An unreleased 35B Qwen model is generating a waitlist in Reddit's local AI community — and the reaction tells us something real: China's open-source LLMs are accelerating on the "small parameters, high performance" path.
The "35B A3B" naming points to a MoE architecture (Mixture of Experts — 35 billion total parameters, only ~3 billion activated at inference), which keeps per-call compute low and VRAM footprint an order of magnitude below a comparably sized dense model. The "3.8" version number has no official confirmation — it remains community speculation — but the Qwen3 series has already shown a clear lightweight product line.
What this is
Qwen is Alibaba Cloud's open-source LLM family, iterated continuously since 2023. On Hugging Face it has long held the #1 spot among Chinese models by developer downloads, and it sits stably in the global top five open-source models. A 35B-class MoE is the natural next step on the "performance-cost balance" line.
Local AI practitioners care about this model for one reason: it can actually run. If the 35B MoE lands, a developer with a 4-GPU consumer server or a single Mac Studio could deploy something close to GPT-4 class — at monthly cost potentially one-fifth to one-tenth of closed-source API pricing.
Industry view
The bull case: China's open-source models genuinely iterate faster than overseas counterparts on the "small but strong" track. MoE gives small teams, for the first time, a "real deployment option that actually runs on-prem." Open weights also enable private fine-tuning to meet compliance requirements in finance and healthcare.
The bear case and the risks we should flag come down to three points. First, Reddit waitlist energy is not the same as real market demand — it may just be the eternal "I'll wait for the next one" crowd expressing itself. Second, MoE models impose specific requirements on inference frameworks and VRAM management; real-world experience may not beat a dense model at the same active-parameter count. Third, the capability gap between Alibaba Cloud's commercial and open-source versions keeps widening, so developers need to manage their own expectations on what they actually get.
Impact on regular people
For enterprise IT: Companies evaluating "build vs. buy" for in-house AI can run a 35B-class open-source model on a single 4-GPU server. The budget bar is lower than most expect — but you still need a dedicated ML ops team, and we'd recommend a 3–6 month POC before commitment.
For individual careers: People willing to tinker can now run near-GPT-4-class models on their own PC or Mac, free of cloud subscriptions. The learning curve remains steep, though — non-technical professionals should think twice.
For the consumer market: No direct impact yet, but the broader trend of AI capability migrating to local devices (PC, phone, car) is accelerating. Within the next 12–18 months, we'd expect "offline-capable" AI endpoint products to hit shelves.