Qwen 4 Coder: Alibaba's Open-Weight MoE Model Hits 82% on SWE-Bench Verified

Alibaba Group has reportedly pushed the boundaries again with the launch of Qwen 4 Coder 32B – an open-weights model that’s making waves in software engineering capabilities. Key Features & Highlights: Record Performance: Achieves 82% on SWE-Bench Verified – the highest score published for any open-source model. This benchmark tests real-world bug fixing on actual GitHub production repositories. Fully Open & Accessible: Apache 2.0 licence with free weights on Hugging Face. Full commercial use, no attribution needed, no usage restrictions. Efficient Architecture: Dense 32B parameter model (not MoE) that runs on a single high-end consumer GPU (~22GB VRAM) or even fits on a MacBook Pro with 64GB unified memory. Lightning-Fast Inference: Smaller 4B variant delivers ~135 tokens/second on M5 Max chips and runs in under 8GB RAM – bringing powerful coding assistance to laptops. Training Edge: Reinforced with verifiable coding tasks – rewarded for code that actually passes tests, resulting in more reliable, production-ready outputs. Easy Deployment: Compatible with vLLM, Ollama, SGLang, and GGUF engines. This marks a significant step where open-source models are not just catching up but outperforming some leading closed-source options on practical software engineering benchmarks. For developers, indie hackers, and small teams, this could dramatically lower barriers to AI-powered coding. https://qwen.ai/qwencode https://lnkd.in/guZsy5QT #AI #OpenSource #Coding #LLM #SoftwareEngineering #AlibabaQwen

I checked this against multiple independent sources on Qwen 4 Coder.

Confirmed accurate:

• Qwen 4 Coder 32B-A3B does score approximately 82% on SWE-Verified, and is reported as the first open-weight, Mac-runnable model to clear 80% on that benchmark — the headline number checks out. Llmcheck
• The model, released June 2, 2026, does ship under the Apache 2.0 license — confirmed, and Apache 2.0 does mean free commercial use with no attribution requirement. Mysummit
• It's compatible with local inference tooling and can be run via Ollama and MLX, consistent with the post's deployment claims (vLLM, Ollama, SGLang, GGUF). Vibe Engines
• The MacBook-runnable framing checks out: it fits comfortably on Macs with 36GB+ unified memory, running around 58-65 tokens/sec on M4 Pro through M5 Max chips at Q4 quantization. Llmcheck

Inaccurate — this is the important one:

• The post's claim that this is a "Dense 32B parameter model (not MoE)" is wrong. Every independent source I found describes it as the opposite: a 32B-total / 3B-active Apache-2.0 Mixture-of-Experts model, and the MoE architecture is explicitly why it only activates 3B parameters per token, which is what makes its inference speed so favorable for its quality tier. The model is commonly named "32B-A3B" specifically to signal this MoE design (32B total, 3B active). The post asserts the exact opposite of the real architecture. Vibe EnginesLlmcheck
• Because it's MoE with 3B active parameters rather than a dense 32B model, the "~22GB VRAM on a single high-end consumer GPU" framing is roughly right in terms of storage footprint, but the reasoning the post gives (implying you need to compute through all 32B parameters) misunderstands why it runs efficiently — it's fast specifically because only 3B parameters activate per token, not despite being dense.

Unverified:

• I couldn't independently confirm the specific claim of "a 4B variant delivering ~135 tokens/second on M5 Max chips in under 8GB RAM." I found comparable tokens/sec figures for other small models (e.g., Phi-4 Mini at ~135 tok/s) and for the 32B-A3B variant itself (58-65 tok/s), but no source specifically confirming a "Qwen 4 Coder 4B" variant at that exact speed and memory footprint. This may be a real but under-documented variant, or a mix-up with a different small model.
• One early social media thread flagged the initial release as "not verified" and asked for a source, though multiple independent review sites subsequently covered it as a real, dated release (June 2, 2026), so this looks like early-hours skepticism that was later resolved rather than an ongoing controversy.