Kefong Studio 科奉工作室

Qwen4 Status Check: What's Officially Confirmed, What Hasn't Shipped, and Is There a 35B?

·7 min read

Written in October 2026. Every "unconfirmed" item is labeled with its source tier (official press release / on-stage slides / engineering-side evidence / third-party speculation). For model specs and licenses, the official model card is authoritative.

In one sentence: Qwen4 has been publicly named by the Qwen team—but only as "in training": no model card, no weights, no API, no pricing, no benchmarks. The only "Qwen4" you can actually get your hands on and run today is the architecture preview Qwen3.8-Flash-Next, open-sourced on August 26, 2026. Also, there is no 35B among the four official tiers.

What Has Been Officially Confirmed

At the Apsara Conference on September 22, 2026, Liu Dayiheng, head of Alibaba Cloud's Tongyi Qianwen (Qwen), announced:

Claim Status
Qwen4 uses an all-new architecture and is already in training Officially confirmed (per the written press release)
Qwen4.5 and Qwen5 are planned to scale to 5–10 trillion parameters Officially confirmed (as a roadmap)
The current flagship, Qwen3.8-Max, is 2.4T total / 95B active parameters Officially confirmed
Zhenwu V900 AI accelerator: 216GB HBM, 1.2TB/s chip-to-chip, FP8/FP4 support, up to 500,000 cards in a single cluster Officially confirmed; mass production targeted for 2027 Q1, with J900 already on the roadmap for 2028 Q3
Global data center capacity target of 20GW by 2032 Officially confirmed
RSI (recursive self-improvement) is already used in training, inference and chip design; Qwen3.8-Max ran 33 automated iterations in just over a month, raising its Artificial Analysis score from 40 → 45 Official statement

⚠️ The easiest thing to misread: the 5–10T parameter figure applies to Qwen4.5 and Qwen5, not Qwen4 itself.

Not Yet Confirmed / Needs a Question Mark

Claim Status
Four-tier naming: Qwen4-Max / Qwen4-Plus / Qwen4-Flash / Qwen4-27B Comes only from on-stage slides at the conference (all four placards marked Coming Soon); the written press release lists no tier names → unconfirmed
"Releasing soon" The only official wording is "currently in training"—no date whatsoever
Parameters, context length, pricing and benchmarks for the four tiers None published
Qwen4-27B will definitely be open-sourced, under Apache 2.0 again Don't assume this. The previous-generation 27B was Apache 2.0, but the same generation's flagship, Qwen3.8-Max (2.4T), switched to a custom license with an annual revenue threshold, and its sibling Qwen3.8-Flash-Next uses qwen-community-1.0. Alibaba decides licenses release by release

The "Qwen4" You Can Actually Touch Today: Qwen3.8-Flash-Next

Weights were released on 2026-08-26. The official model card describes it, verbatim, as "an experimental preview of the architecture that will underpin Qwen4"—an architecture preview of Qwen4.

Specs

Item Value
Parameters 125B backbone + 51B n-gram embedding + 4B MTP ≈ 180B on disk, with only about 6B active per token
Experts 512 experts, 10 routed + 1 shared active
Context 262,144 tokens native, extendable to 1M with YaRN
Modalities Text / image / video in, text out
License qwen-community-1.0 (not Apache 2.0)

Four architectural changes (which are also the foundation of Qwen4)

  1. Hybrid attention: in every 4 layers, 3 use Gated DeltaNet linear attention + 1 uses QSA (Qwen Sparse Attention). The intuition is "GDN handles long-term memory, QSA handles precise retrieval," avoiding the quadratic blowup of global attention at million-token context;
  2. Gated Residual: widens the residual stream into four parallel branches;
  3. N-gram Embedding: adds capacity through deterministic table lookups instead of relying purely on more parameters;
  4. Optimizer: a Muon + AdamW combination.

Officially reported performance and cost (self-reported, not independently reproduced)

  • QSA kernels at million-token scale: up to 7.6× faster prefill and up to 4.9× faster decode;
  • Training cost roughly 1/9 that of Qwen3.7-Plus (about 90% lower);
  • Claims to beat a certain Claude Opus version on several agent / multimodal benchmarks;
  • Pricing for the production Qwen3.8-Flash API: input ¥1 / million tokens, output ¥3 / million tokens.

Engineering-side evidence (more telling than the launch event)

  • Qwen3.8-Flash-Next's config has model_type=qwen4_exp, and HuggingFace Transformers already contains a models/qwen4_exp reference implementation;
  • vLLM has an unmerged "qwen4 fuse op" PR whose operator list closely matches the Flash-Next architecture;
  • SGLang submitted a PR with native support for the same architecture on August 27;
  • vLLM's Rust frontend submitted a PR on September 1 wiring qwen4_exp into the image input pipeline.

