Qwen 4 Is Not Out Yet: What Is Actually Confirmed

What Alibaba actually confirmed

At the Apsara Conference in Hangzhou on September 22, 2026, Qwen project lead Liu Dayiheng said Qwen 4 is in training on a new-generation architecture and will be released "very soon", according to YottaLabs, and that Qwen 4.5 and Qwen 5 will scale to 5 to 10 trillion parameters, according to the same source and Kie.ai. No date, tiers, or specs were given for Qwen 4 itself at that announcement, according to YottaLabs — though Kie.ai notes Alibaba named four tiers at Apsara: Max, Plus, Flash, and a 27B open-weights model.

That is the full confirmed list, as of October 2026:

• Name: Qwen4-Max, flagship tier, shown at Apsara, 22 September 2026.
• Four-tier roadmap: Max, Flash, Plus, and a 27B open-weights model.
• Training status: In training on a new architecture.
• Release timing: "Very soon" / "coming soon" — no date. One community estimate points to early 2027, and a prediction market put a pre-November launch at 83% on 23 October — but neither is a launch date.
• Roadmap scale: 5 to 10 trillion parameters applies to Qwen4.5 and Qwen5, not Qwen 4. A headline saying Qwen 4 will have 5-10T parameters is wrong.

What does NOT exist: no model card, benchmark, context, or price for any Qwen4 tier, and no release date beyond "coming soon". On the evidence available in October 2026, not one published number describes Qwen4-Max itself.

The preview: Qwen3.8-Flash-Next, the architecture window

What does exist is Qwen3.8-Flash-Next — an open-weight preview of the Qwen4 architecture, released late August 2026 (26 August per Eesel.ai, 28 August per Kie.ai/YottaLabs — likely Hugging Face versus ModelScope dates). It's on Hugging Face and ModelScope under the Qwen Community License 1.0: you download and run it yourself. Qwen3.8-Flash, by contrast, is the production version on QwenCloud, with 1M context and built-in tools — the one you hit over an API. Unlike Qwen 3.7 Flash, which shipped closed with no weights or benchmarks, this time you get open weights and a benchmark table on day one. The specs are the story: a 125B mixture-of-experts model activating only 6B per token, bolted to a 51B N-gram table you can offload to ordinary RAM. That's why it runs fast on hardware that could never hold a dense model its size, and why hosted pricing is just $0.15-0.16 in / $0.47 out per million tokens.

Local AI Zone, based on the official model card and launch blog, gives the fuller architecture:

• 125B backbone (6B active, 512 experts, 10 routed + 1 shared) plus a 51B N-gram table (20M bigram+trigram entries, offloadable to RAM) plus a 4B MTP head — ~180B stored, ~6B computed.
• 48 layers as 12 macro-blocks of 3×(Gated DeltaNet → MoE) → 1×(Qwen Sparse Attention → MoE), wrapped in a 4-branch Gated Residual.
• Hybrid attention: Gated DeltaNet (linear) plus Qwen Sparse Attention, per-layer embeddings, gated residuals, MoE routing.
• Multimodal by default: Text, images, video in; text out; thinking mode on by default.
• 1M context on hosted Flash, and listed for Flash-Next.

The architecture signal is real: lean-activation MoE, a large n-gram memory layer, multimodal by default, 1M context. If Flash-Next is the blueprint, expect much lower active counts relative to total size — cheaper serving, but the same MoE rule: every parameter still needs VRAM. The n-gram layer trades transformer compute for lookup, which matters for cost and factual recall. A traditional transformer computes attention over every token; an n-gram table looks up common sequences from a precomputed table in RAM, so the model doesn't recompute them. That saves compute, but relies on memorised sequences for common phrases — helpful for common facts, risky for novel combinations. It's a trade-off, validated only by independent benchmarks, not a launch blog.

Benchmarks: what Flash-Next claims, and what to trust

Per the Alibaba Cloud blog, Flash-Next-Base (the pre-instruction-tuned checkpoint) with 6B activated tops 8 of 14 base-model benchmarks, including MMLU-Pro, SuperGPQA, BBH, GSM8K, EvalPlus, SWEBench-Pretrain, MGSM, and MMMLU, according to Local AI Zone. It remains close to Qwen3.7-Plus-Base on MMLU, MMLU-Redux, GPQA, MATH, and MultiPL-E, according to the same source. The blog does not enumerate the full 14-benchmark list publicly, so the 8-of-14 claim cannot be independently verified beyond those eight names — a transparency gap worth flagging, according to Local AI Zone.

Independent testing backs the hype up to a point, according to Eesel.ai: Artificial Analysis puts Flash-Next in the top handful of models for intelligence at its price. The catch, repeated across every community thread, is that this is an intentionally under-trained R&D preview, not a finished production model, according to Eesel.ai. GLM 5.3 Flash is the more polished rival if you need something stable today, according to the same source.

That catch matters for anyone thinking of running Flash-Next in production. It's an experimental preview, shipped ahead of the complete Qwen 4 rollout, according to YottaLabs quoting Alibaba. It's under-trained by design, to test architecture, not to be a production model. If you need stability today, use Qwen3.8-Max, Qwen3.8-27B, or a rival like GLM 5.3 Flash — not Flash-Next.

