Qwen3.5-397B-A17B: The Open Flagship Behind Qwen3.5-Plus

Qwen3.5-397B-A17B is the model the whole Qwen3.5 family was launched with on 16 February 2026, and it is easy to misread. The headline figure is 397 billion parameters, which sounds like a cluster-only model. The number that decides what it costs to run is the other one in the name: 17 billion active per token.

It is also the open-weight model behind Qwen3.5-Plus. Alibaba's model card describes Plus as "the hosted version corresponding to Qwen3.5-397B-A17B", with a 1M default context and built-in tools added. So the real question for most teams is not whether this model is good, but whether to rent it as Plus or own it as weights.

Three linear-attention layers for every full one

The architecture is the most interesting part of this release. The model has 60 layers arranged as 15 repeating blocks. Each block is three Gated DeltaNet layers followed by one Gated Attention layer, with a mixture-of-experts feed-forward after every one of them.

Gated DeltaNet is a linear-attention mechanism. Its cost grows linearly with sequence length rather than quadratically, and it keeps a fixed-size state instead of a key-value cache that grows with every token. The one-in-four Gated Attention layers are conventional attention, with 32 query heads sharing 2 KV heads, which keeps the cache that does exist small.

The practical effect is that long prompts are cheaper to serve than on an all-attention model of similar size. The trade is that linear attention compresses history into a fixed state, which is why the full attention layers are still there. It is a deliberate compromise rather than a free win.

Native context is 262,144 tokens. The card says it extends to about 1,010,000 tokens with YaRN scaling, and warns that static YaRN can hurt performance on short texts, so only switch it on when you actually need the length.

17B active out of 512 experts

The MoE layers hold 512 experts each. Per token, 10 routed experts and 1 shared expert fire. That is how a 397B model ends up with a forward pass closer to a mid-sized dense model.

Two consequences follow, and they pull in opposite directions. Compute per token is low, so throughput and per-token latency are good for the capability class. Memory is not low: every one of the 397B parameters still has to be resident, because any token might route to any expert. Active vs total parameters works through why both numbers matter and which one each cost line depends on.

The model is also trained with multi-token prediction, and both vLLM and SGLang can use it for speculative decoding. That is a free speed-up if your serving stack supports it.

Reading the vendor benchmarks

Qwen's model card reports, among others:

• Knowledge and reasoning: MMLU-Pro 87.8, GPQA 88.4, HLE 28.7, AIME26 91.3
• Coding: LiveCodeBench v6 83.6, SWE-bench Verified 76.4, Terminal Bench 2 52.5
• Agents and tools: BFCL-V4 72.9, TAU2-Bench 86.7, BrowseComp 69.0
• Vision: MMMU 85.0, MathVision 88.6, OmniDocBench1.5 90.8, Video-MME with subtitles 87.5, OSWorld-Verified 62.2

The shape is broad rather than spiked. It is a generalist with native image and video input, not a coding specialist. The SWE-bench Verified and Terminal Bench figures are respectable but sit below the agentic-coding leaders of mid-2026, and Qwen's own later releases have pushed coding further.

These are vendor-run numbers on vendor harnesses. BrowseComp alone is reported twice on the card, at 69.0 and 78.6, depending on how the agent manages its context. Treat them as a shortlist filter, not as a prediction of how the model handles your repository.

Self-host it, or call Plus

The weights are Apache 2.0, which is about as permissive as model licences get: commercial use, fine-tuning and redistribution without a separate agreement. Open weights vs open source explains why that matters most at the point a product succeeds.

The serving examples on the card use eight-way tensor parallelism. That tells you the intended deployment: a full multi-GPU node, not a workstation. Unless you already run that kind of hardware, or have data that cannot leave your network, the hosted route is simpler.

On OpenRouter the open model is listed at $0.50 per million input tokens and $3.60 per million output at the time of writing, with cached input cheaper. The hosted Plus product is priced separately by Alibaba and adds a 1M default window; the Qwen3.5-Plus guide covers its tiered pricing and snapshots.

Settings that matter on day one

• Thinking is on by default. Output starts with a reasoning block. Pass enable_thinking: false in the chat template arguments to turn it off. The /think and /nothink soft switches from Qwen3 are not supported.
• Sampling differs by mode. The card recommends temperature 0.6, top_p 0.95, top_k 20 for thinking, and temperature 0.7, top_p 0.8, presence penalty 1.5 for non-thinking.
• Do not starve the context. The card advises keeping at least 128K of context available to preserve thinking quality. Capping the window to save memory can quietly degrade reasoning.
• Text-only serving exists. vLLM's --language-model-only flag skips the vision encoder if you never send images.

A test plan before you commit

1. Pick twenty tasks from real work, including some with screenshots or PDFs if vision is part of the appeal.
2. Run them through the hosted API in thinking and non-thinking mode, and record output tokens as well as quality. Thinking mode can multiply output volume.
3. Compare against one smaller sibling, such as the Qwen3.6-35B-A3B, on the same tasks. If the gap is small, the larger model is costing you memory for nothing.
4. Only then price a self-hosted node, using measured tokens rather than estimates.

How to benchmark LLMs on your own work has the harness details. The decision rule is simple: if your workload is broad, multimodal and high-volume, and the licence matters, this is one of the strongest Apache 2.0 options in its family. If you only need text and cost per task matters most, measure a smaller model first.