Qwen 3.5: Four Size Classes, Two Hosted Tiers, and the Eight IDs We Serve

Qwen 3.5 is the February 2026 generation of A​libaba's Q​wen line — the oldest generation still carrying measurable search demand — and its four open-weight size classes are Qwen3.5-397B-A17B, Qwen3.5-122B-A10B, Qwen3.5-35B-A3B and Qwen3.5-27B, with Qwen3.5-Flash and Qwen3.5-Plus as the two hosted tiers that carry the same generation name. If you are choosing it today over Q​wen 3.6, Q​wen 3.7 or Q​wen 3.8, the honest answer is nearly always one of three: it is smaller than what replaced it, it is cheaper to call, or it is already wired into a pipeline and nothing in the newer generations has yet given you a reason to move.

This is a standing page, not a launch report. The dates below — weights for the flagship published on 16 February 2026, the three mid sizes on 24 February — are here to place the generation, not to announce it. What earns the page its place is a different measurement. In the thirty days to 2026-10-13 our own Search Console property recorded 82 impressions and 2 clicks across fourteen qwen 3.5 queries, with 92% of those impressions sitting on qualified queries and the bare term taking 16 impressions and no clicks at all. Fourteen queries across a generation with four size classes is not browsing traffic; it is readers doing homework, most of them holding a specific size or tier and checking what exists. The useful thing to hand them is precision about the ids, so that is what this page is.

Why anyone still picks this generation

Qwen3.5 is the generation that put the architecture the later releases still use on the map. Every open-weight member is a causal language model with a vision encoder, built as a hybrid of Gated DeltaNet blocks and Gated Attention with a sparse mixture-of-experts on top, trained with multi-step multi-token prediction. The vendor's own model cards describe the family as a leap in multimodal learning, architectural efficiency and accessibility, and claim cross-generational parity with the Qwen3 line while beating the Qwen3-VL line across reasoning, coding, agent and visual benchmarks — Qwen's figures, unaudited by anyone outside the lab, and repeated here as the vendor's claims rather than as measurements.

What that buys a reader in October 2026 is a floor, not a frontier. The 3.5 line is where the size classes start, where the prices are lowest, and where the quantised ecosystem has had the longest to mature. If your constraint is a machine you own or a bill you pay per million tokens, 3.5 is the generation you check first; if your constraint is a benchmark, it is not, and the later hubs are one click away at the bottom of this page.

The members, one line each

Four size classes carry the generation name in its open-weight releases, and two hosted tiers carry it on the API side. The MoE notation is the part worth decoding once: 397B-A17B means 397 billion total parameters with about 17 billion active per token, and the same reading applies to 122B-A10B and 35B-A3B.

• Qwen3.5-397B-A17B — the flagship of the generation and the widest MoE: 397B total, 17B active, 512 experts with 10 routed and 1 shared. Weights published 16 February 2026, the earliest member of the family. This is the one to reach for if you are serving the generation rather than running it.

• Qwen3.5-122B-A10B — the middle MoE: 122B total, about 10B active, 256 experts, and on the vendor's own tables the strongest of the three mid sizes on knowledge, document and agentic-search work. It is also the size where the vendor's own numbers put a real gap between this generation and the dense 27B, in the 122B's favour.

• Qwen3.5-35B-A3B — the small MoE and the practical self-hosting pick: 35B total with roughly 3B active per token, which means the memory footprint of a small model with appreciably more capacity than a 3B dense model carries. If you are fitting the generation onto one consumer GPU, this is the member that question is usually about.

• Qwen3.5-27B — the dense member: a single 27-billion-parameter checkpoint with a vision encoder and no expert routing to reason about. Qwen's own tables put it ahead of the 122B on several instruction-following and coding-competition tasks, so dense is not simply the weaker choice here; it is the simpler deployment.

• Qwen3.5-Flash — the hosted cost tier, and per the vendor's own model cards the API-side equivalent of the open weights, listed with a 1M-token default context. Nothing to download.

• Qwen3.5-Plus — the hosted tier above Flash: multimodal chat over text, image and video with a 1M-token context and built-in tools, aimed at coding and agent work. Nothing to download either.

Two lines of housekeeping so nobody goes looking for members that do not exist. First, the generation is wider than four sizes at the small end: Qwen3.5-9B, Qwen3.5-4B, Qwen3.5-2B and Qwen3.5-0.8B are official releases published in late February 2026, each with a -Base pretrained sibling, and they are the sizes to check if you are targeting a laptop. Second, the suffixed names you meet in a provider's catalogue are not extra members: -FP8 and -GPTQ-Int4 are the same checkpoints re-encoded at lower precision for hardware that prefers it, published under the same licence.

Which of those we serve, written out

Eight ids from this generation are live on the OrcaRouter catalogue as of 2026-10-13, verified against our own model list today and listed here exactly as the catalogue spells them. This is the page's real payload, so the list is precise rather than summarised. Four of the eight are the downloadable open-weight members and four are the two hosted tiers, each carried in both an undated and a dated spelling.

• qwen/qwen3.5-397b-a17b — the 397B-A17B flagship.

• qwen/qwen3.5-122b-a10b — the middle MoE.

• qwen/qwen3.5-35b-a3b — the small MoE.

