Qwen 4THE FAMILY GUIDEThe Jade Atlas
Stealth Models ↗MAX IN PREVIEW · ALIBABA · THE QWEN4 FAMILY
Qwen4-Max is out as a preview, built for coding, agents and front-end work. The preview ends on 25 November, and 27B, Flash and Plus are still to come. StealthMark, our test of hard reasoning and visual tasks, now has its first Qwen 4 result.
Max in preview · checked
YOUR GUIDE TO WHAT COMES NEXT
A familiar face. A difficult reflection. Shall we look closer?
A curious mind. An unusually distinguished nose.
OUR STEALTHMARK OUTLOOK
Qwen4-Max-Preview is measured. The gold diamonds are our estimated ranges for the final Max and the three members still to come.
Stealth Models outlook Editorial estimates ·
Every estimate is a range, built from how much earlier Qwen generations improved. Each one starts from a measured model and adds the same gain to its general intelligence, visual intelligence and StealthMark Score. The solid inner border shows the central range; the dashed outer border shows the wider range.
About ten points per version step is the pace Qwen’s 27B line kept from Qwen3.6 to Qwen3.8, as the Qwen climb shows. The preview narrows the range for Max only. It says little about how the smaller models will move, so their ranges stay wide.
Set your own Qwen 4 estimate ↗Measured today: Qwen4-Max-Preview · Qwen3.8-Max · Qwen3.8-27B · Qwen3.8-Flash
QWEN 4 WATCH
On launch day the preview took our benchmark tasks at xhigh reasoning, plus the Latch and Foresight games. Against Qwen3.8-Max it scores higher in visual intelligence, led by Mona Lisa and Selfie, and lower in general intelligence, where both games fell. Our estimate for the final Qwen4-Max now starts from the preview.
Qwen4-Max-Preview vs Qwen3.8-Max ↗The new outlook →Qwen’s main account announced it at 02:01 UTC: “Qwen4-Max is now in preview on Alibaba’s Token Plan.” The API model ID is qwen4-max-preview. Alibaba’s spec lists a one-million-token context, up to 393,216 output tokens, and text, image and video input. It is aimed at coding, agents and front-end work. Early access also runs in QwenWork, Qoder, the Qwen App, QwenStudio and Qwen Chat. The preview ends on 25 November, and this version is retired after it. There are no weights or per-token price yet, and 27B, Flash and Plus are still to come.
Qwen’s post ↗Alibaba’s spec ↗The Qwen4-Max file →A day after the poster, Qoder, Alibaba’s coding platform, launched an anonymous model called Qoduck: “Don’t ask where it came from.” A tokenizer test matches it to Qwen’s tokenizer, and the same tester says it beat every released Qwen on a knowledge check. Its normal Qoder price matches Qwen3.8-Max, and users can switch it to a one-million-token window. It is free each day for individual users in the Qoder app, with no public API. Alibaba has not named it.
Is Qoduck Qwen 4? →Qoder’s announcement ↗Qwen Code v0.26.5 puts every qwen4 model ID on the same reasoning levels as Qwen3.8-Max: low, medium and xhigh, with xhigh as the default. It is the second public release in four days that prepares Alibaba’s own tools for Qwen4, the strongest sign yet that the launch is close.
Read the trace →Pull request #15987 ↗At Alibaba Cloud’s first European Apsara Conference, Qwen’s keynote talk showed a “Qwen4 Series Coming Soon” slide with Max, Flash, Plus and 27B. The lineup matches September’s, and the slide gives no date.
Keynote programme ↗The September slide →Qwen Code v0.26.3 sizes every qwen4 model at one million tokens of context. An Alibaba maintainer asked for the change, gave 1M for the whole family and said the qwen4 IDs are only a matter of time. The release names no model or date.
Read the trace →Pull request #15762 ↗Kerria garbled 3 of 4 tokens that only Qwen models trip over, and none of 52 from 13 other labs. Testers say Kalmia left Arena a day before Kerria arrived. That fits a Qwen model, but other labs build on Qwen’s tokenizer too, and Alibaba and Arena have not said which model it is.
Is Kerria Qwen 4? →Tokenizer test ↗The anonymous model left on 23 October and returned with checkpoints registered that evening and on the 24th. It shares Qwen’s tokenizer quirks, which fits a Qwen model without proving one and says nothing about Qwen 4. Alibaba and Arena have said nothing.
Is Kalmia Qwen 4? →Tokenizer test →A merged Intel-side pull request prepares “Qwen4-27B dense” for one or two GPUs, and Alibaba’s chip unit files a kernel fix “needed for Qwen4.0-Max”. Neither gives a release date.
Read both traces →“big or small?” was followed roughly four hours later by “alright guys, let's do both!” Neither post names a model or launch date.
First post ↗Follow-up post ↗We updated our estimate against the Artificial Analysis frontier.
Estimate and assumptions →The research note gives no launch date.
