An RTX 3090 costs maybe $700 used. A single card with 24GB VRAM. And as of April 17, you can run a model on it that scores 73.4% on SWE-bench Verified — the industry's hardest real-world software engineering benchmark — at 50 to 65 tokens per second.
That's the headline with Alibaba's Qwen3.6-35B-A3B. Not that it exists, but that it runs there.
What Is Qwen3.6-35B-A3B?
The name breaks down like this: 35 billion total parameters, with roughly 3 billion activated per token. The "A3B" is their shorthand for "approximately 3B active." It's a sparse Mixture-of-Experts architecture — 256 total experts baked into the weights, with only 8 routed experts plus one shared expert firing on any given forward pass.
That design is the whole story. The model stores 35B worth of knowledge and capability, but at inference time it's doing the compute work of a much smaller model. Memory bandwidth — the actual bottleneck on consumer hardware — gets slashed dramatically compared to a dense 35B model.
The architecture itself is interesting: Alibaba combined Gated DeltaNet linear attention with traditional self-attention in a hybrid layout, something their team has been iterating on across the Qwen3.x line. Whether that specific combination matters in practice is hard to isolate. What matters is the output: it works, it's fast, and it fits on hardware regular developers actually own.
Context window is 262,144 tokens natively — enough to hold a large codebase in a single prompt. With YaRN scaling enabled, you can push that to about one million tokens, though you'll want multi-GPU for that kind of workload. The model supports both thinking and non-thinking modes. For agentic tasks specifically, Alibaba recommends keeping thinking enabled with their preserve_thinking flag, which maintains the chain-of-thought across multi-step tool calls.
Why Agentic Coding Specifically
If you're running Cursor or building your own coding agent, the frontier model question usually comes down to one of three choices: pay OpenAI/Anthropic API rates, accept latency on remote inference, or compromise on a smaller open-weight model that doesn't quite cut it on complex tasks.
Qwen3.6-35B-A3B is a serious challenge to that third option being a compromise.
On SWE-bench Verified, it posts 73.4%. That benchmark asks models to actually fix real GitHub issues — not toy problems, not curated examples, real production bugs from open-source repos. For context: models scoring in the low 40s were considered impressive a year ago. 73.4% puts this firmly in the "useful for real work" tier.
Terminal-Bench 2.0 — which evaluates multi-step terminal and shell workflows — lands at 51.5%. That's the kind of score that matters if you're building agents that touch CI/CD pipelines, run test suites, or interact with cloud tooling.
The MoE architecture makes local agentic workflows specifically more viable because agentic pipelines are chatty. You're making dozens or hundreds of model calls per task — tool calls, reasoning steps, verification passes. On a per-call basis, latency compounds. Running Q4_K_M quantized on an RTX 4090, you're looking at 75 to 95 tokens per second. A dense model of comparable quality would require significantly more VRAM and run slower.
How It Compares
Against closed frontier models — GPT-5.4, Claude Opus 4.7 — Qwen3.6-35B-A3B doesn't win on raw benchmark scores. Those models occupy a different tier. GPT-5.4 and Opus 4.7 are both better; that's not a knock, that's an accurate statement of where the ceiling is right now.
But the question isn't "does it beat GPT-5.4?" The question is "does it close the gap enough to make local deployment worth it?"
On knowledge benchmarks, it's strong: 85.2% on MMLU-Pro, 93.3% on MMLU-Redux, 86.0% on GPQA. On reasoning, 92.7% on AIME 2026. These are not boutique scores. For developer tasks that are primarily about code understanding and generation rather than cutting-edge reasoning, the gap to frontier closed models is meaningful but not disqualifying.
Against the open-weight field — Llama 4, Mistral Large, earlier Qwen releases — it's in a different class. It clearly leads on agentic coding benchmarks, and that lead isn't marginal. If you're comparing AI models for developer workflows in 2026, Qwen3.6-35B-A3B is the new reference point for what open weights can do.
One multimodal note: it handles text and vision, not just text. That's newer territory for the Qwen line at this weight class. Don't know yet how much that matters in coding agent workflows, but it's there.
Can You Actually Run It?
Yes, and that's the part I keep coming back to. The quantization math:
- Q4_K_M (recommended): ~23GB VRAM. Fits a single RTX 3090 or 4090 (both 24GB). Runs at 50-65 tok/s on a 3090, 75-95 on a 4090.
- Q8_0: ~40GB VRAM. You'll need a 48GB card (A6000, RTX 6000 Ada) or two consumer cards with NVLink.
- IQ3_XS (if you're memory-constrained): ~17GB. Fits older 20-series cards. Expect quality tradeoffs.
Ollama has it available now — ollama pull qwen:Qwen3.6-35B-A3B — and it works with LM Studio and llama.cpp via GGUF. If you're running vLLM or SGLang for production-grade inference, both are supported as of recent versions.
The full Apache 2.0 weights are on HuggingFace with over 220 quantized variants already up. This is about as accessible as a 35B model has ever been.
One honest note on the "runs on a 3090" claim: it's true at Q4_K_M quantization with a moderate context. If you're maxing out the 262K context window, you're going to need more headroom. For typical agentic coding tasks with reasonable context sizes, a single 24GB consumer GPU handles it fine.
Is This a Breakthrough?
For local open-weight deployment, yes. For frontier benchmark leadership, no.
What Alibaba shipped here is the most capable model that runs on hardware a developer can actually buy. The gap to GPT-5.4 and Claude Opus 4.7 is real, but the gap to the next-best locally-runnable option is also real — and it goes the other way.
If you're building coding agents and you want to keep inference local — for latency, cost, privacy, or just control — Qwen3.6-35B-A3B is the clearest answer right now. It doesn't require a server cluster. It doesn't require a cloud API key. It requires a reasonably recent consumer GPU and an afternoon to get your stack set up.
That's genuinely new. A year ago, "local frontier-class coding model" wasn't a coherent phrase. Now it is.
Qwen3.6-35B-A3B is available under Apache 2.0 on HuggingFace. Official Alibaba launch post at qwen.ai.
Top comments (0)