Overview
Qwen3.6-35B-A3B represents a landmark release in efficient open-source AI. As the first open-weight model from the Qwen3.6 family, it employs a Mixture-of-Experts architecture with 35 billion total parameters but only activates approximately 3 billion per token—delivering inference costs and speeds comparable to a 3B dense model while achieving benchmark parity with 27B+ dense architectures. Released April 16, 2026 under Apache 2.0, it is available through Hugging Face, ModelScope, Alibaba Cloud Model Studio (as qwen3.6-flash), and the Qwen Studio playground.
The model's defining strength is agentic coding. On Terminal-Bench 2.0, it scores 51.5—a 27% improvement over its predecessor Qwen3.5-35B-A3B (40.5) and surpassing the much larger Gemma4-31B (42.9). SWE-bench Verified reaches 73.4%, nearly matching the 27B dense model's 75.0% while operating at a fraction of the compute. The QwenWebBench Elo rating surged from 978 to 1397, reflecting substantial gains in front-end code generation. Beyond coding, the model demonstrates remarkable multimodal capability: RealWorldQA (85.3) and OmniDocBench (89.9) both exceed Claude Sonnet 4.5, and spatial intelligence metrics like RefCOCO (92.0) and ODInW13 (50.8) set new benchmarks for models of this scale.
Architecturally, Qwen3.6-35B-A3B uses a 256-expert sparse MoE design where a router selects 9 experts per token (8 routed + 1 shared). This decouples model capacity from compute cost, allowing the 35B parameter storage to encode far more knowledge than a 3B dense model could hold. The model supports both 'thinking' (chain-of-thought reasoning) and 'non-thinking' (direct response) modes, with a critical 'preserve_thinking' feature that maintains reasoning context across multi-turn agentic workflows. Deployment is supported via vLLM, SGLang, and Transformers, with FP8 quantization reducing VRAM requirements to approximately 10GB.
Benchmarks & Performance
## Comprehensive Benchmark Performance
### Agentic Coding (Core Strength)
| Benchmark | Qwen3.6-35B-A3B | Qwen3.5-27B | Qwen3.5-35B-A3B | Gemma4-31B | Qwen3.6-Plus |
|---|---|---|---|---|---|
| SWE-bench Verified | 73.4% | 75.0% | 70.0% | 52.0% | 78.8% |
| SWE-bench Multilingual | 67.2% | 69.3% | 60.3% | 51.7% | 73.8% |
| SWE-bench Pro | 49.5% | 51.2% | 44.6% | 35.7% | 56.6% |
| Terminal-Bench 2.0 | 51.5 | 41.6 | 40.5 | 42.9 | 61.6 |
| Claw-Eval Avg | 68.7 | 64.3 | 65.4 | 48.5 | 74.8 |
| NL2Repo | 29.4 | 27.3 | 20.5 | 15.5 | 37.9 |
| QwenWebBench (Elo) | 1397 | 1068 | 978 | 1197 | 1501.7 |
Key finding: Terminal-Bench 2.0 shows the largest leap (+27.2% over predecessor), demonstrating genuine improvement in real-world terminal-based agentic workflows. SWE-bench Verified (73.4%) approaches the 27B dense model (75.0%) with only 3B active parameters.
### Knowledge & Reasoning
| Benchmark | Qwen3.6-35B-A3B | Qwen3.5-27B | Gemma4-31B | Qwen3.6-Plus |
|---|---|---|---|---|
| MMLU-Pro | 85.2 | 86.1 | 85.2 | 88.5 |
| MMLU-Redux | 93.3 | 93.2 | 93.7 | 94.5 |
| SuperGPQA | 64.7 | 65.6 | 65.7 | 71.6 |
| C-Eval (Chinese) | 90.0 | 90.5 | 82.6 | 93.3 |
| GPQA Diamond | 86.0 | 85.5 | 84.3 | 90.4 |
### STEM & Mathematics
| Benchmark | Qwen3.6-35B-A3B | Qwen3.5-27B | Gemma4-31B | Qwen3.6-Plus |
|---|---|---|---|---|
| AIME 2026 (full) | 92.7 | 92.6 | 89.2 | 95.3 |
| HMMT Feb 2026 | 83.6 | 84.3 | 77.2 | 87.8 |
| HMMT Feb 2025 | 90.7 | 92.0 | 88.7 | 96.7 |
| IMOAnswerBench | 78.9 | 79.9 | 74.5 | 83.8 |
| LiveCodeBench v6 | 80.4 | 80.7 | 80.0 | 87.1 |
| HLE | 21.4 | 24.3 | 19.5 | 28.8 |
