Qwen3.5-27B: The Dense Model That Keeps Up With Bigger Siblings

The medium Qwen3.5 release on 24 February 2026 contained two mixture-of-experts models and one dense one. The dense one, Qwen3.5-27B, has the fewest total parameters of the three, and on several of Qwen's own coding and instruction-following numbers it is level with or ahead of the 122B.

That is the interesting fact about this model. It is not the cheapest per token and not the fastest, but it puts near-122B quality on single-repository work into a footprint that is far easier to own.

Why a dense model in a mostly-MoE family

Mixture-of-experts models are cheap to run per token but expensive to hold in memory, because every expert has to be loaded even though few are used. A dense model inverts that. All 27B parameters run on every token, so compute per token is higher than a 3B-active MoE, but the memory you need is exactly what the parameter count says.

For self-hosting, memory is usually the binding constraint. A 27B dense model at 4-bit is in the mid-teens of gigabytes; the 35B-A3B MoE at the same precision is larger, and the 122B is several times larger. Fine-tuning a dense model also avoids MoE-specific concerns such as expert load balancing. Dense vs mixture-of-experts covers the trade in more depth.

Linear attention in a dense body

The 27B shares the Qwen3.5 hybrid attention design. It has 64 layers arranged as sixteen repeats of three Gated DeltaNet layers followed by one gated full-attention layer, each with a standard feed-forward block rather than experts. Hidden size is 5,120. Like its siblings it is trained with multi-token prediction and is natively vision-language, taking text, images and video.

The linear-attention layers matter for long context. Only one layer in four keeps a conventional KV cache that grows with every token, so a long prompt costs much less memory than it would on an all-attention model of the same size. Native context is 262,144 tokens, extensible to about one million with YaRN according to the model card. The licence is Apache-2.0.

Benchmarks: closer to 122B than its size suggests

The Qwen3.5 model cards share one comparison table across the three medium models. Picking the rows where the 27B stands out:

• SWE-bench Verified: 72.4, the highest of the three (122B: 72.0, 35B-A3B: 69.2).
• IFEval: 95.0, the highest of the three, and IFBench 76.5.
• LiveCodeBench v6: 80.7, ahead of the 122B's 78.9.
• LongBench v2: 60.6, against 60.2 for the 122B.
• CharXiv reasoning: 79.5, the best chart-understanding score of the three.

It trails the 122B where agents run long: 41.6 against 49.4 on Terminal Bench 2, and 68.5 against 72.2 on BFCL-V4. Its CodeForces rating of 1,899 is also the lowest of the three. The shape is a model that is strong at a bounded coding task and at doing exactly what it is told, and less strong at long tool-driven sessions. SWE-bench explained covers what a 72 actually measures.

What dense costs you

Three honest drawbacks.

Throughput. Running 27B parameters per token is about nine times the compute of the 35B-A3B's 3B active. On the same hardware the MoE will generate faster and serve more concurrent users.

Hosted price. At the time of writing OpenRouter lists Qwen3.5-27B at about $0.26 per million input tokens and $2.60 per million output, which is more than the 122B's $2.08 output rate and well above the 35B-A3B's $1.00. Higher compute per token is one plausible reason, though listings are set by providers and can change.

Thinking by default. Like every Qwen3.5 model it reasons before answering unless you send enable_thinking: false in chat_template_kwargs. The old /nothink switch is not officially supported. With thinking on, output tokens and latency climb, and the output column is where the price is.

The serving examples in the model card use eight-way tensor parallelism, but that reflects a high-throughput deployment rather than a minimum. At 16-bit the weights are around 54GB, which fits a single 80GB card; quantised builds bring it within reach of a single 24GB or 32GB card at reduced context. Which open model fits your GPU has the arithmetic.

Qwen3.5-27B or Qwen3.6-27B?

Qwen shipped a successor in April 2026. Qwen3.6-27B is also a dense 27B and reports 77.2 on SWE-bench Verified, about five points above this model's 72.4. The Qwen 3.6 guide covers it in detail.

For new projects, start with 3.6: same size class, same hardware, a better coding score. Stay on 3.5 when you have already fine-tuned it, built an evaluation set around its behaviour, or pinned it in production and have no failing cases that the newer model would fix. Upgrading a working model has a cost of its own, which regression testing a model upgrade covers.

When to pick it

• Pick Qwen3.5-27B when you want to self-host a capable coder on one GPU, need tight instruction following, or plan to fine-tune.
• Pick Qwen3.5-122B-A10B for terminal and browsing agents where its lead on Terminal Bench 2 is real.
• Pick Qwen3.5-35B-A3B when tokens per second or price per token decides it.

The quickest way to confirm the fit is ten real tickets from your repository, run with thinking on and off. If the non-thinking mode solves most of them, you get the dense model's quality without the reasoning bill.