Answer first: Qwen4 is no longer only a speculative name. Qwen now describes Qwen3.8-Flash-Next as a multimodal MoE model and an experimental preview of the architecture that will underpin Qwen4. This confirms an efficiency-first architectural direction, but it does not constitute a Qwen4 product launch. The final model lineup, parameter scale, benchmark scorecard, API identifier, license, pricing, and release schedule remain undisclosed.
What Is Qwen4?
Qwen4 is the next major Qwen architecture now acknowledged by the Qwen team, although no standalone Qwen4 model or product family has been launched. The official Qwen3.8-Flash-Next materials call it an experimental architecture preview for Qwen4. They do not disclose the final parameter count, product lineup, benchmark scorecard, pricing, API identifier, license, or release date. Exact values that cannot be traced to Qwen or Alibaba Cloud should still be treated as unverified.
The most useful way to discuss Qwen4 is still to separate three layers of evidence. Confirmed information now includes current Qwen models, published documentation, and the Qwen3.8-Flash-Next architecture preview. Expected directions are inferences grounded in those signals. Unknowns are final product details that cannot be responsibly filled with estimates. This distinction matters because the preview confirms architectural intent without defining the production Qwen4 family.
| Evidence level | How it should be used | Examples in this article |
|---|---|---|
| Confirmed | State as fact with a direct official link | Qwen3.8-Flash-Next architecture preview; Qwen3.8-Max scale and benchmark baseline |
| Expected | Use cautious language and explain the inference | How the preview architecture may scale into final Qwen4 models |
| Unknown | Do not invent a value or release commitment | Final model lineup, parameters, scorecard, price, license, and API ID |
Why Qwen4 Is the Logical Next Step
The Qwen3 generation established a hybrid approach to reasoning. Its official introduction describes thinking and non-thinking modes that let developers trade response speed against deeper deliberation. The same generation also expanded multilingual support and strengthened tool use, including MCP-oriented agent workflows.
Qwen's roadmap then pointed toward scaling data, extending context, broadening modalities, and environmental-feedback RL. That roadmap matters more than rumor-based release predictions: it frames the next generation as a transition from training models to training agents that can interact with environments and complete long-horizon work.
The Qwen3.8 line now provides two different baselines. Qwen3.8-Max remains the current flagship capability baseline, with a hosted 2.4-trillion-parameter MoE system for coding, office work, visual understanding, long documents, long videos, and autonomous planning. Qwen3.8-Flash-Next plays a different role: Qwen explicitly positions it as the architecture preview for Qwen4. A future Qwen4 must combine the flagship's breadth with the preview's efficiency-oriented design.
What Qwen3.8-Flash-Next Reveals About Qwen4
Qwen3.8-Flash-Next is the first concrete public evidence about Qwen4's foundation. Qwen calls it the first open-weight release under the new architecture and says the design will underpin Qwen4. The preview is therefore stronger than a roadmap inference, while still leaving the production model's scale and configuration open.
The model is a multimodal causal language model with a vision encoder. Its language model contains 125B parameters with 6B activated, plus 51B n-gram embeddings and 4B MTP parameters. It has 48 layers, 512 experts with 10 routed and one shared expert active per token, a native 262,144-token context, and extension support up to 1,000,000 tokens.
| Architecture signal | Confirmed in Qwen3.8-Flash-Next | What it suggests for Qwen4 |
|---|---|---|
| Ultra-sparse MoE | 125B main model; 6B active per token; 512 experts | Capability per unit of active compute may be a central objective |
| Hybrid attention | Three Gated DeltaNet layers for every Qwen Sparse Attention layer | Long-context latency and KV-cache efficiency may matter more than raw window size |
| Gated Residual | Four widened residual branches with data-dependent gates | Qwen is testing more expressive deep scaling with controlled overhead |
| N-gram embedding | 51B bigram and trigram embedding parameters | Memory-offloadable capacity may supplement compute-heavy MoE scaling |
| Native multimodality | Causal language model with a vision encoder | Text and vision are likely architectural foundations, not bolt-on variants |
| Long context | 262K native; extensible to 1M | Qwen4 is likely to emphasize efficient long-horizon agents |
The preview's vendor-reported benchmark results also show why the architecture matters. Despite activating only 6B parameters per token, it improves on the dense Qwen3.8-27B baseline across the selected coding, office-work, tool-use, and multimodal-agent evaluations below. These results are not Qwen4 scores; they are evidence that the new architecture is designed to raise capability per active parameter.
| Benchmark | Qwen3.8-Flash-Next | Qwen3.8-27B |
|---|---|---|
| DeepSWE 1.1 | 58.7 | 42.2 |
| SWE-bench Pro | 62.5 | 61.7 |
| CoWorkBench | 73.9 | 70.7 |
| Toolathlon Verified | 73.5 | 67.1 |
| ClawEval-MM (Pass@3) | 64.4 | 57.4 |
| AndroidWorld | 84.5 | 81.9 |
Interpretation: Flash-Next turns three Qwen4 expectations into stronger signals: ultra-sparse activation, hybrid long-context attention, and native multimodality. It does not reveal Qwen4's final parameter count, exact context configuration, product tiers, or release timing.
