Qwen3.5 27B: The Dense One in a Sparse Generation

What Qwen3.5 27B is

Qwen3.5 27B is the small dense model of Alibaba's Qwen3.5 generation — Apache 2.0, self-hostable, with a 256K token context window, at roughly $0.30 per million input tokens and $2.40 output through third-party providers. The rest of the generation covered on this site is mixture-of-experts, from the 35B-A3B up to the 397B-A17B flagship. A dense 27 billion parameter model activates all of its weights on every token, which makes it more expensive per token than a sparse model of similar total size and considerably simpler to deploy.

Dense and sparse

A mixture-of-experts model needs memory for all its parameters even though it computes with a fraction of them: a 397B-A17B model needs memory for 397 billion parameters and computes with 17 billion. A dense 27B model needs memory for 27 billion and computes with all of them.

What it inherits

It carries the generation's architecture: a unified vision-language foundation trained with early fusion on multimodal tokens, which Alibaba say reaches parity with Qwen3 while outperforming the separate Qwen3-VL line across reasoning, coding, agents and visual understanding. Language coverage runs to 201 languages and dialects, and reinforcement learning was scaled across what Alibaba describe as million-agent environments.

LMArena

On LMArena's text leaderboard Qwen3.5 27B has 1409 Elo (rank 155); by category, 1451 Elo in Coding (rank 159), 1423 Elo in Math (rank 126), 1441 Elo in Expert (rank 137) and 1358 Elo in Creative Writing (rank 174).

Alibaba’s figures

Alibaba's mid-size comparison table covers GPT-5-mini, GPT-OSS-120B, Qwen3-235B-A22B and the three Qwen3.5 mid-size models. Reading the rows the card labels: LongBench v2 gives Qwen3.5-122B-A10B 60.2, Qwen3.5-27B 60.6 and Qwen3.5-35B-A3B 59.0, against GPT-5-mini's 56.8 and GPT-OSS-120B's 48.2; HMMT February 2025 gives 91.4, 92.0 and 89.0 against 89.2 and 90.0; HMMT November 2025 gives 90.3, 89.8 and 89.2 against 84.2 and 90.0; and LiveCodeBench v6 gives 78.9, 80.7 and 74.6 against 80.5 and 82.7. Qwen3-235B-A22B records 54.8, 85.1, 89.5 and 75.1 on the same four. These are Alibaba's own figures.

Architecture

The card lists 64 layers with a hidden dimension of 5,120 and an intermediate dimension of 17,408, arranged as 16 repeats of three Gated DeltaNet layers followed by one gated-attention layer, each feeding a dense feed-forward block rather than mixture-of-experts. The linear-attention layers use 48 value heads and 16 query-key heads of dimension 128; the gated-attention layers use 24 query heads and 4 key-value heads of dimension 256. The repository holds the post-trained model in Hugging Face Transformers format, compatible with Transformers, vLLM, SGLang and KTransformers. The context is 262,144 tokens natively, extensible to 1,010,000. The token embedding is padded to 248,320 entries. Alibaba also describe near-100% multimodal training efficiency compared with text-only training, and asynchronous reinforcement-learning frameworks for large agent scaffolds.

The Qwen3.5 family

The Qwen3.5 open-weight releases covered on this site are the 397B-A17B flagship, the 122B-A10B and 35B-A3B mixture-of-experts models, and the dense 27B. Alibaba name hosted versions for two of them on Alibaba Cloud Model Studio: Qwen3.5-Plus for the 397B-A17B and Qwen3.5-Flash for the 35B-A3B. The models can also be tried in Qwen Chat at chat.qwen.ai.

Official Resources