Qwen 27B: Which 27B You Are Holding, and Which One to Download

Three checkpoints in A​libaba's Q​wen line carry the number 27, and they are not the same model: Qwen3.5-27B, Qwen3.6-27B and Qwen3.8-27B are one size class spread across three generations, separated in the string by nothing but a set of version digits. This is a standing reference page rather than a launch report, and its warrant is first-party search demand we measured on 2026-10-13: 839 impressions and 24 clicks across fifty queries on the 27B stem and its qualifiers, against 72 impressions and a single click across the two bare qwen 27b queries inside that set. Fifty queries chasing one size class, with almost nobody searching the bare stem, is the signature of navigational confusion rather than of a launch. The release dates below date each checkpoint; they are not an announcement. The job here is disambiguation first and specification second — a reader who leaves knowing which 27B to download has been served, and one who leaves knowing only that the line contains several 27B models has not.

The 27B roster, at a glance

Every member of the line that carries 27B in its name, in release order. Full repository id, generation, and the date the weights were published:

• Qwen/Qwen3.5-27B — generation Q​wen3.5 — weights published 24 February 2026.

• Qwen/Qwen3.6-27B — generation Q​wen3.6 — weights published 21 April 2026.

• Qwen/Qwen3.8-27B — generation Q​wen3.8 — weights published 5 August 2026.

That is the whole roster, and three things follow from it that save readers time. First, the same-generation quantisations are not additional members: Qwen/Qwen3.5-27B-FP8 and Qwen/Qwen3.5-27B-GPTQ-Int4, Qwen/Qwen3.6-27B-FP8, and Qwen/Qwen3.8-27B-FP8 are the same weights re-encoded at lower precision for hardware that wants them, published under the same licence. Second, the three chat checkpoints are joined by community derivative work — most visibly the OrcaRouter-hosted orcarouter/Qwen3.8-27B-Uncensored family, which republishes the Q​wen3.8 weights in GGUF, MLX, NVFP4, FP8 and INT8 forms, alongside orcarouter/Qwen3.8-27B-MLX, a conversion of the unmodified base weights. Third, a checkpoint genuinely named 27B cannot predate the Q​wen3.5 generation, because the earlier generations used different sizes throughout: Q​wen3 shipped a dense 32B and a 30B-A3B mixture-of-experts pair, and Q​wen2.5 shipped 32B and 72B, with no 27B anywhere in either line.

How to tell which one you have

The naming trap is that the same number appears in three different generations, and the version digits are the only difference in the string. That is not a cosmetic problem, because the three checkpoints are otherwise described identically at the level a local file gives you. All three report model_type qwen3_5 and architecture Qwen3_5ForConditionalGeneration in their configuration, all three carry 27,781,427,952 parameters — roughly 27.78 billion, counted from the tensor index rather than from the marketing copy — and all three declare the same 55.56 GB of BF16 weights. A directory called 27b on disk, or a config with the vendor name stripped out, therefore identifies the size and nothing else.

What does identify the generation is the repository path, the publication date, and the shard count in model.safetensors.index.json:

• Qwen/Qwen3.5-27B — 11 shards holding 1,199 tensors, a 55.56 GB BF16 checkpoint, index written 24 February 2026. Its index lists 96 float32 tensors totalling 8,448 parameters, about 33 KB: per-layer gating and normalisation values that the later two generations moved into BF16.

• Qwen/Qwen3.6-27B — 15 shards holding 1,199 tensors, a 55.56 GB BF16 checkpoint, index written 21 April 2026. All 1,199 tensors are BF16.

• Qwen/Qwen3.8-27B — 18 shards holding 1,199 tensors, a 55.56 GB BF16 checkpoint, index written 5 August 2026. All 1,199 tensors are BF16.

Shard count is the most useful tell when the repository path has been lost, because it travels with the downloaded files — and the shard filenames differ too, reading model.safetensors-00001-of-00011 for the Q​wen3.5 generation against model-00001-of-00018 for Q​wen3.8. If what you have is a quantised build rather than the BF16 release, read the version digits off the quantiser's repository name instead: that is the one place the generation survives a re-encode.

What actually differs in practical terms

Once you know which of the three you have, the differences that change a download decision are narrower than the version numbers suggest. Every figure below is attributed to whoever published it, and none of it is independently benchmarked by us.

• Context, vendor-reported. All three model cards state 262,144 tokens natively. The Qwen3.5-27B and Qwen3.6-27B cards state extension up to 1,010,000 tokens; the Qwen3.8-27B card states 1,000,000. The native figure is identical in every case — the delta is entirely in the extended claim, and it is a card claim rather than a measured ceiling.

• Licence, vendor-reported. All three carry apache-2.0 on the vendor's own model card. The same licence covers the official FP8 and GPTQ-Int4 re-encodes of each generation, so moving down in precision does not change your terms.

