Qwen 3.6: Which Member You Are Holding, and Which ID to Call

Nobody searches for Qwen 3.6. In the thirty days to 2026-10-13 our own Search Console property recorded 409 impressions and 8 clicks across the entire qwen 3.6 neighbourhood — fourteen queries — and exactly zero on the bare term. It is the only facet in its batch with no bare demand at all. Readers arrive here from somewhere else: a tier name, a comparison against a newer generation, or a served-model list. They arrive holding an id and needing to place it. So this page is the map, not the story. Qwen3.6 is the April 2026 generation, and it has five named members: the API-only Qwen3.6-Plus and Qwen3.6-Max-Preview, the cost tier Qwen3.6-Flash, and the two open-weight releases Qwen3.6-35B-A3B and Qwen3.6-27B. The dates below are used to date each member. None of them is news.

The generation in one sentence

Qwen3.6 is the April 2026 generation of Alibaba's multimodal Qwen line, the step after Qwen3.5 in February 2026 and before the Qwen3.7 numbering, and in Qwen's own framing it is the generation where agentic coding broke through at every scale — from a 3B-active open-weights model to the hosted Plus and Max tiers.

What it introduced, per the two vendor posts we can read in full: an API feature called preserve_thinking, which retains the reasoning content of every preceding turn and is recommended for agentic work, and — on the vendor's own evaluation tables — a dense 27B checkpoint that outscores the previous generation's 397B open-weights flagship on the coding benchmarks Qwen chose to publish. Qwen's numbers there are Qwen's, unaudited; we repeat them as the vendor's claims, not as measurements.

The member map

Five members carry the generation's name. Three are hosted tiers with no published weights; two are open-weight releases you can download today. Each line below gives the member, what it is, and where it lives — this is the whole roster, and it is the part of the page most readers need.

• Qwen3.6-Plus — the flagship hosted tier: multimodal chat over text, image and video, a 1,048,576-token context window and 65,536-token maximum output, with function calling and the coding work Alibaba markets as Vibe Coding. No weights; API only. Qwen's own 35B-A3B post describes the 35B release as following "the launch of Qwen3.6-Plus", so it is the first member of the generation, and our catalogue carries a dated snapshot id of 2026-04-02 against it.

• Qwen3.6-Max-Preview — the preview tier at the top of the range, named alongside Plus as "API-accessible" in Qwen's summary of the 3.6 family. Preview means exactly that: no stable id, no weights, and no separate launch post we can point at.

• Qwen3.6-Flash — the cost tier. This one needs a plain statement, because it is the confusion most readers arrive with: Qwen3.6-Flash is how the Qwen3.6-35B-A3B release is served on Alibaba Cloud Model Studio. Qwen's own post says the open weights can be called "via API as Qwen3.6-Flash on Alibaba Cloud Model Studio API". So Flash is not a sixth architecture — it is the hosted face of the 35B-A3B checkpoint, with the same class of multimodal input and a 1,048,576-token context listing on our catalogue.

• Qwen3.6-35B-A3B — the open-weights mixture-of-experts member: 35 billion total parameters, about 3 billion active per token, natively multimodal, 262,144 tokens of native context extendable to 1,010,000 on the vendor's card. It has its own page in this batch, and the paragraph below hands off to it rather than describing it twice.

• Qwen3.6-27B — the open-weights dense member: a single 27-billion-parameter checkpoint with a vision encoder, built for the scale most people can actually self-host, and in Qwen's framing aimed at developers who want flagship-level coding without MoE routing complexity. 262,144 tokens native, extensible to 1,010,000.

Two things that are not additional members, and are worth naming so you do not go looking for them: Qwen3.6-35B-A3B-FP8 and Qwen3.6-27B-FP8 are the same two checkpoints re-encoded at lower precision for hardware that prefers it, published under the same licence, not new models. And the dated spellings you will meet in a provider's catalogue — qwen3.6-plus-2026-04-02, qwen3.6-flash-2026-04-16 — are frozen snapshots of the aliases above, same capabilities, pinned so a production path does not move under you.

Qwen3.6-35B-A3B, in one paragraph

The 35B-A3B notation is the thing to decode and it is the reason readers land on this member: 35B is the total parameter count of the mixture-of-experts, A3B is how many of them are active for each token — about 3 billion. That gives you the memory footprint of a small model with far more capacity than a 3B dense model carries, which is why it is the member people self-host. Its full specification, the quantisation formats available for it, and the comparison that actually drives its search demand — against Qwen3.8-27B — are all on its own page, Qwen 3.6 35B A3B, in this batch. Read that page for the architecture; this hub only places it.

Which ids we serve, written out

Six ids on the OrcaRouter catalogue belong to this generation today, verified against the live model list on 2026-10-13. All six are available now:

• qwen/qwen3.6-35b-a3b — the open-weights MoE member, listed at a 262,144-token context window and 65,536 tokens of maximum output.