Read the report →Alibaba names the next generation, and the conference screen shows Max, Flash, Plus and 27B. In the opening keynote, Qwen’s LLM lead, Liu Dayiheng, says the series is about to arrive and gives each model a job. Max pushes the upper limit of intelligence, Flash and Plus aim at efficiency and value for money, and 27B is the one local developers are waiting for.
Announcement →Family →Learned n-gram memory joins sparse attention.
Explore the architecture →OUR STEALTHMARK BENCHMARKS
Qwen 3.8 Max, 27B and Flash set the mark on each benchmark below. Qwen4-Max-Preview is the first Qwen 4 we have measured against it.
THE QWEN CLIMB
Every measured Qwen model on one StealthMark scale, with our Qwen 4 estimates in gold.
StealthMark Score out of 100 · Benchmark results: Qwen4-Max-Preview · Qwen3.8-Max · Qwen3.8-27B · Qwen3.8-Flash ↗
THREE TESTS TO WATCH
Physical state. Recursive reflections. Distortions and missing details. Compare the requirements with what the models actually produced.
The layout is close. Now check what the light is doing.
Physical state & light
Reconstruct a scene whose physical relationships, lighting and surface response must agree.
Only Qwen3.8-Max keeps the physical state mostly consistent. Light and shadow detail stay poorly resolved in all four.
Qwen4-Max-Preview left the physical state unclear. Can the final Qwen 4 get it and the light right together?
Physical & recursive reasoning
Keep the physical state, perspective and recursive reflection geometry consistent within one scene.
Qwen4-Max-Preview draws recognizable first mirror images. All four still leave most of the reflection structure missing, so none rebuilds the mirror room.
Qwen4-Max-Preview drew the first reflections. Can Qwen 4 follow them all the way through?
Recall, proportions & detail
Reconstruct the Mona Lisa’s likeness, anatomy, proportions and fine detail from the model’s own visual recall.
Every one is recognizable. Qwen4-Max-Preview adds developed facial shading. The hands still fall short in all four.
An SVG interpretation of the Mona Lisa tests facial likeness, anatomy and the composition of a familiar portrait.
Qwen4-Max-Preview shaded the face. Can Qwen 4 recover the right proportions and the hands too?
GAME BENCHMARKS · MOVE BY MOVE
Every move comes from the model’s own reasoning, with no code and no tools. The replays show the turn where a plan holds or falls apart, and you can play the same levels for a score calculated the same way.
Coordination
Two robots, one goal. One holds a door open on a pressure plate while the other pushes numbered cores onto their sockets. Cores can be pushed but never pulled, so one careless push can strand a level.
Anticipation
Inspired by Pac-Man. Every step has to account for where the ghosts will be next, not where they are now. Chasing a coin and surviving pull in opposite directions.
Qwen3.8-Max, task by task. Pelican results do not count toward model totals.
Every Qwen3.8-Max result ↗LEAK CHECK · 2 NOVEMBER 2026
No Qwen 4 score now circulating leads to a primary source. The new evidence that does trace back is engineering work, and it names models, not results.
Qwen Code, Alibaba’s open-source coding agent, shipped this rule in release v0.26.3: every model ID that starts with qwen4 gets a one-million-token window. An Alibaba maintainer asked for it, gave 1M for the whole family and said the IDs are only a matter of time. Release v0.26.5 followed on 1 November and gives the same IDs Qwen3.8-Max’s reasoning levels. On 11 November Alibaba’s spec for qwen4-max-preview made both official: one million tokens and the same three reasoning levels, with xhigh as the default.
Qwen Code pull request #15762 ↗ Qwen Code pull request #15987 ↗An Intel-side pull request to a fork of vLLM’s Intel GPU kernels adds INT4 paths for one or two GPUs and FP8 for two. It states that no end-to-end run on a real Qwen4 checkpoint was completed.
Pull request #22 ↗T-Head, Alibaba’s chip unit, adds a Gated DeltaNet head configuration to its kernel library, “needed for Qwen4.0-Max tensor-parallel prefill workloads”.
T-Head issue #11 ↗The anonymous model garbled 6 of 8 tokens that trip up Qwen models and none of 40 from other labs. Other labs build on Qwen’s tokenizer too, so this fits the Qwen family without proving it, and names no model.
Tokenizer test ↗New checkpoints ↗Qwen 4 Max (Preview) scores 53 on Artificial AnalysisAbsent from Artificial Analysis’s leaderboard and posts, and still absent on 11 November, the day Qwen4-Max-Preview launched. The bar is solid although the chart’s own legend stripes models without a public API, and the arrow is drawn by hand.
Qwen4-Coder scores 82% on SWE-bench VerifiedThe page that published it calls the figure illustrative and cites no primary source.
Leaked ModelScope repos and SWE-bench tablesNo primary links, and the tables disagree with each other on model size and ID.
Kalmia’s Terminal-Bench resultsThose numbers come from Tencent’s paper on Qwen3.8-27B.