Expected Qwen4 Specifications
Because no final Qwen4 specification exists, the table below uses Qwen3.8-Max as the confirmed production baseline and Qwen3.8-Flash-Next as the confirmed architecture signal. The preview increases confidence in several directions, but every final Qwen4 field remains either expected or unknown until Qwen publishes the model.
| Specification | Qwen3.8-Max confirmed baseline | Qwen4 expectation | Confidence |
|---|---|---|---|
| Status | Released | Architecture preview confirmed; model not launched | Confirmed |
| Architecture | Sparse MoE with hybrid attention foundation | Expected to build on Flash-Next's GDN + QSA, gated residual, and n-gram embedding | High |
| Total parameters | 2.4T | Unknown; may not need to exceed 2.4T | Low |
| Active parameters | About 95B per step in the open-weight model | Likely ultra-sparse routing; exact active compute unknown | Medium-high |
| Context window | 1,000,000 tokens | Long-context optimized; final native and default limits unknown | High |
| Maximum output | 131,072 tokens | Likely maintained or expanded | Medium |
| Hosted inputs | Text, image, and video | Multimodal direction is confirmed; exact audio support is unknown | High |
| Output modality | Text | Text is likely; native media output is unconfirmed | Medium |
| Reasoning | Thinking and faster inference workflows | Thinking controls are likely; final API behavior unknown | Medium-high |
| Tool support | Function calling and structured output | Stronger tool selection, preserved state, and recovery expected | High |
| Weights and license | Open-weight flagship checkpoint exists | Preview is open-weight; final Qwen4 license and checkpoints unknown | Medium |
| API model ID | qwen3.8-max | Not available | Confirmed |
Interpretation: Flash-Next raises confidence in ultra-sparse MoE routing, hybrid long-context attention, gated residuals, n-gram embedding, and native multimodality. It does not prove that a final Qwen4 model will use 125B parameters, activate 6B per token, or ship with the same context limits.
What Architecture Is Qwen4 Expected to Build On?
Ultra-Sparse MoE Is Now the Strongest Signal
The architecture question is now less speculative. The Qwen3.8-Flash-Next model card documents 125B main parameters with only 6B activated per token, plus 51B n-gram embeddings. This ultra-sparse design suggests that Qwen4 may prioritize total task capability per unit of active compute instead of expanding active parameter count in proportion to stored capacity.
The important question is not whether Qwen4 contains more experts, but whether routing, memory access, and expert specialization stay stable across long tasks. Flash-Next's n-gram embeddings also show that Qwen is exploring capacity that is cheaper to compute and easier to offload than conventional MoE scaling.
Hybrid Attention Targets Long-Context Efficiency
Flash-Next replaces the earlier Gated DeltaNet and full-attention pairing with Gated DeltaNet and Qwen Sparse Attention operating at the micro-block level. Its 48 layers repeat three Gated DeltaNet blocks followed by one sparse-attention block. This is the clearest sign that Qwen4 may be designed to reduce long-context latency and KV-cache pressure rather than relying on a larger dense-attention window.
Long Context Should Become Agent Memory, Not Just a Larger Prompt
A million-token window is useful only when the model can retrieve the right evidence, preserve goals, and avoid accumulating contradictory state. Qwen4 should therefore be evaluated on how it uses long context across tool calls, file changes, intermediate results, and error recovery. Raw context length is less informative than retrieval accuracy and task completion over a long trajectory.
Reasoning Control Could Become More Granular
Qwen3's hybrid thinking design already allowed fast and deliberate behavior. Flash-Next strengthens the signal by supporting enable_thinking, preserve_thinking, and reasoning_effort controls. A meaningful Qwen4 upgrade could allocate reasoning dynamically and preserve only the state that improves multi-step consistency, reducing total workflow cost rather than simply producing longer visible reasoning.
Multimodality Could Move Closer to Computer Use
Multimodality is now a stronger expectation because both the hosted Qwen3.8-Max service and the open-weight Flash-Next preview support visual input. Qwen4 could combine that perception with more reliable action: understanding a screenshot, selecting a UI control, checking the result, and correcting the plan. Native audio, image generation, speech output, and video generation remain unconfirmed.
