A Tongyi Qianwen (Qwen) developer publicly told Reddit users "don't wait for the 35B-A3B model" — one sentence that set the overseas open-source community on fire.

What this is

35B-A3B is Qwen's planned "mid-size" model: 35 billion total parameters, with only ~3 billion activated per inference — that's the Mixture of Experts (MoE) architecture (large model capacity, cheap per-query inference). Developers want it because, compared to 70B+ heavyweights, a mid-size MoE like this runs on a single consumer-grade GPU — exactly what small and mid-size teams need for local knowledge bases and private deployments.

One "don't wait" could mean three things: they're holding for something bigger (like 122B), the size is being cut entirely, or resources have been redirected elsewhere.

Industry view

Reddit split into two camps.

Optimists read it as "good stuff is coming." Qwen is the steadiest cadence among China's open-source players — from 7B, 14B, 72B through Qwen3, the positioning has been sharp. If 35B-A3B slips, the team may be pushing for a more competitive size, like a 122B-class mid-weight flagship.

The other reading: Alibaba is pulling back on open-source investment. Earlier this year, the Qwen team was merged into Alibaba Cloud Intelligence, and the industry talk is "fighting wars while slimming down." The open-source update cadence has visibly shifted from "quarterly major releases" to "weekly small steps." A developer who tracks Chinese models long-term told us: "It's not that they're not shipping — what they ship increasingly looks like 'good enough,' totally different cadence from DeepSeek."

The third possibility deserves more of our attention: an overall pivot in open-source strategy — treating open source as a top-of-funnel acquisition channel rather than a commitment, holding flagships for paid cloud users, open-sourcing the mid-tier with no cadence guarantees. For small teams using Qwen as their base, that's a structural risk.

Impact on regular people

  • For enterprise IT: Teams that planned to swap 35B-A3B in for closed-source APIs to cut costs on private deployments now have to replan.
  • For working professionals: The vast majority of daily AI users won't feel this — the volatility only hits the geeks and engineers who run models themselves.
  • For the consumer market: Indirectly slows the capability upgrade cycle for domestic AI in education, customer service, and marketing tools.