What that last one means: image input support in the next-generation Qwen has gone from "a line on a model card" to something concrete in engineering code.

About 35B: Never Mentioned Officially, but This Lineage Does Have a 35B

Bottom line first: there is no 35B in the official Qwen4 messaging. The only small/mid tier named on the on-stage slides is 27B.

35B shows up in two completely different places:

1. The 35B-A3B in Other Qwen Lines

Model Specs
Qwen3.5-35B-A3B 35B total / 3B active; first to use a 3:1 hybrid of linear attention + full attention, with an officially claimed ~40% inference speedup
Qwen3.6-35B-A3B 35B total / ~3B active; 40 layers, 256 experts top-8 + shared, 262K context, a base model with vision
Qwen-UI-Agent (open-sourced 2026-08-20) Cross-platform GUI agent base model, available in 4B / 27B / 35B-A3B

2. The Community Squeezed 35B-A3B into the Qwen4 Architecture (Unofficial)

This is the only place "Qwen4 + 35B" actually exists, and the part local-inference enthusiasts should pay the most attention to:

Project Key points
logic65/Whittle-Qwen-3.8-35B-A3B (2026-09-19) 35.1B = 25.1B core model + 10B n-gram memory table, ~3B active, architecture is exactly qwen4_exp (same as Flash-Next), distilled from Qwen3.8-27B, runs on stock llama.cpp with no patches
empero-ai/Qwen3.8-35B-A3B-Distill Distills the Qwen3.8 flagships (2.4T-Max and Flash-Next) into the Qwen3.6-35B-A3B MoE architecture; Apache-2.0 (inherited from the base); ships with mmproj, supports image input

The Whittle one has a clever trick: the 10B n-gram memory table is a "lookup," not "compute"—use -ot per_layer_token_embd=CPU to keep it in system RAM and put only the 27B-class core model in VRAM, so VRAM needs are those of a 27B, while generation speed is that of a 3B. Its quantizations are roughly Q8_0 37.8GB / Q6_K 29.3GB / Q5_K_M 25.1GB / Q4_K_M 21.3GB / Q3_K_M 16.7GB.

How Much VRAM Do You Need to Run It Locally Now

The thresholds below apply to the 35B-A3B class (MoE sparsity only saves compute; the full weight file still has to fit in VRAM or RAM):

Quantization Requirement
IQ2_M / Q2_K 16GB VRAM, or 16GB system RAM (the smallest tier that stays coherent)
IQ3_M / Q3_K_M 20–24GB VRAM, or 24GB system RAM
IQ4_XS / Q4_K_M Fully loads on a 24GB GPU; or 32GB RAM CPU-only
Q5_K_M / Q6_K 32GB VRAM, or 48GB system RAM
Q8_0 48GB VRAM, or 64GB system RAM
BF16 80GB+ VRAM (for reference only)

With long contexts, the KV cache becomes the main cost and you may need additional offloading. Also, models with this Qwen3.6/Qwen4 architecture require a recent runtime (llama.cpp / vLLM / SGLang must support Gated DeltaNet MoE; older versions will simply fail to load them).

What to Expect on Release Timing

There is no official date at all. Third parties, extrapolating from Alibaba's roughly five-month cadence from "architecture preview → official release," speculate that Qwen4 could land between late 2026 and 2027 Q1—this is only speculation, not a commitment. Several other things are definitely on the way in the same period: Qwen-Image 3.1, a next-generation video generation model (officially said to be November), plus the already-released Qwen3.8-Omni-Flash (omni-modal, not open-sourced) and the Qwen-Audio-3.1 speech series.

Conclusion

  • Qwen4 is "announced, not released": the only things confirmed are "in training," "new architecture," and "the next generation, 4.5/5, will push to 5–10T parameters"; the tier names, parameters, pricing, license and date are all still up in the air.
  • If you want to get hands-on, use Qwen3.8-Flash-Next now: it is itself the Qwen4 architecture preview, and its specs and engineering support are already public.
  • 35B was never mentioned officially: the officially named small/mid tier is 27B. If you want "Qwen4 architecture + 35B size," your only option is community distills/ports, and you have to accept that they have no official benchmarks and each declares its own license.
  • Watch the license: Alibaba's last two generations have clearly used tiered licensing (custom for flagships, qwen-community-1.0 for some open releases), so assuming "it'll be Apache 2.0 like early Qwen3" is not a safe assumption.

References