• qwen/qwen3.5-27b — the dense member.

• qwen/qwen3.5-flash — the Flash tier, undated alias.

• qwen/qwen3.5-flash-2026-02-23 — the Flash tier, pinned to its 23 February 2026 snapshot.

• qwen/qwen3.5-plus — the Plus tier, undated alias.

• qwen/qwen3.5-plus-2026-02-15 — the Plus tier, pinned to its 15 February 2026 snapshot.

The two dated ids are the part of that list worth reading twice, because a dated id and an undated one are not two spellings of one route. qwen/qwen3.5-flash and qwen/qwen3.5-plus are aliases: they track whatever build the provider currently points each name at, so their behaviour can shift under you without the id changing. qwen/qwen3.5-flash-2026-02-23 and qwen/qwen3.5-plus-2026-02-15 are frozen snapshots of a named build. If you are prototyping, the alias is fine and you get fixes for free. If a prompt or a regression test was tuned against a specific build, pin the dated id — that is the entire reason both forms exist in the same list.

Note also what the ids themselves tell you about this generation: four of the eight carry a size class in the name and four do not. The generation is served here as both a set of downloadable weights and a pair of hosted tiers, which is not true of every Qwen generation — 3.7, for instance, survives on this catalogue as hosted tiers only, with no weight-bearing id under it at all.

Where 3.5 sits next to the generations that followed it

The line runs 3.5 in February 2026, then 3.6 in April, 3.7 across May to July, and 3.8 from August onward. Read as a sequence, the later generations did not replace the 3.5 size classes — they kept the architecture and pushed the training, the tooling and the top of the range. That is why 3.5 remains a live choice rather than an archived one. Our Qwen 3.6 hub maps the generation that inherited this architecture and renamed the same 27B and 35B-A3B sizes; Qwen 3.7 covers the hosted-only generation between 3.6 and 3.8; Qwen 3.8 is the current line and the one to read if budget is not the binding constraint; and Qwen 4 collects what is known about the generation after it. This page places 3.5. Those pages run the comparisons.

Licensing, from the licence text

Every open-weight member of this generation is published under Apache-2.0 — the flagship, the 122B, the 35B-A3B, the dense 27B, the small sizes at 9B, 4B, 2B and 0.8B, and the -Base pretrained siblings, all of them, on the vendor's own model cards. The official FP8 and GPTQ-Int4 re-encodes inherit the same terms, so stepping down in precision does not change your legal footing.

Apache-2.0 is permissive: it allows use, modification and redistribution, including commercially, provided attribution and the licence text travel with the work. That single fact is the whole of what this page says about the qualified form of the generation's search demand. Readers arrive on this stem looking for modified builds of these checkpoints, and the answer is a licensing answer — because the weights are open and the licence permits it, derivative checkpoints, quantised re-encodes and fine-tunes are lawful and are exactly what the community repositories under these names are. What a hosted service does with a request is a different question, governed by that service's terms, and nothing on this page speaks to it.

The two hosted tiers are the exception in the roster: Qwen3.5-Flash and Qwen3.5-Plus have no published weights, so there is no licence text to read for them and nothing to self-host. What governs your use of those is the provider's terms of service.

What is not settled

A​libaba's own model cards carry no release-date field — the dates we can stand on are the timestamps of the repositories themselves, which is why this page dates the weights to February 2026 and does not put a day on the generation's launch. We also found no vendor blog post for the generation we can read in full, and the tier naming is thinner than the roster suggests: Flash and Plus are the two hosted names we can source for this generation, and we have found no Max tier carrying the 3.5 label to place alongside them.

Availability, and which size to pick

All eight ids named above are served on OrcaRouter as of 2026-10-13: qwen/qwen3.5-397b-a17b, qwen/qwen3.5-122b-a10b, qwen/qwen3.5-35b-a3b, qwen/qwen3.5-27b, qwen/qwen3.5-flash, qwen/qwen3.5-flash-2026-02-23, qwen/qwen3.5-plus and qwen/qwen3.5-plus-2026-02-15. That is the whole of the served set for this generation, and this page makes no availability claim about any Qwen id outside it.

Which one to call follows from your hardware, not from the version number. If you are downloading weights rather than calling an API, match the size class to the machine: Qwen3.5-35B-A3B is the member built for a single consumer GPU because only about 3B parameters are active per token, Qwen3.5-27B is the one to take if you would rather have a dense checkpoint with no routing to reason about, and Qwen3.5-122B-A10B or the 397B-A17B flagship are serving-side choices rather than desktop ones. Sizes below 27B exist and are worth checking if the machine is small. If you are calling rather than hosting, the two hosted tiers are the shorter path, and the pinned snapshot of whichever tier you settle on is the safer id once something downstream depends on it.

That is also where a routing layer earns its keep on a generation this old. Because all eight ids sit behind one key, a request moves between the 35B-A3B, the dense 27B and the 122B as a config change rather than a second contract or a code rewrite — which is the shape you want when you are still deciding which size class your workload deserves. Pricing here is the provider's own list price passed through with no markup added by us, so a provider-side rate change reaches you the same day it lands, and automatic failover is the cheaper way to test a size you are unsure of than putting it in front of production traffic.