--- title: Speculations about Qwen4 | daily.dev description: A speculative deep-dive compiles every fragment of public evidence about Qwen 4 from Alibaba, including vLLM and SGLang engine code targeting a 'qwen4_exp'... canonical: https://daily.dev/posts/speculations-about-qwen4-elqd6an3x twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Speculations about Qwen4 | daily.dev og:description: A speculative deep-dive compiles every fragment of public evidence about Qwen 4 from Alibaba, including vLLM and SGLang engine code targeting a 'qwen4_exp'... og:url: https://daily.dev/posts/speculations-about-qwen4-elqd6an3x og:image: https://api.daily.dev/og/posts/ELQD6an3X.png og:image:alt: Speculations about Qwen4 og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Speculations about Qwen4 **[Qwen](https://daily.dev/sources/qwen)** · [@ch1cane](https://daily.dev/ch1cane) · 10 min read · 3 upvotes · 6 comments ## Summary A speculative deep-dive compiles every fragment of public evidence about Qwen 4 from Alibaba, including vLLM and SGLang engine code targeting a 'qwen4_exp' checkpoint, a leaked stage slide naming Max/Flash/Plus/27B tiers, an invite-only eval API (invite-2609) closing September 30, 2026, and unmerged PRs about a large PLE table used for deployment. It carefully separates confirmed facts (a Flash-Next preview checkpoint exists, Qwen 4 is 'in training') from high-, medium-, and low-confidence rumors, concluding no model card, pricing, weights, or official spec sheet exists yet, and the most likely release window is early 2027 based on past release cadence. ## Content Perplexity Deep Research ---- Qwen 4 still has no model card, API ID, price, or weights. What exists beyond the Flash-Next card is a stage roster that never made the press release, an invite-only eval API, and a large body of  qwen4_exp  engine work aimed at the preview checkpoint. 1. Confirmed but non-officially packaged facts These are not on the public Flash-Next card you already know, and they are not in Alibaba’s September 22 press release. • The written Apsara release says only that Qwen 4 “is currently in training,” then attaches 5–10T to Qwen 4.5 and Qwen 5. It does not name Max, Plus, Flash, or 27B.[alibabacloud] • vLLM’s current docs (fetched October 7, 2026) ship a  qwen4_exp  package:  Qwen4ExpForCausalLM  plus  Qwen4ExpForConditionalGeneration , described as a Qwen3-VL vision tower on the Qwen4Exp language model. The text path is hybrid (Gated DeltaNet plus a separate PLE short-conv spec), supports MTP, and implements both average-pool and gated HyperConnection. Default HC config in that code is 4 streams, low-rank 16.[vllm] • A community teardown of the public FP8 checkpoint, filed as SGLang issue #37792 on September 3, states the served layout as 176B total / 6B active, 48 layers = 36 Gated DeltaNet + 12 QSA, 512 experts with top-10 routing, and a PLE file of 51,200,245,760 bytes ( ple/ple.f8_e4m3.bin ). That issue is about Flash-Next, not a Qwen 4 product tier.[github] • SGLang’s cookbook labels the same checkpoint “176B-parameter (6B active)” and documents single-node TP4 recipes on H200 / B200 / B300 / GB300 and TP8 on MI350X / MI355X.[sglang] • The invite program is first-party, not a leak.  qwen-latest-series-invite-2609  was a closed eval on  https://qweneval.aliyuncs.com/v1 , OpenAI-compatible, model id exactly that string. The page closed September 30, 2026. It does not say “Qwen 4.”[qianwenai] 1. High-confidence rumors Four names on a stage slide, absent from the press release • On September 22, Liu Dayiheng’s talk “Qwen: Toward Real-World Agents” is reported to have shown “Qwen4 Series Coming Soon” with Qwen4-Max, Qwen4-Flash & Qwen4-Plus, and Qwen4-27B.[kocpc] • The same four names are repeated by 36Kr, DataLearner (citing a LINUX DO morning-session note and an X post), Toutiao recaps, and English conference write-ups.[36kr +2] • NYU Shanghai’s recap is the useful brake: a widely shared X post listed those four names, and Alibaba’s published materials do not. Treat the roster as a captured stage claim, not a spec sheet.[rits.shanghai.nyu] • ZDNet’s on-site report, via KOCPC, assigns roles rather than sizes: Max for the capability ceiling, Flash and Plus for efficiency and price, 27B for local deployment. No parameter count, context, modality list, or price was attached.