Internal testers put its front end near Opus 5.5One subjective forum post, reposted with Tieba watermarks. See the screenshot trail.
Qwen 4 ships this weekOne unsourced X post, repeated by news sites.
Qwen 4 lands at the end of OctoberAn anonymous relay on Reddit. October ended without it.
Qwen4-Exp code means a launch is closeThat code supports the released Qwen3.8-Flash-Next, whose configuration declares qwen4_exp. See the code trace.
The case file · in the order it formed
Four names appeared together, and on 11 November the first of them, Qwen4-Max, arrived as a preview. A fifth is taking shape in the margins. Here is the trail from the first line of code to the conference screen and the rumours that followed.
Transformers documents gated residual paths, sparse attention and n-gram features under an experimental architecture name. Code support is a preparation signal, not a final checkpoint ID.
Transformers integration, merged 26 August ↗Qwen calls Qwen3.8-Flash-Next an early preview of architecture for Qwen4. Its 125B main model and 51B n-gram table describe this precursor.
Qwen's architecture releaseAlibaba names its next-generation model, and Qwen’s LLM lead, Liu Dayiheng, tells the keynote audience that the Qwen4 series is about to arrive. Qwen3.8 already grew from Qwen2.5’s 72B flagship to 2.4T parameters in two years. The 5–10 trillion parameter ambition belongs to the later Qwen4.5 and Qwen5 roadmap.
The same announcement describes automated self-improvement research on Qwen3.8-Max. Its 40-to-45 score gain belongs to that predecessor.
Alibaba's Apsara statement
Inspect the conference photo
An Apsara screen reads “Qwen4 Series Coming Soon” above Max, Flash, Plus and 27B. On stage, Liu Dayiheng gave each one a job. The model cards will give the configurations.
Original attendee postA conference account says the open-weight side is prioritizing Qwen4-27B. The eventual license and model card will settle what ships.
36Kr conference reportA CSC Financial (中信建投) research note says “Qwen4发布在即”: “Qwen4 release is imminent.” First Financial’s report, carried by Eastmoney, gives no release date or Alibaba launch confirmation.
First Financial via EastmoneyA secondhand testing claim points to a small A3B model. Qwen4-35B-A3B is the likely name if the 35B tier returns; the reported test did not specify its total size.
Earlier official 35B-A3B card
Inspect the circulating screenshot
Screenshots circulated through a Tieba-to-X repost chain with a Qwen4 label. They show no model selector, API ID or repeatable transcript. Reposts follow one trail.
Circulating screenshot postA released configuration can establish which Qwen4 members use n-gram tables, what “27B” counts, and the final Flash size. Until then, the precursor is a design clue.
Watch Qwen's official model catalogTHE FAMILY
27B, Flash, Plus and Max appeared at Apsara, and Max is the first one out. A likely small MoE successor joins them on our watchlist.
FOLLOWING Qwen4-27B
Qwen’s LLM lead calls 27B the model local developers have been waiting for, and conference reporting puts it first in line on the open-weight side. Kernel code from 23 October calls it dense.
Read the conference report ↗A small name
can carry a large idea.
Qwen’s LLM lead calls 27B the model local developers have been waiting for; 36Kr reports that it is the first priority on the open-weight side.
The predecessor, Qwen3.8-27B, is a dense multimodal model, and kernel code merged on 23 October calls the new one “Qwen4-27B dense”. Neither name tells us whether 27B counts only the backbone or includes a learned lookup table.
A secondhand testing lead points to an A3B model. Qwen4-35B-A3B is the likely name if the 35B tier returns.
A3B describes roughly three billion active parameters per token. That is different from total stored weights. The 35B total follows the earlier tier; the next checkpoint will settle its size.
Alibaba says Flash, like Plus, aims at top efficiency and value for money. The architecture preview makes this an especially interesting branch of the family.
Qwen3.8-Flash-Next has a 125B main model and a separate 51B n-gram table. Qwen4-Flash’s name does not lock in either number, or reserve that technology for Flash alone.
Alibaba pairs Plus with Flash: both aim at top efficiency and value for money.
Its benchmark ranges are forecasts based on earlier Qwen results.
Alibaba says Max keeps pushing the upper limit of intelligence. It is the first Qwen4 model out: a preview since 11 November, built for coding, office work, front-end design and agent tasks. Alibaba will not change it during the preview and retires it afterwards.
It is also the first Qwen 4 we have measured, and our estimate for the final Max starts from the preview. On 23 October, Alibaba’s chip unit named Qwen4.0-Max in kernel code, along with one of its Gated DeltaNet configurations. Alibaba’s 5–10 trillion parameter ambition refers to Qwen4.5 and Qwen5. It is a later roadmap, not a size announcement for Qwen4-Max.
THE BENCHMARK ATLAS
Compare measured frontier and Qwen models with the estimated Qwen4 family. Each test has its own ranking.
