepSeek V4 and the Apache 2.0 license places it among the most-permissively-licensed frontier-grade releases of the year.

At a glance

• Lab: Alibaba Qwen.
• Released: April 2026.
• Modality: Multimodal. Image, text, and video input; text output. Hour-scale video understanding supported.
• Open weights: Yes. Apache 2.0 license, including commercial use.
• Architecture: Two principal variants. Qwen3.6-27B is a dense 27 billion parameter model. Qwen3.6-35B-A3B is a sparse mixture-of-experts model with 35 billion total parameters, 3 billion active per token, 256 total experts (8 routed plus 1 shared activated), a hybrid Gated DeltaNet plus Gated Attention plus MoE block structure across 40 layers, and a 248,320-token vocabulary. Vision encoder integrated into the same model for multimodal input.
• Context window: 262,144 tokens native, extensible to approximately 1,010,000 tokens with YaRN positional scaling.
• Pricing: Open weights, free to self-host. On Fireworks AI the 35B-A3B variant lists at $1.40 per million input tokens. Alibaba Cloud DashScope and the Qwen API serve the family at first-party pricing tiers that have not been uniformly published.
• Distribution channels: Hugging Face (the 35B-A3B-FP8 card reported approximately 4.3 million monthly downloads at the time of release), vLLM, SGLang, KTransformers, Transformers, Fireworks AI, Alibaba Cloud DashScope, and the consumer Qwen Chat surface.

Origins

Qwen 3.6 follows the Qwen 3 generation released in 2025 and represents the third major iteration in the Qwen family's published progression toward hybrid sparse-and-dense architectures. The 35B-A3B variant continues the active-parameter-light design pattern that the Qwen team has emphasised across the 3.x generations: the architecture retains a large total parameter count for capacity, while routing through a small number of experts per token to keep inference cost competitive with much smaller dense models. The 3-billion-active-parameter figure in particular gives 35B-A3B an inference cost profile closer to a 3-to-5-billion-parameter dense model while retaining the capability headroom of a 35-billion-parameter capacity ceiling.

The release is part of a broader Alibaba Qwen roadmap that across 2025 and 2026 has included the Qwen 3 base generation (8B, 32B, 235B-A22B variants), the Qwen 3 Coder line at 480B-A35B, the Qwen 3 VL vision-specialised line, the Qwen 3.5 generation at 122B-A10B, and the QwQ reasoning-focused 32B model. Qwen 3.6 consolidates the multimodal capability into the flagship line: the 27B and 35B-A3B variants both natively handle image and video input, replacing the previous pattern of separate text and vision SKUs.

The hybrid thinking-or-instruct mode is the second consolidating change. Earlier Qwen generations shipped distinct reasoning-tuned and instruction-tuned variants (the QwQ line for reasoning, the Instruct line for fast-path answers). Qwen 3.6 unifies these into a single model whose default operating mode is the thinking mode, with an instruct mode available via different sampling parameters. The thinking mode emits <think>...</think> blocks before producing the user-facing answer, and the model retains thinking context from historical messages in multi-turn conversations.

The FP8-quantised release form, including the headline 35B-A3B-FP8 variant available through Fireworks AI and other hosted inference providers, reduces the memory footprint enough that the model fits within a single 80-gigabyte data-centre GPU at full context, materially lowering the deployment-cost floor for production inference.

Capabilities

The Qwen 3.6 capability profile spans three principal axes: agentic coding, multimodal understanding, and reasoning.

Agentic coding is the headline differentiator for the 35B-A3B variant. On SWE-bench Verified, the multi-turn software-engineering benchmark that has become the dominant coding evaluation, Qwen3.6-35B-A3B reports 73.4 percent. On Terminal-Bench 2.0, which evaluates terminal-based agentic tasks, the model reports 51.5 percent. The bilingual frontend code generation benchmark QwenWebBench reports 1,397 (a first-party benchmark whose absolute number is harder to compare externally but whose internal consistency suggests strong frontend-engineering capability in both English and Chinese). The model is specifically positioned for repository-level coding workflows and complex agentic coding tasks rather than single-function completion.

Multimodal understanding spans image and video. The model handles variable-aspect-ratio and variable-resolution image inputs, with hour-scale video understanding supported natively (rather than via frame-extraction preprocessing). Benchmark positions include RealWorldQA at 85.3 percent, VideoMMMU at 83.7 percent, and OmniDocBench 1.5 at 89.9 percent (the document-understanding benchmark covering charts, tables, and structured layouts). The OmniDocBench result in particular places Qwen 3.6 in the leading tier for document-OCR-and-structure understanding among open-weights models.

Reasoning and knowledge capability is anchored by MMLU-Pro at 85.2 percent and C-Eval at 90.0 percent. The thinking-mode architecture is most consequential in this dimension: the <think> blocks let the model produce extended reasoning traces before committing to an answer, with measurable benchmark deltas on tasks where step-by-step decomposition pays off. The thinking-preservation feature, which retains historical reasoning context across multi-turn conversations, is novel in the open-weights flagship category.

The tool-calling and agentic capability extend across the foundation-model surface. The model supports OpenAI-compatible function calling, integrates with the Alibaba Cloud agent frameworks, and is positioned as a drop-in replacement for closed-source frontier models in agent and workflow applications.

