Qwen3.6

Qwen3.6 is a generation of large language models from the Qwen team at Alibaba, released in April 2026 as the successor to Qwen3.5. ^[1][2] The line keeps the hybrid architecture that Qwen3.5 had adopted generation-wide, pairing Gated DeltaNet linear attention with a sparse mixture of experts and a native vision encoder, but reorients the release around agentic coding and real-world tool use. ^[2][3] It spans hosted proprietary endpoints (Qwen3.6-Plus, Qwen3.6-Flash, and the Qwen3.6-Max-Preview flagship) alongside two open-weight checkpoints, Qwen3.6-35B-A3B and Qwen3.6-27B, both published under the Apache 2.0 license. ^[1][4][7] The generation held the top of Alibaba's model stack for roughly seven weeks before Qwen3.7-Max replaced it in May 2026, but its two open checkpoints were still drawing more than six million downloads a month each on Hugging Face three months later. ^[4][7][16]

Overview

Qwen3.6 is an incremental but coding-focused update within the broader Qwen family. Where Qwen3.5 had launched under the banner "Towards Native Multimodal Agents," the Qwen team titled the Qwen3.6-Plus announcement "Qwen3.6-Plus: Towards Real World Agents," emphasizing reliability on long-horizon software tasks rather than a new architecture. ^[18][23] The team describes the generation as prioritizing stability and practical utility, shaped by community feedback on the previous release. ^[2]

Two themes run through the launch. The first is agentic coding: the models are tuned to plan, write, test, and iteratively debug code, and to operate inside third-party coding agents such as Claude Code, Cline, and Alibaba's own Qwen Code. ^[2][5] The second is visual reasoning carried over from the native multimodal foundation, so the models can read screenshots, hand-drawn wireframes, and design mockups and turn them into front-end code. ^[5][6] A smaller but notable addition is "thinking preservation," exposed in the model configuration as preserve_thinking, which retains a model's reasoning traces from earlier messages instead of discarding them after each response. ^[2][13]

The Qwen3 line and how 3.6 relates to 3.5

Qwen3.6 is best understood as a point update on the Qwen3.5 platform rather than a clean-sheet generation. Qwen3.5 had debuted on February 16, 2026 with a 397-billion-parameter flagship and established the family's defining engineering choices: a hybrid attention stack that interleaves Gated DeltaNet with periodic full attention, a sparse MoE feed-forward design, and early-fusion multimodal pretraining. ^[1][3] Qwen3.6 inherits all three. Its open-weight model cards describe the same interleave pattern, three Gated DeltaNet blocks for every one gated-attention block, and the same image-text-to-text framing as the Qwen3.5 checkpoints; the Hugging Face repository for Qwen3.6-35B-A3B is even tagged with the qwen3_5_moe architecture identifier. ^[4][7]

The clearest contrasts are in emphasis and packaging. Qwen3.5 led with the largest possible open model and a sweep of eight sizes down to 0.8B, released across three waves between February 16 and March 2, 2026; Qwen3.6 leads with hosted agentic endpoints and ships a narrower open-weight lineup aimed at coding. ^[1][2] The hybrid attention stack itself predates both. It shipped inside the Qwen3 series with Qwen3-Next-80B-A3B on September 11, 2025, whose model card gives the same layout the Qwen3.6 cards use, "12 * (3 * (Gated DeltaNet -> MoE) -> 1 * (Gated Attention -> MoE))," and it was carried into Qwen3-Coder-Next in March 2026. ^[19] What separates the 3.5 and 3.6 lines from the mainline Qwen3 releases, which were conventional dense and MoE transformers with full softmax attention, is that the hybrid became the default across an entire generation and was paired with a native vision encoder rather than a separate vision line like Qwen3-VL. ^[1][3][19]

Release

Alibaba announced Qwen3.6-Plus on April 2, 2026, positioning it as a proprietary, hosted upgrade over the Qwen3.5-Plus endpoint. ^[5][6] The hosted lineup then expanded with Qwen3.6-Max-Preview on April 20, 2026, billed by the team as the most capable model it had shipped to date, and a speed tier, Qwen3.6-Flash, which OpenRouter lists with a release date of April 27, 2026. ^[10][12]

