Qwen3.8-Flash-Next does not exist yet, and the most interesting things known about it were published by accident.
On August 25, 2026, a placeholder appeared on ModelScope under the official Qwen organization: Qwen3.8-Flash-Next, releasing “tomorrow.” The page initially carried a detailed highlights section, then was edited down to a shorter teaser. Enough independent visitors captured the original that its contents are consistent across sources, which puts this launch in an unusual state: the specs are officially written and officially unpublished at the same time.
This page does the same job our Fable 5.1 tracker does: separate what is confirmed from what is captured from what is unknown, with dates. When the model actually ships, we update this page the same day.
- Qwen3.8-Flash-Next is teased but NOT released as of August 25, 2026. The official Qwen organization published a placeholder on ModelScope promising release ‘tomorrow,’ meaning August 26 - no hour, no timezone.
- The page initially exposed detailed specs and was then edited down. Captured copies consistently report: 125B main parameters, 6B active per token, an additional 51B of N-gram embeddings, and roughly 1/9 the training cost of Qwen3.7-Plus at comparable capability.
- The teaser’s most strategic line survived the edit: Flash-Next is built on the next-generation Qwen4 architecture, released early ‘to help the community prepare for the upcoming Qwen4 model family.’ This is an architecture preview as much as a model.
- Nothing about context window, license, benchmarks, API pricing, or Hugging Face availability has been published. Any leaderboard numbers circulating before release are unverified.
- The dominant community question is memory: 6B active does not mean it runs like a 6B model, and whether the 51B N-gram table must sit in RAM or can stream from SSD decides who can run it locally.
- This is a living page: when Flash-Next actually ships, we update it the same day with the confirmed model card, license, and real requirements.
- Is Qwen3.8-Flash-Next released?
- No. As of August 25, 2026 it is a teaser page on ModelScope under the official Qwen organization, promising release on August 26. No weights, model card, license, or benchmarks exist yet.
- What are the confirmed specs?
- Officially teased: a multimodal MoE on the next-generation Qwen4 architecture. Captured pre-edit page text adds 125B main parameters, 6B active per token, and 51B of additional N-gram embeddings - high-confidence but formally unconfirmed.
- What will it cost?
- Unknown. No API pricing, no Alibaba Cloud listing, and no confirmed open-weights license as of August 25. The ModelScope location strongly signals an open-weights release.
- Can I run it locally?
- Unknown until the checkpoint ships. 6B active lowers compute per token, but the full expert weights and the 51B embedding table still need to live somewhere. Wait for the config and real quantizations before buying hardware.
§ 01What is officially confirmed
The teaser text that remains live says two things worth quoting.
First, the identity: “Qwen3.8-Flash-Next is a multimodal MoE model built on the next-generation Qwen4 architecture.”
Second, the strategy: “We are releasing these architectural advancements early to help the community prepare for the upcoming Qwen4 model family.”
That second sentence is the real headline. Flash-Next is not just a cheaper Qwen; it is an architecture preview for the next full model family, shipped early so inference runtimes, quantizers, and tool vendors can build support before Qwen4 proper arrives. The last time a “Next” release played this role, it front-ran a major family transition, and the community treated day-zero runtime support as the actual race.
| Claim | Status |
|---|---|
| Official teaser on ModelScope (Qwen org) | Confirmed, live |
| Release date August 26, 2026 | Promised as “tomorrow”; no hour or timezone |
| Multimodal MoE, Qwen4 architecture | Confirmed in live teaser text |
| 125B main parameters, 6B active | Captured pre-edit page text, since removed |
| Additional 51B N-gram embeddings | Captured pre-edit page text, since removed |
| GDN hybrid layers, Qwen Sparse Attention | Captured pre-edit page text, since removed |
| ~1/9 training cost of Qwen3.7-Plus, comparable capability | Captured pre-edit page text, since removed |
| Context window | Unknown |
| License | Unknown (Apache-2.0 plausible, unconfirmed) |
| Benchmarks | None published |
| API availability and pricing | None published |
| Hugging Face mirror | Not present as of the cutoff |
§ 02The removed text, and why it matters
The captured highlights describe a model that is architecturally unusual in three ways.
The parameter accounting is new. 125B main-model parameters with 6B activated per token is a familiar MoE shape. The unfamiliar part is the “additional 51B N-gram embeddings,” a separate lookup-style parameter block that may total the stored model at something like 176B values. Whether that table behaves like ordinary weights (resident in fast memory) or like a cache-friendly lookup structure (streamable from SSD) is unstated, and it is the single spec that decides what hardware this model really needs.