MAX OUTLOOK · 8 OCTOBER 2026
Our central estimate is a ten-point gain over Qwen3.8-Max 0902 on the Artificial Analysis Intelligence Index.
The current measured leader is Opus 5.5 at 58. The forecast assumes a frontier of 60 and a Qwen gap of about five points. Measured baseline and frontier.
Qwen3.7-Max to Qwen3.8-Max 0902 rises from 29 to 45 on the current index. That sixteen-point gain includes the September refresh and large improvements on agent tasks. Our next-generation estimate is smaller, with further gains across reasoning, coding and knowledge reliability.
| Qwen release | Qwen | Frontier | Gap |
|---|---|---|---|
| 3.6-Max Preview April · reconstructed · v4.0 | 51.8 | Opus 4.7 max 57.3 | 5.5 |
| 3.7-Max 21 May · v4.0 | 56.6 | GPT-5.5 xhigh 60.2 | 3.6 |
| 3.8-Max 6 August · v4.1 | 56 | Opus 5 max 61 | ≈5 |
Qwen improved while the frontier moved too. Each pair uses its period’s index. April combines the April frontier with a later-confirmed Qwen baseline; May and August use dated AA launch evaluations. April frontier · May comparison · August comparison.
Matching today’s Opus on the four main agent evaluations adds about five index points. Reaching 55 needs broader gains too. A stronger release could reach 58–61; a limited advance could land at 49–52.
Stealth Models estimate, 8 October 2026. The 52–58 band is a working forecast range. It assumes evaluation within about two months on AA v4.3.2; no release date is announced. This external index is separate from the StealthMark General Intelligence and Visual Intelligence outlook above. Index methodology.
Forecast bands use the earlier Qwen models as anchors. Each open diamond sits at the centre of its editorial range. Follow the 27B line from Qwen3.6 (reasoning) through Qwen3.8 (xhigh) to the Qwen4 forecast.
Core comparisons: Astra, Opus 5.5 and Fable 5.1 at max effort. A missing result stays visible below its comparison group.
Family focus: all five forecasts
WINDOW 01
Expert questions across many fields.
Text-only · no tools
Estimated ranges · awaiting measured Qwen 4 results
Artificial Analysis methodology
Baseline: · Comparisons updated:
| Model | Score | Setting / forecast basis |
|---|---|---|
| Claude Opus 5.5 (max) | 61.4%measured | Artificial Analysis · Adaptive reasoning · max effort · default fallback Source scorecardChecked 2026-09-29 |
| Claude Fable 5.1 (max) | 59.1%measured | Artificial Analysis · Adaptive reasoning · max effort · default fallback Source scorecardChecked 2026-09-29 |
| Claude Opus 5.5 (high effort) | 56%measured | Artificial Analysis · Adaptive reasoning · high effort · default fallback Source scorecardChecked 2026-09-28 |
| GPT-6 Astra (max) | 54.7%measured | Artificial Analysis · max effort Source scorecardChecked 2026-09-29 |
| Qwen4-Max | 43–56%forecast | Max 0902 to Opus high. Editorial method |
| Qwen3.8 Max (0902) | 43%measured | Artificial Analysis · Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecardChecked 2026-09-28 |
| Qwen4-Plus | 38–48%forecast | Flash-Next to five points above Max 0902. Editorial method |
| Qwen4-Flash | 38–43%forecast | Flash-Next to Max 0902. Editorial method |
| Qwen3.8-Flash-Next | 38%measured | Artificial Analysis · Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecardChecked 2026-09-28 |
| Qwen4-27B | 34–42%forecast | 27B to four points above Flash-Next. Editorial method |
| Qwen4-35B-A3B (likely) | 28–42%forecast | Six below 27B to four above Flash-Next. Editorial method |
| Qwen3.8-27B (xhigh) | 34%measured | Artificial Analysis · xhigh effort Source scorecardChecked 2026-09-28 |
| Qwen3.6-27B (reasoning) | 23%measured | Artificial Analysis · Reasoning enabled · evaluator-listed setting Source scorecardChecked 2026-09-29 |
WINDOW 02
Graduate science questions.
Diamond · two evaluator groups
Artificial Analysis results and Epoch Qwen anchors use separate evaluator groups. Qwen4 forecasts remain with their Epoch anchors.