Expected Qwen4 Features
- More reliable long-horizon agents. Lower tool-call failure rates, stronger goal retention, better recovery, and more consistent deliverables after hundreds of actions.
- Repository-scale software engineering. Improved planning across large codebases, tests, terminals, issue trackers, and deployment environments.
- End-to-end knowledge work. A stronger path from research and document analysis to spreadsheets, presentations, reports, and decision-ready outputs.
- Multimodal tool use. Visual understanding that informs UI interaction, document operations, and environment-grounded reasoning.
- Adaptive reasoning budgets. Automatic escalation from fast responses to deeper reasoning when uncertainty or task complexity increases.
- Higher capability per active parameter. Better expert routing, n-gram capacity, sparse attention, caching, and post-training rather than scale for its own sake.
- A broader model family. Possible Max, Plus, Coder, small MoE, VL, or Omni variants, although none of these names is confirmed.
Qwen4 Benchmark Outlook: What the Qwen3.8-Max Baseline Shows
There are no Qwen4 benchmark scores. Flash-Next and Max answer different questions: the Flash-Next scorecard tests the new architecture's efficiency-oriented capability, while the official Qwen3.8-Max scorecard remains the stronger production capability baseline across coding agents, research reproduction, professional cowork, visual reasoning, mobile use, and computer use. Qwen4 will need to combine both directions under matched evaluations.
| Benchmark | Qwen3.8-Max reported score | What Qwen4 would need to prove |
|---|---|---|
| SWE-Pro | 67.7 | Close the gap in professional repository-level issue resolution |
| TerminalBench 2.1 | 86.6 | Improve terminal-agent reliability under the same harness |
| PaperBench | 93.0 | Maintain research strength with independent replication |
| FrontierSWE | 73.5 | Convert long-horizon reasoning into more completed engineering work |
| CoWorkBench | 74.8 | Produce more dependable professional deliverables |
| OSWorld-Verified | 86.1 | Reduce visual-action errors in computer-use workflows |
The two baselines point to a dual requirement. Flash-Next shows that low active compute can still deliver strong agentic and multimodal results; Max shows the higher capability ceiling that a new flagship must exceed. Qwen4 should therefore be judged on both matched-task success and total workflow cost, with vendor-reported scores kept separate from independent replication.
Official Qwen3.8-Max benchmark results. Source: Qwen3.8-Max official release**.
Qwen4 vs Current Frontier Models: Confirmed Baselines and Unknowns
The most useful comparison is not simply Qwen4 against Qwen3.8-Max. It is Qwen4 against the design choices already visible in Kimi K3, DeepSeek V4 Pro, and GLM-5.3. The following table compares confirmed current-model attributes with the still-unknown Qwen4 column.
| Dimension | Qwen4 | Qwen3.8-Max | Kimi K3 | DeepSeek V4 Pro | GLM-5.3 |
|---|---|---|---|---|---|
| Status | Architecture preview confirmed; model unreleased | Released | Released | Released | Released |
| Scale | Unknown | 2.4T / about 95B active | 2.8T / 16 of 896 experts | 1.6T / 49B active | Not disclosed for this release |
| Context | Expected 1M+ | 1M | 1M | 1M | 1M |
| Input | Multimodal direction confirmed; final interface unknown | Text, image, video | Text and native vision | Text-oriented | Text only |
| Reasoning | Unknown | Thinking and fast workflows | Low / high / max effort | Thinking / non-thinking | Always on; low / high / max |
| Open ecosystem | Final weights and license unconfirmed | Open-weight flagship exists | Open-source positioning | Open weights | Open-source positioning |
| Core focus | Expected efficient multimodal agent | Coding, cowork, visual agents | Coding and knowledge work | Efficient reasoning and coding | Coding and cybersecurity |
Model Scale and Inference Efficiency
Kimi's documentation describes 2.8T parameters and 16 activated experts out of 896. DeepSeek V4 Pro takes a different route with 1.6T total and 49B active parameters. Qwen4 does not need to be the largest model to win this comparison; it needs to convert active compute into more reliable completed work.
Context and Multimodality
One million tokens has become a common flagship baseline, so context length alone will not differentiate Qwen4. Qwen's stronger opportunity is multimodal context use: finding the correct evidence across documents, code, screenshots, and video, then taking the right action. Kimi K3 also has native visual understanding, while GLM-5.3's current interface is text-only with a 1M context and 128K maximum output.