• What the licence permits, and what people build with it. Apache-2.0 permits commercial use, modification and redistribution with attribution, which is why third-party re-releases exist alongside the upstream weights: alternate-precision conversions, derivative quantisations and modified variants are all redistributable on the same terms. That licensing fact is the whole of what this page says about the most common qualified query on this stem, qwen 27b uncensored — modified and re-named derivatives are permitted by the licence, and what any particular one of them does is a question for its own maintainer.

• Quantisation availability after release, third-party download counts. GGUF coverage is not equally mature across the three generations. At the time of writing, the unsloth GGUF repositories for this size class carried 105,370 downloads across 26 files for the Q​wen3.5 generation, 578,462 across 26 files for Q​wen3.6, and 6,650,349 across 30 files for Q​wen3.8 — a spread of more than sixtyfold between the oldest and the newest. The size ladders differ at the tails: 11.21 GB at UD-Q2_K_XL rising to 28.60 GB at Q8_0 on Q​wen3.5, 11.85 GB to 28.60 GB on Q​wen3.6, and 9.83 GB at UD-Q2_K_XL up to 31.46 GB at UD-Q8_K_XL on Q​wen3.8, which reaches both lower and higher than the two older sets. NVFP4 is the sharpest tell of the same drift: the Q​wen3.6 and Q​wen3.8 generations each have several widely-downloaded conversions, led by unsloth at 878,812 and 1,941,507 downloads respectively, while the Q​wen3.5 generation's NVFP4 builds are small community efforts in the hundreds to low thousands.

• Thinking controls, vendor-reported. The control surface moved with each generation, and this is the part most likely to break an upgrade. Qwen3.5-27B and Qwen3.6-27B both think by default and take a disable flag; neither supports the old Q​wen3 /think soft switch. Qwen3.6-27B adds preserve_thinking as an opt-in for retaining reasoning traces across turns. Qwen3.8-27B adds a reasoning_effort setting with xhigh, medium and low levels, defaulting to xhigh, and turns preserve_thinking on by default for all workloads. If your harness hard-codes a chat template written against the Q​wen3.5 generation, that is the difference that will bite first.

• What did not change. The parameter count, the architecture string, the native context window and the licence are identical across all three. Choosing between them is choosing a generation's training and tooling, not a size and not a legal footing.

Which of them we serve

On the OrcaRouter catalogue, three ids carry the 27B class today, and they are the whole of it: qwen/qwen3.5-27b, qwen/qwen3.8-27b and the Obsidian-hosted obsidian/Qwen3.8-27B. The other ids under our Q​wen card are different sizes and different generations and are not summarised here.

• qwen/qwen3.5-27b — the oldest of the three generations, listed with a 32,768-token context window at $0.086 per million input tokens and $0.688 per million output tokens, with higher-context tiers priced above that.

• qwen/qwen3.8-27b — the newest, listed at the full 262,144-token native context and self-hosted on our own infrastructure, at $0.33 per million input tokens and $2.40 per million output tokens.

• obsidian/Qwen3.8-27B — the same Q​wen3.8 weights under a different host, listed at a 262,144-token context window and $0.40 per million input tokens and $4.21 per million output tokens; that host serves the block-FP8 conversion with the vision tower left at full precision.

There is no 27B id on the catalogue corresponding to the Q​wen3.6 generation. If you want the middle generation for comparison work, the weights are open and the download advice above applies; we simply do not serve it.

Availability

All three checkpoints are open-weight releases under Apache-2.0 on the vendor's own Hugging Face organisation, so routing through us is a convenience rather than the only route to the model. Where we do serve the 27B class, it is the three ids above and nothing else, and the two Q​wen3.8 routes differ in host and price rather than in the weights they serve. Nothing here should be read as a claim about the wider Q​wen catalogue: this size class is three ids out of a much longer list, and we have named only those three.

For readers who want this size class in a footprint that fits a single smaller accelerator, the OrcaRouter-hosted orcarouter/OrcaSAQ-2-27B is a sensitivity-aware three-bit mixed-precision re-encode of the Qwen3.8-27B weights: 54 GB of BF16 reduced to 12.3 GB at 3.21 bits per weight, with a measured perplexity of 5.6482 against 5.6468 for the source, 93.2% top-1 agreement, and the 262,144-token context preserved. Those figures are ours, published on the repository and reproducible from it — which makes them vendor-reported numbers of exactly the kind this page has been labelling as such throughout.

What is not settled

One thing is genuinely open, and it is worth stating rather than guessing. The Qwen3.8-27B model card carries a vendor note that a hosted version with a 1M-token context window and official built-in tools is coming, with no date attached to it. Until that ships, the extended-context figures above remain card claims rather than service levels, and the native 262,144 tokens is the number to plan against.

Where to go next

This page exists to route you to the right one of three, not to restate what its neighbours already cover in depth. If the Q​wen3.8 generation is the one you have, its specification, context window and local-run recipes are documented in their own pages — start from the full specification there rather than from the summary above. If what you actually want is the generation after this one rather than this size class, that is a separate and differently-warranted question, and the settled-versus-unsettled position on it belongs on that page, not here. For everything in between, the three lines in the roster section are the map: match the version digits, then confirm with the shard count.