Claude Opus 5.5 (max)No published Artificial Analysis result. Model scorecard
Estimated ranges · awaiting measured Qwen 4 results
Epoch AI methodology · Artificial Analysis methodology
Baseline: · Comparisons updated:
| Model | Score | Setting / forecast basis |
|---|---|---|
| GPT-6 Astra (max) | 96.1%measured | Artificial Analysis · max effort Source scorecardChecked 2026-09-29 |
| Claude Fable 5.1 (max) | 93.7%measured | Artificial Analysis · Adaptive reasoning · max effort · default fallback Source scorecardChecked 2026-09-29 |
| Qwen3.8-27B (xhigh) | 90.5%measured | Artificial Analysis · xhigh effort Source scorecardChecked 2026-09-29 |
| Qwen3.6-27B (reasoning) | 84.2%measured | Artificial Analysis · Reasoning enabled · evaluator-listed setting Source scorecardChecked 2026-09-29 |
| Claude Opus 5.5 (max) | —No published result | Artificial Analysis · No published Artificial Analysis result. Source scorecardChecked 2026-09-29 |
| Qwen4-Max | 92–97%forecast | Max 0902 plus zero to five points. Editorial method |
| Qwen3.8 Max (August) | 93%measured | Epoch AI · Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecardChecked 2026-09-28 |
| Qwen3.8 Max (0902) | 92%measured | Epoch AI · Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecardChecked 2026-09-28 |
| Qwen4-Plus | 89–95%forecast | Max 0902 minus three to plus three. Editorial method |
| Qwen4-27B | 87–95%forecast | A3B predecessor plus two to ten. Editorial method |
| Qwen4-Flash | 85–93%forecast | A3B predecessor to August Max. Editorial method |
| Qwen4-35B-A3B (likely) | 80–94%forecast | A3B predecessor minus five to plus nine. Editorial method |
| Qwen3.6-35B-A3B | 85%measured | Epoch AI · Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecardChecked 2026-09-28 |
WINDOW 03
Research-level mathematics.
Tiers 1–3 · v2 private set
Estimated ranges · awaiting measured Qwen 4 results
Qwen3.8-27BNo published Epoch Tiers 1–3 v2 result found. Model scorecard
Claude Opus 5.5 (max)No published Epoch Tiers 1–3 v2 result found. Model scorecard
Epoch AI methodology
Baseline: · Comparisons updated:
| Model | Score | Setting / forecast basis |
|---|---|---|
| GPT-6 Astra (max) | 93.7%measured | Epoch AI · max effort Source scorecardChecked 2026-09-28 |
| Claude Fable 5.1 (max) | 90.2%measured | Epoch AI · max effort Source scorecardChecked 2026-09-28 |
| Qwen4-Max | 66–94%forecast | Max 0902 to rounded Astra max. Editorial method |
| Qwen3.8 Max (August, xhigh) | 74.7%measured | Epoch AI · xhigh effort Source scorecardChecked 2026-09-28 |
| Qwen3.8 Max (0902) | 66%measured | Epoch AI · Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecardChecked 2026-09-28 |
| Qwen4-Plus | 45–78%forecast | A3B predecessor plus 25 to 58 points. Editorial method |
| Qwen4-27B | 35–70%forecast | A3B predecessor plus 15 to 50 points. Editorial method |
| Qwen4-Flash | 30–66%forecast | A3B predecessor plus ten to Max 0902. Editorial method |
| Qwen4-35B-A3B (likely) | 20–70%forecast | A3B predecessor plus zero to 50 points. Editorial method |
| Qwen3.6-27B | 35%measured | Epoch AI · Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecardChecked 2026-09-29 |
| Qwen3.6-35B-A3B | 20%measured | Epoch AI · Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecardChecked 2026-09-28 |
| Qwen3.8-27B | —No published result | Epoch AI · No published Epoch Tiers 1–3 v2 result found. Source scorecardChecked 2026-09-29 |
| Claude Opus 5.5 (max) | —No published result | Epoch AI · No published Epoch Tiers 1–3 v2 result found. Source scorecardChecked 2026-09-29 |
WINDOW 04
Scientific programming problems.
Artificial Analysis implementation
Estimated ranges · awaiting measured Qwen 4 results
Artificial Analysis methodology
Baseline: · Comparisons updated:
| Model | Score | Setting / forecast basis |
|---|---|---|
| Claude Opus 5.5 (max) | 66.9%measured | Artificial Analysis · Adaptive reasoning · max effort · default fallback Source scorecardChecked 2026-09-29 |
| Claude Fable 5.1 (max) | 63.1%measured | Artificial Analysis · Adaptive reasoning · max effort · default fallback Source scorecardChecked 2026-09-29 |
| Claude Opus 5.5 (high effort) | 60%measured | Artificial Analysis · Adaptive reasoning · high effort · default fallback Source scorecardChecked 2026-09-28 |
| GPT-6 Astra (max) | 56.5%measured | Artificial Analysis · max effort Source scorecardChecked 2026-09-29 |
| Qwen4-Max | 52–60%forecast | Max 0902 to Opus high. Editorial method |
| Qwen4-Plus | 51–57%forecast | Flash-Next to five points above Max 0902. Editorial method |
| Qwen4-Flash | 51–54%forecast | Flash-Next to two points above Max 0902. Editorial method |
| Qwen3.8 Max (0902) | 52%measured | Artificial Analysis · Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecardChecked 2026-09-28 |
| Qwen3.8-Flash-Next | 51%measured | Artificial Analysis · Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecardChecked 2026-09-28 |
| Qwen4-27B | 47–54%forecast | 27B to three points above Flash-Next. Editorial method |
| Qwen3.8-27B (xhigh) | 47%measured | Artificial Analysis · xhigh effort Source scorecardChecked 2026-09-28 |
| Qwen4-35B-A3B (likely) | 40–54%forecast | Seven below 27B to three above Flash-Next. Editorial method |
| Qwen3.6-27B (reasoning) | 43%measured | Artificial Analysis · Reasoning enabled · evaluator-listed setting Source scorecardChecked 2026-09-29 |
WINDOW 05
Practical work inside a terminal.
