Alibaba Just Dropped Qwen 3.5 Medium — The 35B Model That Outruns Giants

It feels like we’re in the middle of a weight-class revolution. Just a week after Alibaba dropped their massive 397B flagship, they followed up yesterday with the ‘Medium’ series, and honestly? The numbers are more impressive than the titan.

The 35B Giant Slayer

The standout of the release is Qwen3.5-35B-A3B. It’s a sparse Mixture-of-Experts (MoE) model with only 3 billion active parameters per token. Here’s the kicker: it’s seven times smaller than the previous flagship Qwen3-235B, yet it’s outperforming it on major benchmarks. We’re talking about a model that can run on consumer hardware while delivering frontier-level reasoning.

I’ve been watching the LocalLLaMA community have a collective meltdown over this. The ‘intelligence-per-watt’ here is setting a new floor for what we should expect from open weights. Imagine running something smarter than GPT-4o-mini locally on your RTX 4090. That’s the promise here.

Efficiency over Scaling

Alibaba is clearly pivoting. Instead of just throwing more compute at the wall, they’re leaning into architectural efficiency. The series also includes:

• Qwen3.5-122B-A10B: An MoE model that matches the 397B flagship and even edges it out on specific reasoning tasks like HLE. Ideally suited for high-throughput inference where you need massive intelligence but can’t afford the latency of a 400B+ model.
• Qwen3.5-27B: A dense model that is a monster at instruction following, hitting 95.0% on IFEval. This is the workhorse for code generation and strict format adherence.

Why This Matters

We’re moving into the ‘agentic era’ where context and tool use matter more than raw parameter counts. With the Flash variant supporting a 1 million token context window and native tool use, Alibaba isn’t just releasing models; they’re releasing a runtime for agents.

Think about it: agents don’t just need to be smart once; they need to loop, reason, and tool-call hundreds of times per task. A massive 400B model is too slow and expensive for that loop. But a 35B model with 3B active params? That’s fast enough to think in real-time. If early 2027 is really when we see recursive self-improvement—as Anthropic’s new roadmap suggests—these efficient ‘medium’ models are exactly the kind of scaffolding that will get us there.

The Community Verdict

Early tests are showing the 35B model handles nuanced prompts surprisingly well. It’s not perfect—it still hallucinates on obscure facts compared to the 70B+ class—but for reasoning and coding, it feels like a generational leap for its size. The 122B is being hailed as the new open-source king for those with the VRAM to run it (likely needing 2x3090s or similar).

My take? The era of ‘bigger is better’ is officially on life support. The future belongs to the lean, the fast, and the agentic. And right now, Qwen is leading that charge.