Qwen 4THE FAMILY GUIDEThe Jade Atlas
Stealth Models ↗UPCOMING · ALIBABA · THE QWEN4 FAMILY
Qwen4-27B, Max, Flash and Plus are named. Qwen4-35B-A3B is the likely fifth. Explore release updates, benchmark forecasts, architecture and VRAM planning.
YOUR GUIDE TO WHAT COMES NEXT
Welcome, curious one. Let us see what the clues have brought.
A curious mind. An unusually distinguished nose.
01 / THE FAMILY
Meet the announced names and the likely small MoE sibling. Each has its own file and a place in the benchmark outlook.
FOLLOWING Qwen4-27B
Conference reporting puts Qwen4-27B first in line on the open-weight side.
Read the conference report ↗A small name
can carry a large idea.
The conference lineup puts 27B on the map; 36Kr reports that it is the first priority on the open-weight side.
The predecessor, Qwen3.8-27B, is a dense multimodal model. The new generation’s name does not yet tell 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.
Flash appears alongside Plus and Max in the conference lineup. 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.
Plus was displayed with Max, Flash and 27B on the Apsara screen. Its place between the familiar tier names makes it a natural model to compare across the five benchmark windows.
The chart treats its position as an editorial forecast anchored to earlier Qwen results. Its eventual price, speed and architecture will make this file more concrete.
Max is the largest-sounding name on the announced family screen. The benchmark atlas places its outlook alongside the frontier comparison models.
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.
Forecast bands use the earlier Qwen models as anchors. Each open diamond sits at the centre of its editorial range.
Family focus: all five forecasts
WINDOW 01
Expert questions across many fields.
Text-only · no tools
Artificial Analysis methodology
Snapshot:
| Model | Score | Setting / forecast basis |
|---|---|---|
| Claude Opus 5.5 (max) | 61.4%measured | Adaptive reasoning · max effort · default fallback Source scorecard |
| Claude Fable 5.1 (max) | 59.1%measured | Adaptive reasoning · max effort · default fallback Source scorecard |
| Claude Opus 5.5 (high effort) | 56%measured | Adaptive reasoning · high effort · default fallback Source scorecard |
| Qwen4-Max | 43–56%forecast | Max 0902 to Opus high. Editorial method |
| Qwen3.8 Max (0902) | 43%measured | Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecard |
| 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 | Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecard |
| 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 | xhigh effort Source scorecard |
WINDOW 02
Graduate science questions.
Diamond · Epoch administered
Epoch AI methodology
Snapshot:
| Model | Score | Setting / forecast basis |
|---|---|---|
| Qwen4-Max | 92–97%forecast | Max 0902 plus zero to five points. Editorial method |
| Qwen3.8 Max (August) | 93%measured | Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecard |
| Qwen3.8 Max (0902) | 92%measured | Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecard |
| 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 | Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecard |
WINDOW 03
Research-level mathematics.
Tiers 1–3 · v2 private set
Epoch AI methodology
Snapshot:
| Model | Score | Setting / forecast basis |
|---|---|---|
| GPT-6 Astra (max) | 93.7%measured | max effort Source scorecard |
| Claude Fable 5.1 (max) | 90.2%measured | max effort Source scorecard |
| Qwen4-Max | 66–94%forecast | Max 0902 to rounded Astra max. Editorial method |
| Qwen3.8 Max (August, xhigh) | 74.7%measured | xhigh effort Source scorecard |
| Qwen3.8 Max (0902) | 66%measured | Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecard |
| 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 35B-A3B | 20%measured | Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecard |
WINDOW 04
Scientific programming problems.
Artificial Analysis implementation
Artificial Analysis methodology
Snapshot:
| Model | Score | Setting / forecast basis |
|---|---|---|
| Claude Opus 5.5 (max) | 66.9%measured | Adaptive reasoning · max effort · default fallback Source scorecard |
| Claude Fable 5.1 (max) | 63.1%measured | Adaptive reasoning · max effort · default fallback Source scorecard |
| Claude Opus 5.5 (high effort) | 60%measured | Adaptive reasoning · high effort · default fallback Source scorecard |
| 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 | Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecard |
| Qwen3.8-Flash-Next | 51%measured | Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecard |
| Qwen4-27B | 47–54%forecast | 27B to three points above Flash-Next. Editorial method |
| Qwen3.8 27B (xhigh) | 47%measured | xhigh effort Source scorecard |
| Qwen4-35B-A3B (likely) | 40–54%forecast | Seven below 27B to three above Flash-Next. Editorial method |
WINDOW 05
Practical work inside a terminal.
Version 4.0 · mini-swe-agent · pass@1, 3 repeats
Artificial Analysis methodology
Snapshot:
| Model | Score | Setting / forecast basis |
|---|---|---|
| Claude Opus 5.5 (max) | 59.6%measured | Adaptive reasoning · max effort · default fallback Source scorecard |
| GPT-6 Astra (xhigh) | 59.6%measured | xhigh effort Source scorecard |
| Claude Opus 5.5 (high effort) | 57%measured | Adaptive reasoning · high effort · default fallback Source scorecard |
| Qwen4-Max | 39–57%forecast | Max 0902 to Opus high. Editorial method |
| Qwen3.8 Max (0902) | 39%measured | Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecard |
| 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 | Evaluator-listed checkpoint; effort not specified in this scorecard. Source scorecard |
| 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 | xhigh effort Source scorecard |
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.
HLE uses Artificial Analysis text-only results without tools. GPQA uses Epoch only. 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.
The case file · updated 28 September 2026
Four names appeared together. A fifth is taking shape in the margins. Follow the dated trail from public announcement to engineering trace to rumor.
Alibaba names its next-generation model. The 5–10 trillion parameter ambition belongs to the later Qwen4.5 and Qwen5 roadmap.
Alibaba's Apsara statement
Inspect the conference photo
An Apsara screen reads “Qwen4 Series Coming Soon” above Max, Flash, Plus and 27B. It gives names, not final model configurations.
Original attendee postQwen 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 releaseTransformers 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 ↗A conference account says the open-weight side is prioritizing Qwen4-27B. The eventual license and model card will settle what ships.
36Kr conference reportA 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 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 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 catalogArchitecture 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.
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
Cast the voices, switch languages and score the scene. Create it all locally in Demodokos Foundry.
Press play. The lab is listening.
Some secrets are difficult to keep.
Master QwenAt the pavilion
Some secrets are difficult to keep.
Master Qwen
Some secrets are difficult to keep.
The lab is listening.
RELEASE WATCH
Qwen4 is in training. Follow the official release channels for weights, API model IDs and the first configuration files.
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 ↗Release dates and availability belong here first. The 22 September Apsara statement is the current family announcement in this guide.
Qwen’s official site ↗Qwen3.8-Flash-Next already offers an open look at the new design. Its published model card and technical report make useful reading while Qwen4 trains.
Read the preview model card ↗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.
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.
KEEP EXPLORING