• qwen/qwen3.6-flash — the hosted flash tier, listed at 1,048,576 tokens of context and 65,536 of maximum output.

• qwen/qwen3.6-flash-2026-04-16 — the frozen 16 April 2026 snapshot of the id above, same capabilities, same context listing. Use it when you want the version you tested to stay the version you call.

• qwen/qwen3.6-plus — the flagship hosted tier, 1,048,576 tokens of context, 65,536 maximum output, multimodal input.

• qwen/qwen3.6-plus-2026-04-02 — the frozen 2 April 2026 snapshot of the Plus alias.

• obsidian/Qwen3.6-35B-A3B — the same 35B-A3B checkpoint under a different host on our catalogue. The host's own model card describes it as a mixture-of-experts build of Alibaba's Qwen3.6 with 35B total and about 3B active parameters, a 262,144-token context window and text, image and video input, and lists it as an independent hosted entry under its own provider name and its own listing date. If you have seen that id somewhere else and expected it to be a different model, it is not: the weights are the 35B-A3B weights, the difference is who serves them and what they charge for it.

Note the one discrepancy rather than resolving it: Qwen's post identifies Flash as the API name for the 35B-A3B release, while our catalogue lists the weights member at 262,144 tokens and the flash ids at 1,048,576. Both figures are reported here as we read them.

That is also where a router earns its keep on an older generation. Because all six ids sit behind one key, placing a request in this family is a config change rather than a second contract, and the flash and weights spellings of the same checkpoint are directly substitutable — which is exactly the shape you want when you are comparing a hosted tier against the same weights served by a different host. Everything above is priced at the provider's own list price passed through, with no markup added by us, so a provider-side price change reaches you the same day it lands. For a generation this old the honest advice is the same as for any long-lived checkpoint: put the tier you are unsure about behind automatic failover rather than in front of production, and let the fallback absorb the first bad week.

Licensing, from the licence text

Both open-weight members carry Apache-2.0 on the vendor's own model card — Qwen3.6-35B-A3B and Qwen3.6-27B alike — and the official FP8 re-encodes of each inherit it, so stepping down in precision does not change your terms.

Apache-2.0 is a permissive licence: it allows use, modification and redistribution, including commercially, with attribution and the licence text carried along. In practice that is what the ecosystem around this generation looks like — quantised re-encodes, alternate-precision conversions and community derivative checkpoints all redistributable under the same terms they came in under.

The hosted tiers are a different situation and it is worth being exact about it: Qwen3.6-Plus, Qwen3.6-Max-Preview and the hosted Qwen3.6-Flash route have no published weights, so there is no licence text to read and nothing to self-host. What governs them is the provider's terms of service, not a model licence.

The question the biggest qualified query is really asking

The largest single query in this facet's neighbourhood over the measured window is a licence question wearing an adjective: readers search for an abliterated build of this generation, and the useful answer is the licensing one. Because the 35B-A3B and 27B weights are published under Apache-2.0 and that licence permits modification and redistribution, modified and fine-tuned derivatives are lawful and are exactly what the community repositories and the alternate-host listing above are. That is the whole of what this page says on the subject: what the licence and the published weights allow, and nothing about how any hosted service behaves.

What is not settled

Two members of this generation have no launch post we can read and no sourceable date. Qwen3.6-Plus launched before 16 April 2026 — that much follows from the 35B-A3B post referring back to it — and the only date we can put on it is the snapshot id on our own catalogue, so treat 2026-04-02 as the file naming rather than as a confirmed release date. Qwen3.6-Max-Preview has no dated artefact at all, and "Preview" is the vendor's own label. We have not found a vendor post for either, and we are not going to invent one.

If you arrived from a comparison

Most of the fourteen queries in this facet's neighbourhood are not about a member of this generation at all — they are comparisons against a newer one, and the 35B-A3B comparisons in particular belong to the 35B page in this batch. If that is what brought you here: the map above tells you which Qwen3.6 member the comparison is about, and the answer to "which of these should I run" is a size and a generation question, not a name question. If you came in holding a tier name — Plus, Flash, Max-Preview — you are in the right place and the roster above is the answer. If you came in holding an id with a date on the end of it, it is a snapshot of one of the aliases, and the alias is the thing to read about.

Availability

All six ids listed above are served on OrcaRouter as of 2026-10-13: qwen/qwen3.6-35b-a3b, qwen/qwen3.6-flash, qwen/qwen3.6-flash-2026-04-16, qwen/qwen3.6-plus, qwen/qwen3.6-plus-2026-04-02, and obsidian/Qwen3.6-35B-A3B. That is the served set and the whole of it; this page makes no claim about any Qwen id outside it. Nothing here is exclusive to us — the two open-weight members are downloadable from the vendor and the hosted tiers are reachable elsewhere — but they are one key away here, which for a generation this size is the entire convenience.