Benchmarks and standing

Qwen3.6-35B-A3B reports the following benchmark positions at release:

• SWE-bench Verified: 73.4 percent
• Terminal-Bench 2.0: 51.5 percent
• MMLU-Pro: 85.2 percent
• C-Eval: 90.0 percent
• RealWorldQA: 85.3 percent
• VideoMMMU: 83.7 percent
• OmniDocBench 1.5: 89.9 percent
• QwenWebBench: 1,397 (first-party)

The combined profile places Qwen 3.6 in the top tier of open-weights models across coding, multimodal, and reasoning axes as of May 2026. Direct comparison against DeepSeek V4, the principal open-weights peer, is mixed: DeepSeek V4 Pro retains a small edge on certain reasoning-heavy benchmarks while Qwen 3.6 leads on the agentic-coding and multimodal axes. Against closed-source frontier models, Qwen 3.6 is competitive on many task categories but trails on the highest-difficulty reasoning evaluations like ARC-AGI Challenge and the most-recent AIME competitions.

Benchmark leadership is point-in-time. The competitive set turns over on roughly a quarterly cadence in the 2025-to-2026 cycle, and the next significant Qwen, DeepSeek, GLM, and Gemma releases are expected through the second half of 2026.

Access and pricing

Qwen 3.6 ships as open-weights under Apache 2.0, allowing both research and commercial use without per-token licensing. Distribution and deployment options:

• Hugging Face Hub at Qwen/Qwen3.6-35B-A3B-FP8 and sibling repositories for the 27B variant. Direct download for self-hosting.
• Inference frameworks: vLLM, SGLang, KTransformers, Transformers. The FP8 quantisation is supported across the major serving stacks.
• Fireworks AI hosts the 35B-A3B variant at $1.40 per million input tokens at the time of release.
• Alibaba Cloud DashScope offers the first-party API surface, OpenAI-compatible.
• Qwen Chat is the consumer chat interface, free at the Standard tier.
• Local deployment: vLLM and SGLang scripts for single-GPU 80-gigabyte and multi-GPU configurations; Docker Model Runner integration.

Recommended sampling parameters from the model card: thinking mode for general tasks uses temperature 1.0, top-p 0.95, top-k 20, presence penalty 1.5. Coding tasks in thinking mode use temperature 0.6. Instruct mode uses temperature 0.7, top-p 0.80.

Comparison

• DeepSeek V4 (DeepSeek). The closest open-weights peer. DeepSeek V4 Pro at 1.6 trillion total parameters and 49 billion active sits at a substantially larger capacity ceiling; Qwen 3.6-35B-A3B sits at much lower inference cost. The competitive picture depends on which axis matters most: capability ceiling favours DeepSeek, inference economics favours Qwen.
• GLM-5.1 (Z.ai). The principal Chinese open-weights peer focused on agentic-engineering performance. GLM-5.1's 754 billion total parameters and its agentic-engineering benchmarks (SWE-Bench Pro at 58.4 percent, Terminal-Bench 2.0 at 63.5 percent) overlap with Qwen 3.6's agentic-coding positioning.
• Gemma 4 (Google DeepMind). The principal open-weights peer from a Western frontier lab. Gemma 4 31B IT, the closest size-and-modality match, is a dense model rather than MoE; the architectural trade-off favours Qwen 3.6 on inference cost and Gemma 4 on per-active-parameter capability per FLOP.
• Llama 4 (Meta AI). The Llama 4 generation sits at comparable scale but operates as a different architectural philosophy. The competitive question is research-and-fine-tuning ecosystem depth, where Llama has historically led and Qwen has been closing the gap rapidly.
• Kimi K2.6 (Moonshot AI). Another Chinese-origin open-weights peer with a different reasoning-versus-coding emphasis.

Outlook

Open questions for the next 6 to 18 months:

• Successor cadence. Qwen 3.6 lands roughly four months after the 3.5-generation 122B-A10B variant. Whether the cadence continues at this pace, and what scale jump (a Qwen 3.7 or a Qwen 4 generation) appears next, is the central technical-roadmap question.
• Independent multimodal-leaderboard reproduction. Qwen 3.6's RealWorldQA, VideoMMMU, and OmniDocBench numbers are first-party reports. Independent reproduction on the standard open multimodal leaderboards will determine whether the headline positions hold.
• Deployment-cost trajectory. The FP8 quantisation lowers the inference cost floor materially. Whether further INT4 or INT2 quantisations preserve capability at additional cost reductions, and how the open-source serving stacks adopt them, is a structural question for the open-weights deployment economy.
• Reasoning-mode adoption patterns. The unified thinking-or-instruct mode is the most-distinctive architectural change in 3.6. Whether developer applications standardise on always-thinking, always-instruct, or context-aware switching will indicate the practical fit of the unified design.
• Western export-control posture toward the model. The Apache 2.0 release and the Chinese-origin lab status interact with the evolving US-and-EU export-control posture on advanced-AI components. Whether Qwen 3.6 weights, derivative work, or hosted inference become subject to additional review will shape the model's enterprise-adoption pattern in Western markets.

Sources

• Hugging Face: Qwen3.6-35B-A3B-FP8. Primary model card with architecture, benchmark, and distribution details.
• Fireworks AI: model catalog. Pricing for the hosted 35B-A3B variant.
• Companion profile: Alibaba Qwen for the broader Qwen family roadmap, the Alibaba Cloud DashScope distribution surface, and the lab's positioning relative to DeepSeek, Moonshot AI, and Z.ai in the Chinese open-weights cohort.
• Companion model: Qwen 3 for the previous-generation profile.
• Companion model: DeepSeek V4 for the principal open-weights peer.