Qwen3.5-35B-A3B: 3B Active Parameters, Medium-Model Scores

Most model comparisons ask which model is best. For high-volume work the better question is which model is good enough at the lowest cost per completed task. Qwen3.5-35B-A3B is built for that question. It activates about 3B parameters per token, roughly a ninth of what the dense Qwen3.5-27B runs, and on a large share of Qwen's published benchmarks it lands within a few points of both larger siblings.

It was released on 24 February 2026 alongside the 27B and the 122B-A10B, under Apache-2.0, as a native vision-language model that accepts text, images and video.

What 3B active means in practice

The model has 35B parameters in total, spread across 40 layers. Each layer's feed-forward block is a mixture of 256 small experts, and each token goes to 8 routed experts plus 1 shared one. Attention follows the Qwen3.5 hybrid pattern: three Gated DeltaNet linear-attention layers for every gated full-attention layer, repeated ten times.

Two consequences follow. Compute per token is tiny, so generation is fast and one GPU can serve many concurrent requests. Memory is not tiny: all 35B parameters must be resident, roughly 70GB at 16-bit or around 20GB at 4-bit before KV cache. The linear-attention layers keep that cache small at long context, which is what makes the native 262,144-token window affordable. Mixture of experts explained covers why sparse models behave this way.

The benchmark profile: strong agents, thinner long context

The shared table in the Qwen3.5 model cards shows where the small active count costs something and where it does not.

Where it leads or ties its larger siblings:

• TAU2-Bench: 81.2, the best of the three medium models (122B: 79.5, 27B: 79.0).
• AndroidWorld: 71.1, also the best of the three.
• ERQA (embodied reasoning): 64.8, ahead of both.
• BrowseComp-zh: 69.5, close to the 122B's 69.9.

Where it falls behind:

• AA-LCR (long-context reasoning): 58.5, against 66.9 and 66.1. This is the clearest gap.
• SWE-bench Verified: 69.2, against about 72 for both siblings.
• LiveCodeBench v6: 74.6, against 80.7 for the 27B.
• IFBench: 70.2, against about 76 for both.

One reading of the pattern: customer-service style tool use and phone-GUI tasks reward short, well-formed actions, while reasoning across a very long document asks for more than the small active count appears to supply. If your workload is long-document analysis, this is the model in the family to be most careful with. Long context degradation covers how to test that before it bites.

Base weights and FP8: built to be adapted

The 35B-A3B has a published base model, Qwen3.5-35B-A3B-Base, alongside the instruction-tuned weights, and Qwen also publishes an FP8 checkpoint, Qwen3.5-35B-A3B-FP8. That combination makes it a natural starting point for domain fine-tuning: you can train from the base rather than fighting the instruct model's existing habits, and serve in FP8 to halve the weight memory.

The usual MoE caveat applies. Fine-tuning a sparse model can unbalance expert routing if the dataset is narrow, so evaluate on held-out general tasks as well as your target domain. Load balancing in MoE explains what goes wrong.

Hosted: Qwen3.5-Flash and third-party endpoints

The model card identifies Qwen3.5-Flash as the hosted version of this model on Alibaba Cloud Model Studio. The Qwen3.5-Flash guide covers that endpoint. Third-party providers serve the open weights directly.

At the time of writing OpenRouter lists Qwen3.5-35B-A3B at about $0.15 per million input tokens and $1.00 per million output, with cached input at $0.05. That is roughly 40 per cent of the 27B's output rate and half of the 122B's. The listing shows a maximum output of 16,384 tokens on that route, lower than the 32,768 Qwen recommends for thinking mode, so check the cap before relying on long reasoning through a given provider.

As with every Qwen3.5 model, thinking is on by default. Send enable_thinking: false in chat_template_kwargs for direct answers; the Qwen3 /nothink switch is not officially supported. For a cheap workhorse, non-thinking mode is often the configuration that makes the economics work.

Its successor arrived in April

Qwen released Qwen3.6-35B-A3B in April 2026, keeping the same total and active size. The Qwen3.6-35B-A3B guide covers what changed. For a new deployment, test the 3.6 model first; the hardware requirements are the same, so switching costs little beyond re-running your evaluation.

Stay on 3.5 if you have fine-tuned from its base weights or validated it in production. A newer model that is better on average can still be worse on the specific cases your pipeline depends on.

A routing role, not a flagship role

The best use of this model is usually as the first stop in a routed setup:

1. Send every request to the 35B-A3B in non-thinking mode.
2. Validate the output with something cheap: a schema check, a test run, a confidence threshold.
3. Escalate failures to Qwen3.5-27B for coding or Qwen3.5-122B-A10B for long agent runs.

Measure the escalation rate. If most requests stay on the small model, the blended cost sits close to its price while quality on hard cases comes from the larger one. Model routing and fallbacks covers the details.