### General Agent & Tool Use
| Benchmark | Qwen3.6-35B-A3B | Qwen3.5-27B | Gemma4-31B | Qwen3.6-Plus |
|---|---|---|---|---|
| τ³-Bench | 67.2 | 68.4 | 67.5 | 70.7 |
| MCPMark | 37.0 | 36.3 | 18.1 | 48.2 |
| DeepPlanning | 25.9 | 22.6 | 24.0 | 41.5 |
| Tool Decathlon | 26.9 | 31.5 | 21.2 | 39.8 |
| WideSearch | 60.1 | 66.4 | 35.2 | 74.3 |
### Multimodal (Vision & Video)
| Benchmark | Qwen3.6-35B-A3B | Claude Sonnet 4.5 | Qwen3.5-27B | Qwen3.6-Plus |
|---|---|---|---|---|
| MMMU | 81.7 | 79.6 | 82.3 | 86.0 |
| RealWorldQA | 85.3 | 70.3 | 83.7 | 85.4 |
| OmniDocBench 1.5 | 89.9 | 85.8 | 88.9 | 91.2 |
| MMBench EN-DEV | 92.8 | 88.3 | 92.6 | — |
| RefCOCO (avg) | 92.0 | — | 90.9 | 93.5 |
| ODInW13 | 50.8 | — | 41.1 | 51.8 |
| VideoMMMU | 83.7 | 77.6 | 82.3 | 84.0 |
| VideoMME (w/ sub.) | 86.6 | 81.1 | 87.0 | 87.8 |
Key multimodal finding: Despite having only 3B active parameters, Qwen3.6-35B-A3B matches or exceeds Claude Sonnet 4.5 on nearly all vision benchmarks, and achieves 97% average performance parity with the much larger Qwen3.6-Plus on multimodal tasks.
Detailed Comparison
## Head-to-Head Comparisons
### vs Qwen3.5-27B (Dense, Open-Source)
The closest internal comparison. Both are open-source under Apache 2.0, but the architecture difference is fundamental:
| Dimension | Qwen3.6-35B-A3B | Qwen3.5-27B | Winner |
|---|---|---|---|
| Active Parameters | 3B | 27B | — |
| SWE-bench Verified | 73.4% | 75.0% | 27B (+1.6) |
| Terminal-Bench 2.0 | 51.5 | 41.6 | 35B-A3B (+23.8%) |
| GPQA Diamond | 86.0 | 85.5 | 35B-A3B (+0.6) |
| AIME 2026 | 92.7 | 92.6 | Tie |
| Inference Speed (Q4, RTX 4090) | ~100-122 tok/s | ~32-40 tok/s | 35B-A3B (~3×) |
| VRAM (FP16) | ~18GB | ~54GB | 35B-A3B (3× less) |
| Context Window | 128K–262K | 128K–262K | Tie |
**Verdict:** Qwen3.6-35B-A3B is the superior choice for speed-sensitive agentic workflows and consumer hardware deployment. Qwen3.5-27B retains a marginal edge in complex multi-step planning and sustained coding quality where the full 27B parameter utilization provides deeper reasoning.
### vs Gemma4-31B (Google, Dense, Open-Source)
| Dimension | Qwen3.6-35B-A3B | Gemma4-31B | Winner |
|---|---|---|---|
| Active Parameters | 3B | 31B | — |
| SWE-bench Verified | 73.4% | 52.0% | 35B-A3B (+41%) |
| Terminal-Bench 2.0 | 51.5 | 42.9 | 35B-A3B (+20%) |
| QwenWebBench (Elo) | 1397 | 1197 | 35B-A3B (+200 Elo) |
| NL2Repo | 29.4 | 15.5 | 35B-A3B (+89.7%) |
| GPQA Diamond | 86.0 | 84.3 | 35B-A3B |
| MMLU-Pro | 85.2 | 85.2 | Tie |
**Verdict:** Qwen3.6-35B-A3B dominates Gemma4-31B on coding and agentic tasks with dramatically less compute. The gap on general knowledge benchmarks is minimal. This comparison powerfully validates the MoE architecture's efficiency advantages.
### vs Qwen3.6-Plus (Alibaba, Closed-Source API)
| Dimension | Qwen3.6-35B-A3B | Qwen3.6-Plus | Gap (Achievement Rate) |
|---|---|---|---|
| SWE-bench Verified | 73.4% | 78.8% | 93.1% |
| Terminal-Bench 2.0 | 51.5 | 61.6 | 83.6% |
| SkillsBench Avg | 28.7 | 45.7 | 62.8% |
| GPQA Diamond | 86.0 | 90.4 | 95.1% |
| MMLU-Pro | 85.2 | 88.5 | 96.3% |
| AIME 2026 | 92.7 | 95.3 | 97.3% |
| DeepPlanning | 25.9 | 41.5 | 62.4% |
| Multimodal Avg | ~97% parity | — | — |
**Verdict:** The 35B-A3B achieves 85–97% of Plus capability on most tasks, but the gap widens to ~37% on deep planning and practical skills. For budget-conscious teams, a hybrid strategy—using 35B-A3B for 80% of tasks and escalating to Plus for complex planning—optimizes cost and quality.