The open-weight checkpoints landed between the two hosted announcements. According to the Qwen GitHub release notes, Qwen3.6-35B-A3B reached Hugging Face Hub and ModelScope on April 16, 2026, and the dense Qwen3.6-27B arrived on April 22, 2026. ^[1] Alibaba said it would continue to release selected Qwen3.6 models to the open-source community in "developer-friendly sizes" even as the top-tier Plus and Max-Preview endpoints stayed closed. ^[5][17] No Qwen3.6-Coder or separate Qwen3.6-VL checkpoint was released; the vision encoder is built into both open models, so there is no separate vision variant to publish. ^[1][4][7]

Variants and sizes

The Qwen3.6 release is split between proprietary hosted endpoints and open-weight checkpoints. The hosted tier follows Alibaba's usual naming (Max for the flagship, Plus for the balanced workhorse, Flash for the low-latency option), while the open releases are named by parameter count, with an "A3B" suffix marking 3 billion active parameters on the mixture-of-experts model. ^[1][3][10]

Scroll sideways for more →

Both open checkpoints ship in BF16 alongside fine-grained FP8 versions, Qwen3.6-27B-FP8 and Qwen3.6-35B-A3B-FP8, whose quantization_config uses the e4m3 format with a 128 by 128 weight block size, and both carry a single-layer multi-token-prediction head (mtp_num_hidden_layers: 1) for speculative decoding. Those four repositories are the entire official Qwen3.6 lineup on Hugging Face; standalone multi-token-prediction builds exist only as community conversions. ^[4][7][21] Sources for the table: the Qwen GitHub release notes, the official Hugging Face model cards, Alibaba Cloud's launch announcement, Decrypt's report on Max-Preview, and OpenRouter's Flash listing. ^{[1][4][5][7][10][12]}

Architecture as disclosed

The two open-weight model cards spell out the architecture in detail; Alibaba has not published comparable internals for the hosted Plus, Flash, or Max-Preview endpoints, and has not disclosed their parameter counts. ^[4][7][9]

Qwen3.6-35B-A3B is a sparse mixture-of-experts model with 35 billion total parameters and roughly 3 billion active per token, built from 40 layers with a hidden dimension of 2,048. ^[4] Its MoE blocks hold 256 experts, of which a router activates 8 routed experts plus 1 always-on shared expert per token. ^[4] Attention follows the Qwen3.5 hybrid template: the layers are arranged as 10 repeated blocks, each laid out as three Gated DeltaNet sublayers followed by one gated-attention sublayer, with a MoE feed-forward network after every sublayer. ^[4] The Gated DeltaNet sublayers use 32 linear-attention heads for values and 16 for queries and keys (head dimension 128), while the gated-attention sublayers use 16 query heads and 2 key/value heads (head dimension 256) with a 64-dimensional rotary position embedding. ^[4] Gated DeltaNet is a linear-attention mechanism whose memory cost stays constant as a sequence grows, which is what lets the model serve long contexts cheaply, while the periodic full-attention layers preserve precise long-range recall. ^[4]

The dense Qwen3.6-27B uses the same hybrid pattern at a different scale: 64 layers organized as 16 blocks of three Gated DeltaNet sublayers plus one gated-attention sublayer, a hidden dimension of 5,120, a padded token embedding of 248,320, and standard feed-forward networks in place of the MoE blocks. ^[7][13] Both checkpoints are described as causal language models with a vision encoder, exposed as image-text-to-text systems, and both are trained with multi-token prediction to support speculative decoding. ^[4][7]

Both open models carry a native context window of 262,144 tokens, which the model cards say can be extended to roughly 1,010,000 tokens with YaRN rope scaling (factor 4.0, rope_theta 10000000). ^[4][7] The hosted Qwen3.6-Plus and Qwen3.6-Flash endpoints are offered with a 1-million-token context window by default, while Max-Preview is served at 256K. ^[6][8][12]