Qwen’s agent roadmap centres on finishing multi-step work: use tools, keep track of progress and recover from errors. These scores measure task success; time, human interventions and cost per completed task are the next useful comparisons.
Version 4.0 · mini-swe-agent · pass@1, 3 repeats
Estimated ranges · awaiting measured Qwen 4 results
Artificial Analysis methodology
Baseline: · Comparisons updated:
| Model | Score | Setting / forecast basis |
|---|---|---|
| Claude Opus 5.5 (max) | 59.6%measured | Artificial Analysis · Adaptive reasoning · max effort · default fallback Source scorecardChecked 2026-09-29 |
| GPT-6 Astra (xhigh) | 59.6%measured | Artificial Analysis · xhigh effort · additional setting Source scorecardChecked 2026-09-28 |
| GPT-6 Astra (max) | 59.1%measured | Artificial Analysis · max effort Source scorecardChecked 2026-09-29 |
| Claude Opus 5.5 (high effort) | 57%measured | Artificial Analysis · Adaptive reasoning · high effort · default fallback Source scorecardChecked 2026-09-28 |
| Claude Fable 5.1 (max) | 52%measured | Artificial Analysis · Adaptive reasoning · max effort · default fallback Source scorecardChecked 2026-09-29 |
| Qwen4-Max | 39–57%forecast | Max 0902 to Opus high. Editorial method |
| Qwen3.8 Max (0902) | 39%measured | Artificial Analysis · Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecardChecked 2026-09-28 |
| Qwen4-Plus | 25–46%forecast | Flash-Next to seven points above Max 0902. Editorial method |
| Qwen4-Flash | 25–39%forecast | Flash-Next to Max 0902. Editorial method |
| Qwen3.8-Flash-Next | 25%measured | Artificial Analysis · Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecardChecked 2026-09-28 |
| Qwen4-27B | 6–29%forecast | 27B to four points above Flash-Next. Editorial method |
| Qwen4-35B-A3B (likely) | 3–31%forecast | Three below 27B to six above Flash-Next. Editorial method |
| Qwen3.8-27B (xhigh) | 6%measured | Artificial Analysis · xhigh effort Source scorecardChecked 2026-09-28 |
| Qwen3.6-27B (reasoning) | 0%measured | Artificial Analysis · Reasoning enabled · evaluator-listed setting Source scorecardChecked 2026-09-29 |
These are our editorial estimates, dated 28 September 2026. Each range starts from a measured Qwen predecessor or the named comparison model. The score tables give the exact anchor and percentage-point adjustment for every estimate. The open diamond sits halfway between the endpoints.
Max spans the latest Max checkpoint to a strong comparison result. Plus assumes a step above Flash-Next. Flash stays near the preview. 27B starts near the existing 27B where a matched result is available. The likely 35B-A3B model gets a wider range because both its final size and performance remain open. GPQA and FrontierMath use Epoch predecessors; their 27B and Flash positions are broader tier assumptions. No cross-benchmark average is calculated. The core comparison set is GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 at max effort. Additional effort settings are labeled; missing results have no plotted point.
HLE uses Artificial Analysis text-only results without tools. GPQA keeps Artificial Analysis frontier and 27B results separate from Epoch Qwen anchors and forecasts. FrontierMath uses Epoch Tiers 1–3 v2. SciCode and Terminal-Bench use Artificial Analysis results. Checkpoint and published effort labels stay attached to every measured entry; a scorecard without an effort label is marked accordingly.
Qwen4-35B-A3B (likely) is a working name for the rumored A3B successor; 35B is a size hypothesis.
Architecture preview · Qwen3.8-Flash-Next
Qwen opened a working architectural precursor in August. It reveals the engineering direction; each finished Qwen4 model still needs its own configuration sheet.
These are Qwen3.8-Flash-Next's published figures. Final Qwen4 family sizes are the next piece of this puzzle.
A normal embedding looks up a vector for one token. The preview also uses the current token and several preceding tokens to select learned vectors for local patterns. The model then uses those vectors while doing its main work.
A short sequence supplies the lookup address.
Host RAM: Qwen says the full table can stay here and the needed entries can be prefetched.