[kocpc] Qwen4-27B as the open-weight track • 36Kr’s English write-up of the speech says Qwen 4 has entered training and “Qwen4-27B will be prioritized for the open source side.”[36kr] • Chinese aggregators repeat that sequencing (“开源侧优先安排 Qwen4-27B”) and cite 36Kr plus Bilibili.[post.smzdm] • This is high-confidence as a conference-recap claim, not as a license or a date. Nothing says Apache-2.0, and nothing says the 27B is dense rather than a small MoE. “Very soon,” with no date that survived contact with October • Multiple Chinese recaps quote “很快发布” / “coming soon.” DataLearner records that phrase and explicitly says it is not a schedule.[kocpc +1] • The July rumor that Qwen 4 would ship at Apsara was graded “not confirmed” after the conference. What shipped that day was a status line, RSI numbers on Qwen3.8-Max, and other modalities.[alibabacloud +1] 1. Medium-confidence rumors Family follows the preview by months, not weeks • CellCog’s tracker uses the only clean precedent: Qwen3-Next in September 2025, Qwen3.5 family on February 17, 2026, about five months. Applied to the August 26 architecture preview, that arithmetic lands in early 2027.[cellcog] • Speculative part: that cadence still holds after Liu took the public lead, and that “very soon” does not mean October. As of October 6–7 write-ups, no Qwen 4 card had appeared.[yottalabs] • DataLearner’s page shows a release date of October 22, 2026. The same page says that date is an inferred convention, not an official plan. Do not treat October 22 as a leak.[datalearner] PLE table size and host-staging are the deployment constraint • OrcaRouter’s September 18 write-up of unmerged SGLang PR #40235 ( [Qwen4-Exp] Add host staging for file-backed PLE , branch  task/ple-host-staged , author Dev-Jahn) says community teardowns put the n-gram table file at 47.7 GiB. The PR author’s own numbers, unreproduced: about 0.5–1 ms extra per decode step versus a pinned table; on an RTX PRO 6000 96 GB, host RAM for the table could be capped from 71 GB pinned down to 6 GB at a 24 GB cap, with concurrency-1 latency moving from 8.62 ms to about 9.1–9.2 ms.[orcarouter] • Speculative part: that Qwen 4 tiers keep this table at this size. The PR targets  model*type=qwen4*exp , which today means Flash-Next. A larger Max could ship a larger table, or Alibaba could fold it into the API so self-hosters never see it. • Related engine work, from the same reporting chain and from vLLM/SGLang issue text: vLLM #53909 “Add qwen4 fuse op” (HyperConnection, QSA, PLE kernels), open since August 26; llama.cpp #27742  qwen4exp  merged August 27; llama.cpp #28136 lazy PLE reads, with one DGX Spark report of roughly 300 tok/s rising to 750–800 tok/s when the table stays on NVMe.[cellcog +2] Invite-2609 is a Qwen 4 checkpoint, not a named tier • The eval page, X posts, and Chinese forums all appeared in the same week as Apsara and all guess “this is the Qwen 4 early build.” The page itself only says “Qwen preview model.”[qianwenai +2] • Speculative mapping to Plus or Max is not supported. The scored tasks are coding, agentic collaboration, and web development, which matches the talk title more than it matches a throughput tier. A Flash-class checkpoint would be an equally legal guess. • Naming:  2608  and  2609  match August and September 2026 invite waves.  qwen-latest-series-invite-2608  has its own closed page at  qianwenai.com/drops/qwen-0828  with the same task list. I found no public diff between the two checkpoints.[qianwenai +1] Tier economics copied from 3.8 • English explainers (Yotta Labs, OrcaRouter, AI Tools Review) fill empty cells from the previous generation: Max as the hosted flagship, Plus as the mid tier, Flash as the cheap high-volume tier, 27B as the download.[yottalabs +2] • The only part with on-site support is the three-track split above. “Plus has the widest multimodality” and any price like “$2 / $6” are extrapolation. Nobody has published a Qwen 4 price. 1. Low-confidence rumors / wildcards • Qwen 4 Coder 32B-A3B, June 2026, Apache-2.0, 82% SWE-bench Verified. Debunked. No such repo exists. LLMCheck says the row was ingested in a July data incident and removed in an August reconciliation. If a chart still shows it, that is the provenance.