Qwen3.6 (35B-A3B and 27B)

2026-04-15 · Publisher : Alibaba · Domain : Artificial Intelligence

Alibaba publie Qwen3.6 en poids ouverts : Qwen3.6-35B-A3B (MoE) le 15 avril 2026 puis Qwen3.6-27B (dense) le 22 avril, sous Apache 2.0, avec un code agentique qui dépasse celui du fleuron ouvert précédent.

Alibaba releases Qwen3.6 with open weights: Qwen3.6-35B-A3B (MoE) on April 15, 2026, then Qwen3.6-27B (dense) on April 22, under Apache 2.0, with agentic coding that surpasses that of the previous open flagship. ### Context Two months after Qwen3.5-397B-A17B, the Qwen team is trying to concentrate agentic capability in models that can be deployed on few GPUs. The hosted version Qwen3.6-Plus is offered through the API only, while the compact models are open. ### What the model brings - Qwen3.6-27B (dense) reaches 77.2 on SWE-bench Verified and 53.5 on SWE-bench Pro, ahead of Qwen3.5-397B-A17B (76.2 and 50.9) on the same tests in the model card. - Qwen3.6-35B-A3B (3B active) improves markedly on agentic coding compared with Qwen3.5-35B-A3B: SWE-bench Verified 73.4 versus 70.0, Terminal-Bench 2.0 51.5 versus 40.5. - Both models are native vision-language (text, image, video). - Thinking control: a thinking mode enabled by default, with preservation of historical reasoning (`preserve_thinking`). - FP8 weights published in parallel to reduce memory. ### Architecture and specifications - Qwen3.6-35B-A3B: 35B total, 3B active; 40 layers; 256 experts. - Qwen3.6-27B: dense, 27B; 64 layers. - Hybrid Gated DeltaNet and Gated Attention architecture, carried over from Qwen3.5. - Context: 262,144 tokens natively, extensible up to 1,010,000 for the 35B-A3B; the model card recommends at least 128k to preserve thinking. - License: Apache 2.0. - Weights: Qwen/Qwen3.6-35B-A3B and Qwen/Qwen3.6-27B (Hugging Face). Knowledge cutoff: not disclosed. ### Cost and access | Model | Input (USD / M tokens) | Output (USD / M tokens) | |---|---|---| | qwen3.6-35b-a3b (Qwen Cloud) | 0.375 | 2.25 | | qwen3.6-27b (Qwen Cloud) | 0.6 | 3.6 | Pricing recorded on September 29, 2026 on Qwen Cloud (not necessarily the launch pricing). The service lists 262K of context and 65K of maximum output. Hardware: the model card shows a typical launch on 8 GPUs with tensor parallelism, with no minimum disclosed. ### Benchmarks | Test | Claude 4.5 Opus | Gemma4-31B | Qwen3.5-397B-A17B | Qwen3.6-27B | |---|---|---|---|---| | SWE-bench Verified | 80.9 | 52.0 | 76.2 | 77.2 | | SWE-bench Pro | 57.1 | 35.7 | 50.9 | 53.5 | | Terminal-Bench 2.0 | 59.3 | 42.9 | 52.5 | 59.3 | | GPQA Diamond | 87.0 | 84.3 | 88.4 | 87.8 | | LiveCodeBench v6 | 84.8 | 80.0 | 83.6 | 83.9 | | AIME26 | 95.1 | 89.2 | 93.3 | 94.1 | Source: Qwen3.6-27B Hugging Face model card, scores published by the vendor. ### Versus closed models The reference closed model in this comparison is Claude 4.5 Opus. On Terminal-Bench 2.0, Qwen3.6-27B matches this model (59.3 on both sides). It remains behind on SWE-bench Verified (77.2 versus 80.9) and SWE-bench Pro (53.5 versus 57.1). On GPQA Diamond it is slightly ahead (87.8 versus 87.0). This comparator is a model from the previous generation, released in late 2025, and not the best closed model of April 2026, which the model card does not cite. On price, the Qwen API for the 27B is at 0.6 and 3.6 USD per million tokens, and the model fits in a self-hostable 27B dense model under Apache 2.0. ### Key takeaways Qwen3.6 shows that a 27B dense model can surpass the previous generation's open flagship on agentic coding. For companies, self-hosting a model of this size becomes credible without a restrictive license.

Tags : LLM, Open Source, Alibaba, Qwen, Apache 2.0, MoE, Multimodal, Code