Something is afoot in the land of Qwen
Qwen’s latest open-weight models, especially Qwen3.5, are impressing many developers for local, “agentic” coding and tooling despite still lagging top commercial models like Claude and GPT-4. The thread centers on reports that key researchers behind Qwen have left Alibaba, raising concerns about the future of one of the strongest open model families and prompting speculation about internal politics, open-vs-closed strategies, and potential poaching by Western labs. Participants also debate the economics and geopolitics of Chinese open models, including how they’re funded, how they compete with US proprietary offerings, and whether researchers will prefer to stay in China given US immigration and policy risks.
Qwen team shake-up and project direction
- Multiple commenters report that key Qwen researchers have left after internal tensions with the parent company, possibly over KPIs (e.g., DAU for the Qwen app) and product vs. research priorities.
- Some speculate about demotion, power struggles, or a shift toward closed, proprietary models, but specific causes remain unclear.
- Many see this as a major loss for the open/local LLM ecosystem, given Qwen’s recent progress.
Model capabilities and comparisons
- Qwen3.5 models, especially 35B-A3B and 27B, are widely praised as state-of-the-art among local/open weights, with strong coding, planning, and tool use for their size.
- Experiences vary: some find Qwen3.5-35B-A3B better than Qwen3-Coder-Next, others the reverse, often attributing differences to model size, quantization quality, chat template, and serving stack.
- Compared to frontier cloud models (Claude, Gemini, etc.), Qwen is still seen as roughly “a year behind,” but impressively close for self-hosted use.
Agentic coding, harnesses, and behavior
- Harness / orchestrator quality (e.g., Zed’s agentic features, Pi-style minimal setups, Qwen’s own harness, Antigravity, OpenCode) strongly affects outcomes.
- Tools often go unused unless the system prompt clearly defines them; explicit tool descriptions and formats significantly improve behavior.
- Users report both tenacious problem-solving and frustrating looping, shortcutting, or ignoring instructions mid-task. Lower temperature helps but temperature=0 can be counterproductive.
Performance, hardware, and quantization
- People share practical setups: consumer GPUs (3070 Ti, 5080, AMD AI Max), Macs, large-RAM CPUs, and various 4–6 bit quants via llama.cpp / vLLM.
- Token speeds in the ~20–70 tok/s range are common; context length and quant choice heavily impact tool-calling reliability and looping.
- Small Qwen3.5 models (0.8–9B) are noted as surprisingly capable for OCR and vision, but weaker for complex coding and coherent prose.
Geopolitics, talent, and economics
- Long subthreads debate why top Chinese researchers might stay in or return to China vs. joining US or EU labs, citing nationalism, quality of life, immigration enforcement, and government expectations.
- Some view Chinese open releases (Qwen, GLM, Kimi) as strategically subsidized to pressure US proprietary vendors.
- Business sustainability is questioned: training is costly, while models are released for free; suggested motives include VC funding, hosted inference revenue, and national strategic goals.
Vendor conflicts and “distillation”
- Discussion of Anthropic’s complaints centers on use of Claude as “LLM-as-a-judge” and for generating training data.
- Commenters argue this is more like RL with model-based feedback than true weight-level distillation, and note that similar cross-model training behavior is widespread.