[cellcog +2] • July 20 “World of AI” leak (Geeky Gadgets): Qwen 4.0 in September 2026, strong 3D coding, stealth codenames “Caleb” and “Terrania Alpha,” claimed to beat GLM 5.2. The September ship date is false. The 3D and codename claims were never corroborated.[yottalabs +1] • Ox Alpha on OpenRouter was an early Qwen 4. Debunked. Z.ai identified it as GLM-5.3-Flash on August 26.[cellcog] • “Qwen 4 Max is 5–10T.” Mis-attached slide. The press release puts that range on Qwen 4.5 and Qwen 5. Reuters’ “up to four times larger than the flagship” is the same 2.4T → 5–10T comparison, not a Qwen 4 size.[alibabacloud +1] • 35B-A3B as the Qwen 4 local shape. A comment under the Apsara coverage asked whether 35B was being killed. That is a reaction to the 27B name, not a spec. A separate August “Qwen3.8-35B-A3B leak” is about the 3.8 line, not Qwen 4.[kocpc +1] • Qwen 4 trained by RSI with no humans. RSI numbers at Apsara (33 cycles, Artificial Analysis 40 → 45, 96% throughput gain on a new GPU, EDA area −42%) are company-reported results for Qwen3.8-Max, not a description of the Qwen 4 run.[alibabacloud +1] 1. qwen-latest-series-invite-2609 What is documented • First-party page title: “Qwen preview-model evaluation, round 2,” status “ended,” deadline September 30, 2026. Tasks: programming, intelligent collaboration, web development.[qianwenai] • Access was application-gated. After approval, callers used an API key against  https://qweneval.aliyuncs.com/v1  with  model="qwen-latest-series-invite-2609" .[qianwenai] • Incentive bands, from the same page: 20 million tokens plus a ¥100 voucher for one real-work case; 50 million tokens plus ¥300 if process logs were attached; 100 million tokens plus an unspecified exclusive reward for a self-built benchmark and full report.[qianwenai] • Chinese posts add “released in three batches.” I could not verify that on the official page.[ai345] • Sibling  qwen-latest-series-invite-2608  uses the same task boilerplate on an earlier drop page. No source I found says 2609 is Max and 2608 is Flash, or the reverse.[qianwenai] What people believe • X and Chinese forums, September 25: “this is probably the Apsara Qwen 4 early build.” That is a timing inference, not a tier mapping.[x +1] • One X post guessed a DeepSeek-style internal-expire codename and expected news in the first week of October. That week has passed with no model card.[x] Anecdotal tester impressions • I found no independent posts of the form “stronger than Qwen3.8-Max on coding” or “worse than DeepSeek-V4 on tools.” The official feedback section on the eval page did not expose scored comments in the fetch. • The only concrete run I found is a September 30 Chinese video recap on techgogogo.com. The author labels the target “suspected Qwen 4,” gives the real id as Invite 2609, and says they are not sure it is Qwen 4. Qualitative notes from that write-up, translated: strict instruction-following had misses; an SVG clock had reversed or wrong hands; a two-cannon SVG was structurally fine but visually off; a browser-OS UI was mostly usable with window-placement bugs; a 3D racing game started but camera and steering were incomplete; a no-library Trello board was fine. The author calls the level “not bad” next to recent Claude Opus 5.5 and GPT 5.1 Sol runs, without scores. Single source, hedged, not reproduced.[techgogogo] 1. Timeline and most likely story • August 26: architecture preview ships as Qwen3.8-Flash-Next,  model*type=qwen4*exp . Engines implement that name within days. Confirmed. • August 26–September 18: the public engineering problem is the n-gram / PLE table, not a secret dense Qwen 4. Host mmap, UVA offload, and file-backed staging all target that table. Confirmed for the preview; unknown for later tiers. • September 22: press release confirms training and the 4.5 / 5 scale path. A stage screen, not the release, is what put Max, Flash, Plus, and 27B in circulation. High-confidence roster, zero specs. • Same week: invite-2609 opens on an Alibaba eval endpoint, aimed at coding and agents, and closes September 30. Official program; Qwen 4 identity is a community guess. • Role split that best matches on-site reporting: Max is the ceiling model, Flash and Plus are the price/latency API pair, 27B is the local/open track and is described as first in the open-source queue. High-confidence as intent, not as a ship order with a date. • Do not copy 125B / 6B / 51B onto Max or 27B. DataLearner’s own caveat is right: Flash-Next specs are a 3.8-generation preview and are not a Qwen 4 tier card.