Alibaba Cloud / Qwen Team model product
Qwen3.6-35B-A3B is the first open-weight variant of the Qwen3.6 series, a multimodal Mixture-of-Experts model with 35B total parameters and 3B activated. It pairs a vision encoder with a hybrid 40-layer language model that interleaves Gated DeltaNet linear-attention blocks and Gated Attention blocks (10 × (3 × DeltaNet + 1 × Attention)) over 256 experts (8 routed + 1 shared, expert dim 512). The release prioritizes stability and real-world utility, with substantial gains in agentic coding (frontend workflows, repo-level reasoning) and a new option to preserve reasoning context across turns. Native context length is 262K tokens, extensible to ~1M via YaRN, and the model thinks by default.
Updated Aug 10, 2026. Default version: Qwen3.6-35B-A3B
Structured fields from the published default version.
This profile uses the latest version under this unique model that has a calculated LLMBoard score. Arena and price are excluded.
Published benchmark records for the scored version Qwen3.6-35B-A3B.
| Hallusion Bench | 0.7 | 2 | 16 | 93.3% | C | |
| MMBench-V1.1 | 0.9 | 2 | 18 | 94.1% | C | |
| CC-OCR | 0.8 | 3 | 18 | 88.2% | C | |
| ODinW | 0.5 | 3 | 16 | 86.7% | C | |
| ZEROBench-Sub | 0.3 | 3 | 5 | 50.0% | C | |
| DeepPlanning | 0.3 | 4 | 9 | 62.5% | C | |
| EmbSpatialBench | 0.8 | 4 | 8 | 57.1% | C | |
| OmniDocBench 1.5 | 0.9 | 4 | 16 | 80.0% | C | |
| RefCOCO-avg | 0.9 | 4 | 7 | 50.0% | C | |
| TAU3-Bench | 0.7 | 4 | 5 | 25.0% | C | |
| VideoMME w/o sub. | 0.8 | 4 | 10 | 66.7% | C | |
| ZClawBench | 0.5 | 4 | 4 | 0.0% | C | |
| MathVista-Mini | 0.9 | 5 | 23 | 81.8% | C | |
| MLVU | 0.9 | 5 | 10 | 55.6% | C | |
| MVBench | 0.7 | 5 | 17 | 75.0% | C | |
| RefSpatialBench | 0.6 | 5 | 6 | 20.0% | C | |
| RealWorldQA | 0.9 | 6 | 26 | 80.0% | C | |
| VITA-Bench | 0.4 | 6 | 10 | 44.4% | C | |
| SkillsBench | 0.3 | 7 | 7 | 0.0% | C | |
| VideoMME w sub. | 0.9 | 7 | 10 | 33.3% | C | |
| AI2D | 0.9 | 8 | 32 | 77.4% | C | |
| AIME 2026 | 0.9 | 8 | 17 | 56.3% | C | |
| MCP-Mark | 0.4 | 8 | 8 | 0.0% | C | |
| WideSearch | 0.6 | 8 | 9 | 12.5% | C | |
| C-Eval | 0.9 | 9 | 18 | 52.9% | C | |
| HMMT25 | 0.9 | 9 | 25 | 66.7% | C | |
| LVBench | 0.7 | 9 | 24 | 65.2% | C | |
| HMMT Feb 26 | 0.8 | 10 | 11 | 10.0% | C | |
| SimpleVQA | 0.6 | 10 | 13 | 25.0% | C | |
| MMMU | 0.8 | 11 | 63 | 83.9% | C | |
| VideoMMMU | 0.8 | 11 | 26 | 60.0% | C | |
| MMLU-Redux | 0.9 | 12 | 48 | 76.6% | C | |
| SuperGPQA | 0.6 | 12 | 34 | 66.7% | C | |
| Claw-Eval | 0.5 | 13 | 13 | 0.0% | C | |
| NL2Repo | 0.3 | 14 | 14 | 0.0% | C | |
| HMMT 2025 | 0.9 | 16 | 33 | 53.1% | C | |
| IMO-AnswerBench | 0.8 | 16 | 19 | 16.7% | C | |
| MMLU-Pro | 0.9 | 16 | 129 | 88.3% | C | |
| LiveCodeBench v6 | 0.8 | 20 | 53 | 63.5% | C | |
| MCP Atlas | 0.6 | 22 | 30 | 27.6% | C | |
| SWE-bench Multilingual | 0.7 | 22 | 34 | 36.4% | C | |
| CharXiv-R | 0.8 | 24 | 47 | 50.0% | C | |
| Toolathlon | 0.3 | 31 | 31 | 0.0% | C | |
| MMMU-Pro | 0.8 | 32 | 65 | 51.6% | C | |
| Terminal-Bench 2.0 | 0.5 | 33 | 49 | 33.3% | C | |