Capabilities

The Qwen3.6 launch centers on agentic software work. Alibaba's launch release says Qwen3.6-Plus "is optimized for the 'capability loop', the ability to perceive, reason, and act within a single workflow," and that the model plans, tests, and iterates on code toward production-ready solutions. ^[17] The open checkpoints are tuned for front-end workflows and repository-level reasoning, and they support tool use and function calling so they can drive an AI coding agent end to end. ^[2][4] Alibaba lists compatibility with third-party agents including OpenClaw, Claude Code, and Cline; the Qwen3.6 repository separately points at Alibaba's own Qwen Code, "an open-source AI agent for the terminal, optimized for Qwen models." ^[2][17] The Max-Preview endpoint is exposed through an API that accepts both the OpenAI and the Anthropic request formats, which is what lets it slot into Claude Code and similar clients without a custom adapter; Qwen3.6-Plus had already shipped with Anthropic protocol compatibility at its April launch. ^[10][14]

Because the models inherit the early-fusion vision encoder from Qwen3.5, visual reasoning is a first-class capability rather than an add-on. Alibaba's launch release says Qwen3.6-Plus is engineered to combine cross-modal information for "high-density document parsing, physical-world visual analysis, and long-form video reasoning," and that it "can now interpret user interface screenshots, hand-drawn wireframes, or product prototypes and instantly generate functional frontend code." ^[17] Caixin's report of the same launch describes the model as producing code "from screenshots, design drafts and text prompts." ^[6] The open checkpoints are likewise multimodal, accepting text, image, and video input. ^[4][7] The hosted flagship is the exception: Max-Preview launched as a text-only endpoint despite sitting at the top of a multimodal generation. ^[10]

The preserve_thinking feature lets a model keep its chain-of-thought context across the messages in a conversation instead of resetting it each turn, which the team presents as helping continuity on multi-step agentic sessions. ^[2][13] Thinking mode is on by default in both open checkpoints and can be switched off with "enable_thinking": False. ^[4][7]

Vendor-reported benchmarks

Quantitative results are published for the open-weight checkpoints on the Hugging Face model cards. All of the figures below are Alibaba's own measurements, including the comparison columns for rival models, which the Qwen team selected and ran itself.

Scroll sideways for more →

Sources: the Qwen3.6-27B and Qwen3.6-35B-A3B Hugging Face model cards. ^[4][7]

The pattern Alibaba highlights is that the dense 27B beats not only its own predecessor but also the 397-billion-parameter Qwen3.5 flagship on every agentic-coding row, a gap that is widest on SkillsBench (48.2 against 30.0). ^[7][13] The MoE checkpoint shows the same direction of travel against its own predecessor, with SkillsBench moving from 4.4 to 28.7 and Terminal-Bench 2.0 from 40.5 to 51.5. ^[4] Knowledge and math scores barely move; the generation's gains are concentrated in software and tool use.

For the hosted flagship, Alibaba said Qwen3.6-Max-Preview took first place on six coding-oriented benchmarks (SWE-bench Pro, Terminal-Bench 2.0, SkillsBench, QwenClawBench, QwenWebBench, and SciCode), reported a roughly 2.3 percent gain on the SuperGPQA reasoning test and about 5.3 percent on a Chinese-language benchmark over Qwen3.6-Plus, and claimed it beat Claude 4.5 and GLM 5.1 on agentic tasks. ^[10] Alibaba published no full result table for Max-Preview, and two of the six benchmarks it topped are Qwen's own internal evaluations, so those figures should be read as marketing rather than measurement.

Independent evaluation

Artificial Analysis is the main third party publishing comparable numbers across all four Qwen3.6 endpoints. Its Intelligence Index v4.1 aggregates nine evaluations (GDPval-AA v2, tau-cubed Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR) and is run independently of the vendor. Scores are not comparable across index versions, so every figure below is v4.1 as measured on 2026-07-27. Output speed is a live measurement that drifts between readings.