GPU: selected entries move to the accelerator while the main network computes. RAM capacity, transfer bandwidth and concurrency still matter.
“The book stays on the shelf. I bring the useful page to the desk.”
The 51B entries are stored parameters, not 51B more activated matrix multiplications on every token. Qwen's embedding explanation
In the precursor, three of every four layers use Gated DeltaNet to compress history into a fixed-size state. The fourth uses Qwen Sparse Attention to choose useful earlier micro-blocks, limiting long-context work while retaining targeted access.
Gated Residual expands one residual path into four branches. Content-dependent gates control what each layer reads and writes; these are paths inside one model, not four models.
Qwen uses Muon on suitable weight matrices in attention, GDN and experts, while embeddings, the router and low-rank residual weights use AdamW. This is a training method, not an inference setting.
The precursor trains extra prediction steps to improve speculative decoding acceptance. Its MTP layers also use QSA. Final Qwen4 serving paths have yet to be specified.
All four mechanisms are described for Qwen3.8-Flash-Next, the Qwen4 architecture preview. Kernel code from 23 October already ties both halves of GDN + QSA to named Qwen4 models: Gated DeltaNet prefill for Qwen4.0-Max and a Qwen Sparse Attention path for Qwen4-27B.
27B ON YOUR DESK
Start with Q5_K for quality. Reach for IQ3_K when memory is tight. Keep Q8_0 for a roomier GPU and disk. Then choose how much memory to give the conversation.
Qwen4-27B planning guide · Qwen3.8-27B reference sizes and cache layout.
At 128K context, this reference cache drops from 8 GiB in F16 to 2.25 GiB in q4_0. That frees 5.75 GiB while leaving the model’s weight quantization untouched.
The cache estimate uses Qwen3.8-27B’s configuration: 16 full-attention layers, four KV heads and 256 values per head, with both K and V cached. Its recurrent state sits in the runtime allowance. Quantization scales are included: q8_0 uses 34 bytes per 32 values; q4_0 uses 18.
Weight storage approximates full GPU weight residency. Vision components, multiple sessions, larger work buffers or an additional Qwen4 lookup table need their own budget. Host-resident weights and lookup tables consume system RAM. Partial CPU offload can fit a larger file on a smaller GPU, with a speed tradeoff.
Files use decimal GB; memory uses GiB. IQ3_K’s 12–13 GB range allows for tensors stored at higher precision. Final Qwen4 GGUF sizes and its cache layout will replace these reference estimates.
Qwen 27B q4_0 cache quality tests · llama.cpp cache settings
Advertisement Demodokos Foundry
So does Demodokos Foundry. Tap a language and listen: every voice here was generated on a home PC. No credits, no queue, nothing uploaded.
Full access. Cancel anytime.
Now speaking中文Chinese · Tap a language
Now make it say your words. Try Foundry in your language
Ten Foundry narrations, one in each language
Audio could not load. Hear it at Demodokos ↗
RELEASE WATCH
Qwen4-Max came first, as a preview from 11 to 25 November 2026. Qwen4-27B, Flash and Plus have no date yet. Alibaba named all four at Apsara on 22 September, and any open weights, model cards and configuration files will land in the channels below.
Release checked 11 November 2026. QwenCloud lists qwen4-max-preview on Token Plan only, with no pay-as-you-go access or per-token price. OpenRouter, Vercel AI Gateway and models.dev list no Qwen4 model, and Qwen’s Hugging Face organization has no Qwen4 weights or model card. Current API models.
The preview runs in Qwen Chat, the Qwen App, QwenStudio, QwenWork and Qoder. Its API runs on Alibaba’s Token Plan, a monthly subscription for interactive use in compatible AI tools. Alibaba does not allow scripts or app backends on it.
Qwen Chat ↗Qwen’s official model collection is the place to watch for downloadable checkpoints, licenses and the parameter breakdown.
Qwen on Hugging Face ↗The Qwen4-Exp integration exposes the architecture preview. A final model configuration will turn the remaining architecture questions into specifications.
Explore Qwen4-Exp ↗Qwen3.8-Flash-Next already offers an open look at the new design. Its published model card and technical report make useful reading until Qwen4-Flash arrives.
Read the preview model card ↗Qwen4-Max is. Alibaba opened it as a preview on 11 November 2026, through its Token Plan API and in Qwen Chat, the Qwen App, QwenStudio, QwenWork and Qoder. The preview ends on 25 November, and the preview model is retired after that. Qwen4-27B, Qwen4-Flash and Qwen4-Plus have no release date yet, and as of 11 November there are no Qwen4 weights or model card. Alibaba said on 22 September that Qwen4 was in training and about to arrive. By 23 October hardware teams were preparing kernels for Qwen4-27B and Qwen4.0-Max, and on 28 October and 1 November Alibaba’s Qwen Code agent added support for qwen4 model IDs. On 6 November Qwen posted a teaser poster: “The becoming begins.” See where to try it.