| SWE-Bench Pro | 0.5 | 43 | 44 | 2.3% | C | |
| SWE-Bench Verified | 0.7 | 44 | 104 | 58.3% | C | |
| GPQA | 0.9 | 47 | 233 | 80.2% | C | |
| Humanity's Last Exam | 0.2 | 57 | 92 | 38.5% | C |
Preference and agent-evaluation signals from published Arena datasets.
The default version has no published Arena rows, or its source alias has not been resolved.
Official vendor API PAYG pricing is summarized first. The table then lists individual provider offerings without treating their minimum as the official price.
| QVAC | qwen3.6-35b-a3b | global | N/A | N/A | 262.1K | |
| Cortecs | qwen3.6-35b-a3b | global | $0.167 | $0.557 | 262K | |
| LLM Gateway | qwen3.6-35b-a3b | global | $0.248 | $1.49 | 262.1K | |
| Scaleway | qwen3.6-35b-a3b | global | $0.25 | $1.5 | 128K | |
| GreenPT | qwen3.6-35b-a3b | global | $0.342 | $2.05 | 262.1K |
Official prices use only the vendor's configured official Provider and positive standard USD PAYG rates. Third-party offers remain explicitly labeled.
All published versions linked to this unique model. The score columns identify the version used by the current overall ranking.
| Qwen3.6-35B-A3B | 48.2 | 35B | 262.1K | 65.5K | Yes | Apache 2.0 |
A concise description based on the published model registry.
Qwen3.6-35B-A3B is the first open-weight variant of the Qwen3.6 series, a multimodal Mixture-of-Experts model with 35B total parameters and 3B activated. It pairs a vision encoder with a hybrid 40-layer language model that interleaves Gated DeltaNet linear-attention blocks and Gated Attention blocks (10 × (3 × DeltaNet + 1 × Attention)) over 256 experts (8 routed + 1 shared, expert dim 512). The release prioritizes stability and real-world utility, with substantial gains in agentic coding (frontend workflows, repo-level reasoning) and a new option to preserve reasoning context across turns. Native context length is 262K tokens, extensible to ~1M via YaRN, and the model thinks by default.
Use the benchmark, Arena and pricing sections above as separate evidence. A missing field means the current data snapshot does not support that claim.
Data snapshot: 2026-08-07. Editorial model content is not available in the backend.
Open a comparison with the three ranked models immediately above and below this model.
Recommendations prioritize the same model type and family, then the closest published LLMBoard score.
Common questions about Qwen3.6 35B A3B.
Qwen3.6 35B A3B's default version was released on Apr 16, 2026.
No official standard PAYG price is currently available for Qwen3.6 35B A3B. The lowest tracked third-party offer starts at $0.167 input and $0.557 output via Cortecs.
Qwen3.6 35B A3B is published under Alibaba Cloud / Qwen Team in the model registry.
The default version has a 262.1K token context window.
Yes. The default version is marked as open weight under Apache 2.0.
5 published provider offerings are linked to the default version.
Nearby ranked alternatives include Gemini 1.5 Pro, Llama 4 Maverick, MiniMax M1 40K.