[datalearner] • Most plausible size story, still speculative: Qwen 4 stays on the new stack (GDN + QSA + gated residual + n-gram table) at unknown scale; 5–10T is the generation after that, and it needs the Zhenwu V900 / supernode build-out rather than today’s API footprint.[alibabacloud] • Most plausible window, speculative: not September 2026, not a silent October 7 drop. The five-month preview-to-family precedent points at early 2027, unless “很快” means a 27B-only drop first. Nobody has a dated artifact for either. • What would falsify this picture quickly: a  Qwen/Qwen4-*  repo, a QwenCloud model id, or a tester log that maps invite-2609 onto a named tier. None of those existed in the sources checked through October 7, 2026. ## Community discussion Top comments from developers on daily.dev. **@nerdalytics** · 3 upvotes > Okay Perplexity deed research marked as such. I like that, good thinking. **@nerdalytics** · 3 upvotes > Some speculations by Perplexity are interesting and worrying. The b-gram table hosted only by Alibiba could indeed be possible. Let's see when Qwen4 finally drops. --- Tags: [#llm](https://daily.dev/tags/llm), [#deep-learning](https://daily.dev/tags/deep-learning), [#qwen](https://daily.dev/tags/qwen), [#vllm](https://daily.dev/tags/vllm), [#alibaba](https://daily.dev/tags/alibaba) [View this post on daily.dev](https://daily.dev/posts/speculations-about-qwen4-elqd6an3x) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"DiscussionForumPosting","mainEntityOfPage":"https://daily.dev/posts/speculations-about-qwen4-elqd6an3x","headline":"Speculations about Qwen4","text":"A speculative deep-dive compiles every fragment of public evidence about Qwen 4 from Alibaba, including vLLM and SGLang engine code targeting a 'qwen4_exp' checkpoint, a leaked stage slide naming Max/Flash/Plus/27B tiers, an invite-only eval API (invite-2609) closing September 30, 2026, and unmerged PRs about a large PLE table used for deployment. It carefully separates confirmed facts (a Flash-Next preview checkpoint exists, Qwen 4 is 'in training') from high-, medium-, and low-confidence rumors, concluding no model card, pricing, weights, or official spec sheet exists yet, and the most likely release window is early 2027 based on past release cadence.","url":"https://daily.dev/posts/speculations-about-qwen4-elqd6an3x","datePublished":"2026-10-07T19:21:29.393Z","dateModified":"2026-10-07T19:22:02.214Z","author":{"@type":"Person","name":"Chico","url":"https://daily.dev/ch1cane","interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"EndorseAction"},"userInteractionCount":1150}},"interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":6}],"comment":[{"@type":"Comment","text":"Okay Perplexity deed research marked as such. I like that, good thinking.","datePublished":"2026-10-07T19:28:04.470Z","url":"https://daily.dev/posts/ELQD6an3X#c-bozY9DEiA","author":{"@type":"Person","name":"nerdalytics","url":"https://daily.dev/nerdalytics","image":"https://avatars.githubusercontent.com/u/97166791?v=4"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3}},{"@type":"Comment","text":"Some speculations by Perplexity are interesting and worrying. The b-gram table hosted only by Alibiba could indeed be possible. Let’s see when Qwen4 finally drops.","datePublished":"2026-10-07T19:29:31.336Z","url":"https://daily.dev/posts/ELQD6an3X#c-haGtXwWBS","author":{"@type":"Person","name":"nerdalytics","url":"https://daily.dev/nerdalytics","image":"https://avatars.githubusercontent.com/u/97166791?v=4"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3}}],"isPartOf":{"@type":"WebPage","url":"https://daily.dev/squads/qwen","name":"Qwen"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Qwen","item":"https://daily.dev/squads/qwen"},{"@type":"ListItem","position":3,"name":"Speculations about Qwen4"}]} ```