Scroll sideways for more →

Sources: Artificial Analysis model pages, measured 2026-07-27. The two hosted endpoints are ranked against the full 190-model field; the two open checkpoints are ranked within Artificial Analysis's open-weights class. ^{[8][9][11][15]}

Two findings cut against the launch messaging. The hosted flagship and the mid-tier Plus endpoint both land on 40, fiftieth and fifty-third of 190 models, even though Max-Preview costs 2.6 times as much per input token; whatever edge the flagship holds sits inside the rounding. And the hosted pair was measured as slow for its price band, as was the dense 27B, which Artificial Analysis calls notably slow at 55.6 tokens per second; only the 3-billion-active-parameter MoE checkpoint was fast, at 156.1 tokens per second against a median of 98.6 for comparable open-weight models. ^{[8][9][11][15]} Artificial Analysis also flagged Plus as unusually verbose, consuming 100 million output tokens to complete the index against a 63 million median, and it now marks Max-Preview as deprecated, saying it continues performance benchmarking only for the default 10,000-token input workload. ^[8][9]

Pricing

Scroll sideways for more →

Caixin reported the launch price for Qwen3.6-Plus on Alibaba's Bailian platform as 2 yuan per million input tokens and 12 yuan per million output tokens, about $0.29 per million input tokens at the time. ^[6] International list prices are higher than the mainland ones. The OpenRouter figure for Flash is a discounted rate: the listing shows $0.1875 in and $1.125 out after a 25 percent promotional discount that was still live on 2026-07-27, implying $0.25 and $1.50 undiscounted, and the listing carries separate cache-read and cache-write pricing. ^[12] The open checkpoints have no list price of their own, since anyone can serve them; the figures above for 27B and 35B-A3B are Alibaba's own API rates as recorded by Artificial Analysis, which notes that pricing varies by provider. ^[11][15]

License and availability

The open-weight checkpoints are released under the permissive Apache 2.0 license, with license files in their respective Hugging Face repositories; the hosted endpoints are proprietary and reachable only through Alibaba's API. ^[2][4][7]

Scroll sideways for more →

The open checkpoints run under common serving stacks including vLLM 0.19.0 or later, SGLang 0.5.10 or later, KTransformers, and Hugging Face Transformers, with community builds for llama.cpp and MLX. ^[1][4][7]

Adoption

The open checkpoints were downloaded heavily and quickly. As of 2026-07-27, three months after release, the Hugging Face page for Qwen3.6-35B-A3B reported 6,187,853 downloads in the previous month and 2,530 likes, and Qwen3.6-27B reported 6,101,101 downloads and 2,070 likes across 48 open discussions. Downloads on Hugging Face are a rolling 30-day count, so the figure moves day to day. ^[4][7]

Derivative activity is the better measure of an open weights release, and by that standard Qwen3.6 was absorbed into the ecosystem within weeks. The counts are kept per repository rather than per generation: on the same date Hugging Face listed 658 quantizations, 185 finetunes, 221 adapters, and 22 merges derived from Qwen3.6-35B-A3B, and separately 650 quantizations, 312 finetunes, 330 adapters, and 6 merges derived from Qwen3.6-27B, with more than 100 community Spaces running the 27B model. ^[4][7] Third-party quantizers including Unsloth published GGUF and FP8 conversions; Unsloth's 4-bit GGUFs of the 27B run from 15.4 GB for IQ4_XS to 17.6 GB for the dynamic UD-Q4_K_XL build, small enough to fit a 24 GB consumer card. ^[21]

That sits inside a much larger Qwen ecosystem. More than 200,000 variants of open-source Qwen models have been published on Hugging Face, and the Qwen consumer app reported 234 million users by May 2026. ^[3]

Position in the Chinese open-weight field

Qwen3.6 continued a pattern Alibaba had settled into by 2026: permissive licensing for mid-size checkpoints, closed API access for the flagship. The two open models sit in a size band chosen for practical deployment, a 27-billion-parameter dense model that fits on a single high-memory accelerator and a 35B mixture of experts that activates only 3 billion parameters per token, rather than at the frontier. Reporting on the Max-Preview launch treated the closed flagship as the notable move, since the Qwen brand had been built on open releases. ^[10]