Unconfirmed for all three. Kalmia is an anonymous model on Arena, first seen on 21 October. A tokenizer test found Qwen’s quirks in 6 of 8 test tokens and in none of 40 tokens from other labs. Kalmia returned with new checkpoints on 23 and 24 October, and testers say it left around 26 October. Kerria arrived on 27 October and garbled 3 of 4 tokens that only Qwen models trip over, and none of 52 from 13 other labs. Both point to the Qwen family, but other labs build on Qwen’s tokenizer too, and the test cannot tell Qwen 4 from a Qwen3.8 refresh. Their demos are unscored one-shot battles, verdicts differ, and labels such as “Max” or “27B” are testers’ guesses. Qoduck has gathered the most clues so far. It appeared on 7 November on Qoder, Alibaba’s own coding platform, which says only “Don’t ask where it came from.” A tokenizer test matches it to Qwen’s tokenizer, and the tester says it beat every released Qwen on a knowledge check and is likely newer than Qwen3.8. Some users guess Max from its price, others 27B or Flash. Since 11 November, Alibaba lists Qoder among the places with early access to Qwen4-Max, and 53AI reports that Qoduck was an early Qwen4-Max checkpoint. Alibaba, Qwen, Arena and Qoder have not named any of them. See the tokenizer evidence or Qoduck’s test.
We measured Qwen4-Max-Preview on StealthMark on 11 November. Against Qwen3.8-Max it scores higher in visual intelligence and lower in general intelligence, where the two games fell most. See the two side by side. On 11 November, Artificial Analysis and Arena had no score for it, and none of the scores that circulated before the launch traces back to a source. Our estimate keeps the final Max close to the preview. Flash and 27B start from their Qwen 3.8 scores plus about ten points, the pace of earlier Qwen version steps, with a central range of seven to thirteen, and Plus sits halfway between Max and Flash. The Qwen climb and the StealthMark map show where that lands against today’s models. Think we are too cautious or too generous? Set your own Qwen 4 estimate.
No. Kibitz, the anonymous Arena model that appeared beside Kalmia, garbled all 11 Moonshot-specific test tokens, and testers report that it called itself Kimi. Moonshot released Kimi K3.1 Pro on 23 October.
None traces back to a primary source. The “Qwen 4 Max (Preview) 53” chart is absent from Artificial Analysis’s own leaderboard and posts, which still had no Qwen4-Max score on launch day, 11 November. The 82% SWE-bench figure for “Qwen4-Coder” comes from a page that calls it illustrative, and the Terminal-Bench numbers shared beside Kalmia come from a paper on Qwen3.8-27B. Our own Qwen 4 figures on this page are labeled forecasts. Read the full leak check.
Kernel code says so. A pull request merged on 23 October adds kernels for “Qwen4-27B dense”, with INT4 on one or two GPUs and FP8 on two. It comes from Intel-side engineers preparing hardware support rather than from a Qwen model card, and it states that no end-to-end run on a real Qwen4 checkpoint was completed. The predecessor, Qwen3.8-27B, is also dense. See the kernel trace.
Qwen4-Max-Preview reads up to one million tokens and writes up to 393,216, with thinking and answer sharing that output limit. Alibaba published that spec on 11 November, confirming the one million tokens its Qwen Code agent had set for every qwen4 model ID since 28 October. The other members have no published spec yet. Qwen’s Flash-Next architecture preview supports 262,144 tokens natively and reaches one million with extension, and the hosted Qwen3.8-Flash defaults to one million. See Alibaba’s spec.
That has not been established. Qwen documents n-grams in Qwen3.8-Flash-Next, an early Qwen4 architecture preview. No released Qwen4 family configuration assigns them exclusively to Flash or to every member. Read the Qwen preview.
It may continue the Flash role and architectural direction. The published 125B main model plus 51B n-gram table belongs to Qwen3.8-Flash-Next. Qwen has not published a Qwen4-Flash parameter count. See the precursor specifications.
No final model identifier or total size has been published. An A3B test is a secondhand lead; 35B follows the earlier Qwen3.6-35B-A3B naming pattern. It stays in the fifth watchlist position as a likely hypothesis.
Alibaba’s API catalog lists text, image and video input for Qwen4-Max-Preview, with videos up to two hours long, and Qwen Chat also lists audio among its features. The other Qwen4 members have no published spec yet. Qwen3.8-Omni-Flash already understands text, images, audio and video, and works with tools on tasks such as video editing and translation. Qwen’s Omni report.
Start planning around Q5_K_M on a 24 GB-class GPU, IQ3_K on 16 GB, or Q8_0 on 32 GB. The current 27B reference puts Q5_K_M weights at about 18.4 GiB; a q4_0 cache and runtime allowance bring a 32K text session to roughly 21 GiB. These are Qwen3.8-based planning figures for the upcoming Qwen4 model. Compare quants, context and KV-cache memory.
QWEN ON OUR BENCHMARKS