### Deployment & Cost Summary
| Model | VRAM (FP16) | VRAM (Q4) | Speed (RTX 4090) | API Availability |
|---|---|---|---|---|
| Qwen3.6-35B-A3B | ~18GB | ~10GB | 100-122 tok/s | qwen3.6-flash |
| Qwen3.5-27B | ~54GB | ~17GB | 32-40 tok/s | qwen3.5-turbo |
| Gemma4-31B | ~62GB | ~19GB | ~45 tok/s | Open |
| Qwen3.6-Plus | N/A | N/A | N/A | qwen3.6-plus |
| Claude Sonnet 4.5 | N/A | N/A | N/A | API-only |
Community Feedback
Developer and researcher reception has been overwhelmingly positive, with several recurring themes emerging:
**Efficiency benchmarks reset expectations.** Multiple Chinese tech commentators have highlighted the Terminal-Bench 2.0 leap from 40.5 to 51.5 as a 'generational jump' rather than incremental improvement. The YOMXXX review noted that '3B active parameters beating 22B dense models on coding benchmarks validates MoE as the optimal architecture for open-source models.' The Alibaba Cloud Community blog post emphasized that 'Qwen3.6-35B-A3B demonstrates sparse MoE models can achieve remarkable agentic coding and reasoning capability.'
**Local deployment enthusiasm is high.** The CSDN comparative review between 27B dense and 35B-A3B MoE generated significant discussion. Users on RTX 4090 report 100-122 tok/s inference speeds (vs ~35 tok/s for the dense 27B), making real-time interactive coding assistance feasible on consumer hardware. One user noted: '35B completed tasks in half the time, but the results didn't reflect 27B's depth of understanding.' Another countered: 'I need fast responses for routine queries—35B is perfectly sufficient.'
**Agentic coding integration drove immediate adoption.** The model's compatibility with Claude Code via Anthropic protocol, OpenClaw, and Qwen Code was highlighted as a key differentiator. DataLearner reported that Qwen3.6-Plus (the closed-source sibling) broke OpenRouter's single-day record with 1.4 trillion tokens on launch day, suggesting strong demand for the Qwen3.6 family. The 35B-A3B's open-source release is expected to capture a significant portion of the self-hosted agent coding market.
**Multimodal performance surprised many.** Several reviewers noted surprise that a 3B-active model matches Claude Sonnet 4.5 on vision tasks. The cnblogs review called it 'the most unexpected part,' noting RealWorldQA 85.3 vs Claude's 70.3 as particularly striking. ODInW13's jump from 42.6 to 50.8 was cited as evidence of meaningful spatial intelligence advances.
**Concerns about general agent tasks persist.** While coding benchmarks impress, community members noted that DeepPlanning (25.9) and Tool Decathlon (26.9) scores suggest the model's MoE routing may struggle with complex multi-step tool orchestration requiring sustained cross-domain reasoning. The Lixx Blog comparison concluded that 'perception tasks are nearly at flagship level, but deep reasoning and long-horizon planning still require larger models.'
**Context window and thinking preservation praised.** The `preserve_thinking` feature for maintaining reasoning chains across multi-turn agent sessions received specific developer praise. Combined with 262K context support (for self-hosted deployments), this enables 'feeding an entire mid-size code repository in one pass'—described as 'essential for long-horizon coding tasks.'
Use Cases
### 1. Agentic Coding & IDE Integration
**When to choose:** Daily software development workflows involving bug fixing, code generation, refactoring, and multi-file editing. This is Qwen3.6-35B-A3B's strongest domain.
**Example:** A developer connects the model to Claude Code or OpenClaw via the Anthropic API protocol. The model reads an issue description, navigates the repository structure, identifies relevant files across multiple directories, generates patches, and runs test suites—all within a single session. With Terminal-Bench 2.0 score of 51.5 (surpassing models 10× its active parameter count) and SWE-bench Verified at 73.4%, it handles real GitHub issue resolution reliably.
**Why over alternatives:** At 100-122 tok/s on RTX 4090 (Q4 quantization), responses arrive 3× faster than the 27B dense alternative. For teams making hundreds of coding queries daily, this translates to hours of saved wait time. The `preserve_thinking` feature maintains reasoning context across iterative debugging sessions, reducing redundant re-analysis.