The 35B-A3B checkpoint is the clearest illustration of where the sparse-plus-linear-attention design leads. It activates about 3 billion of its 35 billion parameters per token, roughly 8.6 percent, yet on Alibaba's own numbers it lands within a few points of the 397-billion-parameter Qwen3.5 flagship on agentic coding (SWE-bench Verified 73.4 against 76.2, Terminal-Bench 2.0 51.5 against 52.5), and it was the fastest of the four Qwen3.6 endpoints by a wide margin in independent measurement. Other Chinese labs, DeepSeek among them, were pursuing the same high-sparsity route in the same window. ^[4][7][11] Alibaba's open lineup this generation is narrower than the eight-size Qwen3.5 sweep, however, with no very large or very small open checkpoint, and other Chinese labs shipping in the same window, among them GLM-5.1, Kimi K2.6, DeepSeek V4, and MiniMax M2.7, competed directly for the same self-hosting audience.

Succession

Qwen3.6 had an unusually short run at the top. Qwen3.7-Max was announced at the Alibaba Cloud Summit in Hangzhou on May 20, 2026, followed by Qwen3.7-Plus in June 2026, both proprietary. ^[16] (Wikipedia dates Qwen3.7-Max to May 18, 2026; the summit reporting is the more specific source. ^[3]) At that event Liu Weiguang, senior vice-president of Alibaba's cloud computing unit, said "What we're building is China's AI factory," and the company described itself as the only AI and cloud company in China operating all five layers of the full AI stack: chips, agentic cloud, AI models, model service platforms, and agentic applications. Zhou Jingren, chief AI architect of the newly formed Alibaba Group Technology Committee and previously chief technology officer of Alibaba Cloud, said Qwen3.7-Max "consistently ranked among the top tier on various benchmarks and outperformed all other AI models in China." ^[16] Artificial Analysis scores Qwen3.7-Max at 46 on Intelligence Index v4.1, thirty-third of 190 models, against 40 and fiftieth for Qwen3.6-Max-Preview as of 2026-07-27. ^[8][20] The line continued with a preview of the 2.4-trillion-parameter Qwen3.8-Max in July 2026, which Alibaba said it would release with open weights. ^[3]

Qwen3.6's successors in the flagship line have been hosted-only so far, so its two Apache 2.0 checkpoints remained the newest open-weight releases in Qwen's general-purpose model line as of 2026-07-27, which is part of why their download numbers stayed high months after release. The Qwen team kept publishing open weights in other lines during the same window, including the Apache 2.0 Qwen-AgentWorld-35B-A3B environment-simulation model in June 2026 and the Qwen3-ASR speech models in July 2026, so Qwen3.6 is not the last open Qwen release of any kind. ^{[3][4][7][22]}

Limitations

Most published internals and benchmarks cover only the two open-weight checkpoints. Alibaba has disclosed nothing about the parameter counts or architecture of the hosted Plus, Flash, and Max-Preview endpoints, and the Max-Preview claims come from the company's own announcement without a published result table. ^[4][7][10] The headline gains are concentrated in software and tool-use tasks: knowledge and math scores moved by roughly a point over Qwen3.5, so the generation's improvements do not obviously transfer to other domains. ^[4][7] The long-context claim of roughly one million tokens for the open models depends on YaRN rope scaling rather than native training at that length, a method that can degrade quality on the longest inputs, and the model cards advise reducing the served context if memory is tight. ^[4][7] Independent measurement also found the hosted endpoints slow and verbose, and gave the flagship no intelligence advantage over the cheaper Plus tier. ^[8][9]