For context on the current Qwen line, according to Kie.ai and Local AI Zone:

• Qwen3.8-Max (August 3, 2026): The 2.4T-total / 95B-active flagship, priced at $2.00 / $6.00 per M tokens, according to Local AI Zone, scoring 58 on the Artificial Analysis Intelligence Index, according to Kie.ai, with a 1M-token context.
• Qwen3.8-27B (August 14, 2026): A dense, hybrid-thinking, Apache-2.0 model that runs in roughly 17GB of VRAM and turned out to be surprisingly good at single-file app generation, according to Local AI Zone.
• Qwen3.8-Flash-Next (August 26-28, 2026): The cheapest SKU in the Qwen3.8 family and the architectural preview for Qwen4, according to Local AI Zone, at roughly $0.15-0.16 / $0.47 per M tokens — roughly 12× cheaper than Qwen3.8-Max on both axes, and substantially cheaper than DeepSeek-V4-Flash and Claude Opus 4.6 Max, according to Local AI Zone.

That price gap — $0.15 versus $2.00 for input — is the point of Flash-Next's architecture. If Qwen 4 inherits that lean-activation design, it could be dramatically cheaper to serve than current flagships, even if it's not the most capable model. Cost, not just capability, is Qwen's bet.

The rumour: Kalmia, the arena checkpoint

As of October 26, 2026, Qwen 4 still hasn't shipped and Alibaba hasn't given a date, according to YottaLabs. An unreleased model codenamed Kalmia is being tested on Arena and testers are attributing it to Qwen 4, but nothing is confirmed, according to the same source. A video deep dive notes the same: an anonymous model nicknamed Kalmia has been building interactive 3D demos in a blind testing arena, and people suspect it is a Qwen 4 checkpoint — but the Kalmia-to-Qwen-4 link is unconfirmed, and that video labels every number as first-party or independent and marks leak figures as reported.

What the leak claims — 3D generation, gaming, and architecture applications — is unverified and not something the confirmed architecture speaks to either way, according to YottaLabs. Benchmarks don't exist yet, obviously; when they arrive they'll be vendor-reported first, and the Qwen 3.8 benchmarks tracker shows how long independent replication takes for this family, typically weeks, according to the same source.

Treat Kalmia as a rumour, not a Qwen 4 benchmark. If Kalmia turns out to be Qwen 4, it would suggest Qwen 4 is strong at interactive 3D and app generation — which would match Qwen3.8-27B's surprising strength at single-file app generation, according to Local AI Zone — but until Alibaba confirms it, Kalmia is an anonymous arena model, not a Qwen 4 model.

Other unconfirmed claims to treat with scepticism, according to Kie.ai's evidence check: whether Qwen4-Max inherits any Qwen3.8-Max specifications — by Alibaba's own framing, it does not inherit by default. Whether the Flash-Next architecture details (176B/6B, 512 experts, N-gram embeddings) transfer to the Max tier — unconfirmed, because Flash-Next is a Flash-tier preview, not a Max-tier model. A Max-tier model will likely be larger, with more active parameters, and possibly a different architecture balance — so don't assume Max will be 6B active just because Flash-Next is.

What to do while you wait for Qwen 4

If you're waiting for Qwen 4, here's what to do now, based on the confirmed evidence:

• If you need an open model today: Use Qwen3.8-27B (Apache-2.0, runs in ~17GB VRAM) for local use, or Qwen3.8-Max via API for flagship capability at $2/$6. Both are live, with model cards and benchmarks, unlike Qwen 4.
• If you want to test the Qwen 4 architecture: Run Qwen3.8-Flash-Next, under the Qwen Community License 1.0, with the understanding that it's an under-trained R&D preview, not a production model. Offload the 51B N-gram table to host RAM, and expect fast, cheap inference at $0.15/$0.47 hosted — but don't expect polished, stable output for production workloads.
• If you need stability today: GLM 5.3 Flash is the more polished rival, according to Eesel.ai, if you need something stable today at a similar price point.
• If you're building for Qwen 4: Build against the architecture signals — lean-activation MoE, n-gram memory, multimodal input, 1M context — but don't hardcode a parameter count, price, or benchmark, because none exists yet. When Qwen 4 ships, check Alibaba's model card first, then independent benchmarks (Artificial Analysis, LMArena) weeks later, before committing to it in production.
• If you see a Qwen 4 benchmark, price, or release date: Ask for the source. If it's from Alibaba's model card, it's confirmed. If it's from a blog, a video, or an arena leak, it's reported or rumoured — and if it's a 5-10 trillion parameter figure, it's about Qwen4.5/Qwen5, not Qwen 4.

Qwen 4 will probably be important — the architecture preview suggests a cheaper, leaner, multimodal model family, with a novel n-gram memory bet, and a four-tier lineup from 27B open weights to Max flagship. But as of October 2026, Qwen 4 is not a model you can download, call via API, or benchmark. It's a name, a training claim, a "very soon" promise, and a preview model that shows you the architecture.

That's not nothing — it's actually more transparency than many unreleased models get, because Alibaba shipped the architecture as an open model you can run today. But it's not Qwen 4, and any article that presents Flash-Next specs, Kalmia demos, or roadmap figures as Qwen 4 specs is misleading you. Wait for the model card. Then judge Qwen 4 on its own numbers.