### 2. Multimodal Document & Video Analysis
**When to choose:** Processing visual documents, OCR tasks, chart analysis, video understanding, and spatial reasoning at scale with cost constraints.
**Example:** An enterprise deploys the model to extract structured data from invoices, contracts, and technical diagrams (OmniDocBench 89.9). Simultaneously, it analyzes security camera footage for spatial events (RefCOCO 92.0, ODInW13 50.8). Video understanding capabilities (VideoMMMU 83.7) enable automated video summarization and content moderation.
**Why over alternatives:** RealWorldQA score of 85.3 exceeds Claude Sonnet 4.5 (70.3) at a fraction of the cost. For organizations processing thousands of documents daily via API, the qwen3.6-flash pricing tier offers dramatic savings compared to proprietary alternatives. Local deployment ensures sensitive documents never leave the premises.
### 3. Local/Private Deployment for Data-Sensitive Environments
**When to choose:** Healthcare, finance, legal, or government environments where data cannot leave the organization's infrastructure.
**Example:** A hospital deploys Qwen3.6-35B-A3B on a single RTX 4090 (Q4 quantization, ~10GB VRAM) behind their firewall. Clinicians use it for medical document understanding (OmniDocBench 89.9), radiology image analysis (MMMU 81.7), and clinical note summarization—all while maintaining full HIPAA compliance. The Apache 2.0 license permits unlimited commercial use without royalties.
**Why over alternatives:** The 35B MoE model fits in 10GB VRAM quantized, versus 17GB+ for comparable dense models. This enables deployment on existing workstation GPUs without additional hardware procurement. GPQA Diamond (86.0) demonstrates expert-level knowledge retrieval suitable for professional domains.
### 4. Cost-Optimized High-Volume API Workflows
**When to choose:** Startups and teams running large-scale AI-powered services (chatbots, content generation, automated code review) where API costs are a primary concern.
**Example:** A SaaS company processes 10,000+ code review requests daily through the qwen3.6-flash API. For routine reviews (80% of traffic), the 35B-A3B model handles them at flash-tier pricing. For complex architectural decisions or multi-repo refactoring (20% of traffic), requests are escalated to Qwen3.6-Plus. This hybrid strategy achieves ~87% of flagship quality at approximately one-third the cost.
**Why over alternatives:** Compared to Qwen3.6-Plus, the 35B-A3B achieves 93%+ performance on standard coding tasks (SWE-bench Verified, QwenClawBench) at significantly lower per-token pricing. Compared to dense open-source alternatives, the MoE architecture delivers 2-3× faster inference, reducing time-to-first-token for interactive applications.
Latest News
### April 2026 Releases
- **April 16, 2026:** Qwen3.6-35B-A3B officially open-sourced with Apache 2.0 license. Available on Hugging Face (Qwen/Qwen3.6-35B-A3B), ModelScope, and Qwen Studio (chat.qwen.ai). This is the first open-weight model in the Qwen3.6 family, following the closed-source Qwen3.6-Plus which launched via API on March 30, 2026.
- **API Access:** The model is available through Alibaba Cloud Model Studio as `qwen3.6-flash`, supporting both OpenAI-compatible (chat completions/responses) and Anthropic-compatible API protocols. The `preserve_thinking` feature was introduced for maintaining reasoning context across agentic multi-turn sessions.
- **Integration Partners:** Immediate compatibility announced with three coding agent frameworks: OpenClaw (formerly Moltbot/Clawdbot), Qwen Code (Qwen-optimized terminal agent), and Claude Code (via Anthropic API protocol compatibility).
### Ecosystem Expansion
- **Qwen3.6-27B followed on April 22-24, 2026** as a dense 27B model targeting 'flagship-level coding in a 27B dense model,' expanding the Qwen3.6 open-source family. Official benchmarks show Qwen3.6-27B SWE-bench Verified at 77.2%, Terminal-Bench 2.0 at 59.3%.
- **Deployment Framework Support:** Day-0 support for vLLM (≥0.19.0 recommended), SGLang with multi-token prediction (MTP) acceleration, and Transformers. AMD Instinct MI300X/MI325X/MI350X/MI355X GPU support confirmed with ROCm integration.
- **Quantized Variants:** FP8 (GPTQ) quantization available, reducing VRAM requirements to ~10GB. GGUF 4-bit quantization through Unsloth enables deployment on 24GB Mac M3 Max systems.
### Personnel Context
- The Qwen team experienced significant leadership changes in early 2026, with core Qwen3-Max and Qwen3.5 series lead Lin Junyang departing Alibaba alongside several senior team members. Qwen3.6-35B-A3B represents the first open-source release under the restructured team.
- **Looking Ahead:** Alibaba has indicated the Qwen3.6 open-source family will continue expanding with additional model sizes and variants.