The efficiency claim is aggressive. Roughly one-ninth the training cost of Qwen3.7-Plus at comparable capability, per the removed text, with coding and “cowork” called out as the areas where it is better rather than merely comparable. No benchmark table backs this yet; it is a marketing claim with a deleted timestamp. But “cowork” is a notable word for a model teaser to use: it points at multi-step, agentic, collaborative work as a first-class training target rather than an emergent bonus.
The attention stack is new. GDN hybrid layers and Qwen Sparse Attention are the kind of changes that suggest long-sequence efficiency was a design goal, which is why community guesses about a very large context window are plausible. They are still guesses; no number exists.
§ 03What nobody knows yet
The unknown list is long enough to matter for anyone planning to use the model rather than just read about it: context window, output limits, exact modalities, language coverage, license, Base versus Instruct variants, thinking modes, tool-calling format, expert count and routing, official quantizations, hardware requirements, Hugging Face timing, API availability, and price. Every one of those was unpublished as of August 25, 16:50 UTC.
That did not stop the questions. The dominant thread across Hacker News and LocalLLaMA is not “how smart is it” but “will it fit”: single 5090 owners, dual-3090 builders, Strix Halo and 128GB unified-memory buyers all asking the same thing. The honest answer is that 6B-active tells you about speed, not size. MoE sparsity cuts compute per token; it does not shrink the checkpoint. Until the config, the checkpoint files, and real runtime support exist, every RAM and tokens-per-second prediction is a guess, including the optimistic ones.
The practical read for tool builders: Unsloth has said it is aiming for day-zero support while warning the architecture is genuinely new, and llama.cpp support for novel attention mechanisms has historically taken days to weeks, not hours. If your plan is “run it locally on release morning,” have a fallback.
§ 04Where this sits for buyers
The model layer is having a loud month: an anonymous stealth model soaking up free-window usage, an architecture preview shipping ahead of its own family, and price cuts landing across the frontier tier. The pattern underneath is that models are rotating faster than the work they power. That layer above the models is where we live, so discount accordingly: CellCog is the employee layer, and the point of a standing AI employee is that the rotation does not matter to you. The employee keeps its role, its context, and its continuity; whether the model underneath is a Qwen4 preview or next month’s stealth mystery is an implementation detail that improves on its own schedule. Our harness ranking covers how the layers fit together.
§ 05The tracker
This page updates when facts change, not when threads get louder. As of August 25, 2026: an official teaser exists, a release is promised for August 26, the detailed specs were published and then removed, and nothing about license, context, benchmarks, or price is confirmed. When Flash-Next ships, this page gets the confirmed model card, the real requirements, and an honest read on whether the removed claims survived contact with the release, the same day.
Q1When does Qwen3.8-Flash-Next come out?
The official teaser on ModelScope, published August 25, 2026, promises release ‘tomorrow’ - August 26, 2026. No hour or timezone was stated, and Qwen has not said whether weights, a Hugging Face mirror, and an API endpoint arrive together.
Q2What did the removed ModelScope text say?
Three bullets captured by multiple independent visitors before the page was edited down: a redesigned multimodal MoE with 125B main parameters plus 51B N-gram embeddings and 6B active per token; architectural upgrades across attention, residual, embedding, and optimization, including GDN hybrid layers and Qwen Sparse Attention; and roughly 1/9 the training cost of Qwen3.7-Plus at comparable capability, with stronger coding and cowork performance. Treat these as prematurely published official information, not final specifications.
Q3Is Flash-Next the same as Qwen4?
No, it previews it. The teaser says Flash-Next is built on the next-generation Qwen4 architecture and is being released early so the community and inference runtimes can prepare for the Qwen4 family. The product name stays in the Qwen3.8 line.
Q4Does it replace Qwen3.8-Flash or compete with Qwen3.8-Max?
Neither is established. No deprecation notice, migration document, or comparison against Qwen3.8-Max has been published. The only official capability anchor is the comparison against Qwen3.7-Plus in the removed text.
Q5What hardware will it need to run locally?
Unknown, and worth waiting on. MoE sparsity lowers compute per token, not storage: the full 125B of expert weights plus the 51B embedding table have to live in some mix of VRAM, RAM, and possibly SSD. Whether the N-gram table can stream from disk is the single question that most changes the answer, and only the shipped checkpoint and runtime support will settle it.
Q6Should I build on it the day it ships?
Evaluate on day one, depend on it later. Day-zero support in llama.cpp, vLLM, and friends is hoped for but not guaranteed for a genuinely new architecture, and free or cheap launch access has a way of changing terms - the same advice we gave for the Ox Alpha window.