References

1. ^"QwenLM/Qwen3.6 repository and release notes." GitHub, accessed 2026-07-25. github.com/...Qwen3.6
2. ^"Qwen3.6 README." QwenLM/Qwen3.6 on GitHub, accessed 2026-07-25. github.com/...README.md
3. ^"Qwen." Wikipedia, accessed 2026-07-25. en.wikipedia.org/...Qwen
4. ^"Qwen/Qwen3.6-35B-A3B model card." Hugging Face, 2026-04-16. huggingface.co/...Qwen3.6-35B-A3B
5. ^"Alibaba Unveils Qwen3.6-Plus to Accelerate Agentic AI Deployment for Enterprises and Alibaba's AI Applications." Alibaba Cloud Community, 2026-04-02. alibabacloud.com/...%80%99s-ai-applications_603000
6. ^"Alibaba Releases Qwen 3.6-Plus AI Model With Enhanced Coding Capabilities." Caixin Global, 2026-04-02. caixinglobal.com/...-coding-capabilities-102430395
7. ^"Qwen/Qwen3.6-27B model card." Hugging Face, 2026-04-22. huggingface.co/...Qwen3.6-27B
8. ^"Qwen3.6 Max Preview: Intelligence, Performance and Price Analysis." Artificial Analysis, accessed 2026-07-25. artificialanalysis.ai/...qwen3-6-max
9. ^"Qwen3.6 Plus: Intelligence, Performance and Price Analysis." Artificial Analysis, accessed 2026-07-25. artificialanalysis.ai/...qwen3-6-plus
10. ^"Alibaba Drops Qwen 3.6 Max Preview, Its Most Powerful Model Yet." Decrypt, 2026-04-20. decrypt.co/...-3-6-max-preview-most-powerful-model
11. ^"Qwen3.6 35B A3B: Intelligence, Performance and Price Analysis." Artificial Analysis, accessed 2026-07-25. artificialanalysis.ai/...qwen3-6-35b-a3b
12. ^"Qwen3.6 Flash: API Pricing and Providers." OpenRouter, accessed 2026-07-25. openrouter.ai/...qwen3.6-flash
13. ^"Alibaba Qwen Team Releases Qwen3.6-27B: A Dense Open-Weight Model Outperforming 397B MoE on Agentic Coding Benchmarks." MarkTechPost, 2026-04-22. marktechpost.com/...e-on-agentic-coding-benchmarks
14. ^"Alibaba's Qwen launches new flagship LLM with Qwen 3.6-Plus." Constellation Research, 2026-04-02. constellationr.com/...ew-flagship-llm-qwen-36-plus
15. ^"Qwen3.6 27B: Intelligence, Performance and Price Analysis." Artificial Analysis, accessed 2026-07-25. artificialanalysis.ai/...qwen3-6-27b
16. ^"Alibaba unveils new Qwen model, custom chips in bid to become China's AI factory." South China Morning Post, 2026-05-20. scmp.com/...tom-chips-bid-become-chinas-ai-factory
17. ^"Press release: Alibaba unveils Qwen 3.6 Plus to accelerate agentic AI deployment for enterprises and Alibaba's AI applications." Treasury Today, 2026-04-02. treasurytoday.com/...-and-alibabas-ai-applications
18. ^"Qwen3.6-Plus: Towards Real World Agents." Qwen blog, mirrored on Alibaba Cloud Community, 2026-04-02. alibabacloud.com/...wards-real-world-agents_603005
19. ^"Qwen/Qwen3-Next-80B-A3B-Instruct model card." Hugging Face, 2025-09-11. huggingface.co/...Qwen3-Next-80B-A3B-Instruct
20. ^"Qwen3.7 Max: Intelligence, Performance and Price Analysis." Artificial Analysis, accessed 2026-07-27. artificialanalysis.ai/...qwen3-7-max
21. ^"unsloth/Qwen3.6-27B-GGUF." Hugging Face, accessed 2026-07-27. huggingface.co/...Qwen3.6-27B-GGUF
22. ^"Qwen/Qwen-AgentWorld-35B-A3B model card." Hugging Face, June 2026. huggingface.co/...Qwen-AgentWorld-35B-A3B
23. ^"Qwen3.5: Towards Native Multimodal Agents." Alibaba Cloud Community, 2026-02-17. alibabacloud.com/...ative-multimodal-agents_602894