Coding, Agents, and Specialized Strengths
Qwen3.8-Max brings a broad coding, research, cowork, and visual-agent profile. Qwen3.8-Flash-Next adds an efficiency-oriented multimodal architecture with stronger capability per active parameter. Kimi K3 emphasizes long-horizon coding and knowledge work, DeepSeek V4 Pro emphasizes efficient reasoning and agentic coding, and GLM-5.3 adds a distinctive cybersecurity focus. A credible Qwen4 launch would need to combine broad agent reliability with specialist-level software-engineering and visual-computer-use performance.
Open Weights Are Not the Whole Deployment Story
Open weights can improve control, customization, and provider choice, but the practical comparison also includes license terms, hardware requirements, quantization quality, fine-tuning support, and the availability of optimized serving stacks. The word open should not be used as a substitute for checking the exact license and deployment conditions of each release.
Comparison result: Qwen4's strongest potential position is a broad multimodal agent that combines Qwen3.8-Max's visual and cowork strengths with Flash-Next's low-active-compute architecture and better software-engineering reliability. It cannot be declared the winner until matched evaluations and production behavior are available.
Potential Use Cases for Qwen4
Autonomous Software Engineering
A stronger Qwen agent could inspect a repository, reproduce an issue, edit multiple files, run tests, review failures, and prepare a pull-request-ready change. The important metric would be the percentage of tasks completed without human rescue, not the amount of code generated.
Enterprise Knowledge Work
Long-context and multimodal understanding could support workflows that begin with contracts, PDFs, spreadsheets, diagrams, and meeting records and end with a decision memo, analysis, or structured action plan. Enterprises would still need retrieval controls, audit logs, and source verification around the model.
Multimodal Computer-Use Agents
Qwen4 could be useful where an agent must interpret screenshots, navigate applications, verify UI state, and recover when a click or form submission produces an unexpected result. This is a natural extension of Qwen3.8-Max's strong visual and OSWorld-style baseline, but native computer-control support has not been announced.
Scientific and Engineering Research
Research workflows combine literature review, code, simulation, data analysis, and report writing. A future Qwen model could orchestrate these steps and maintain a longer experimental history. PaperBench performance makes this a credible direction, but reproducibility and independent verification would remain essential.
What We Don't Know Yet
Although Qwen has acknowledged the architecture through Flash-Next, the following production details remain unavailable and should stay explicitly unresolved until an official Qwen4 announcement:
- The exact product names and whether Qwen4 launches as one model or a family.
- The release date or preview schedule.
- Whether the final models retain Flash-Next's exact GDN/QSA ratio, gated residual design, n-gram embedding scale, and expert routing.
- Total parameters, active parameters, expert count, and training data scale for each Qwen4 model.
- Native context length, default context, maximum output, and supported modalities for each route.
- Native audio, image-generation, speech, or video-generation capabilities.
- Official benchmark results and the evaluation harness used for each score.
- Open-weight availability, license terms, and commercial-use conditions.
- API pricing, regional availability, rate limits, and the final model ID.
Verification rule: Treat any exact Qwen4 specifications, benchmark charts, pricing tables, or API identifiers as unverified unless they can be traced directly to Qwen, Alibaba Cloud, or an official model repository.
When Will Qwen4 Be Released?
There is no defensible public release date. Qwen3.8-Flash-Next confirms that Qwen is testing the architecture that will underpin Qwen4, but the public materials provide no schedule for a production Qwen4 model. Training completion, serving efficiency, safety evaluation, weight licensing, and product integration could all affect the timing and shape of the release.
The safest publication approach is to avoid a month or quarter prediction. Readers should watch the official Qwen site, Alibaba Cloud Model Studio documentation, verified model repositories, and the CometAPI model catalog. A Qwen4 model page should be added only after the model is actually listed.
How Developers Can Prepare for Qwen4
- Establish two baselines. Use Qwen3.8-Flash-Next to measure architecture efficiency and Qwen3.8-Max to measure current flagship capability.
- Separate model IDs from business logic. Keep routing and configuration outside application code so models can be swapped safely.
- Build task-level evaluations. Measure completed software fixes, accurate research outputs, successful tool chains, and usable deliverables.
- Log total workflow cost. Track latency, input and output tokens, cache use, retries, failed actions, and human rework.
- Preserve fallback routes. Use model routing and provider fallbacks instead of making one unreleased model a single point of failure.
- Do not invent a model ID. Wait for the official identifier before adding qwen4 or qwen4-max to production configuration.
Try Qwen3.8-Flash-Next and Qwen3.8-Max Through CometAPI
Developers can now test Qwen3.8-Flash-Next through CometAPI and keep Qwen3.8-Max as the flagship comparison. Create an API key, keep it in an environment variable, and call the existing model through an OpenAI-compatible client. The example intentionally uses qwen3.8-flash-next; it does not assume a future Qwen4 identifier.
Conclusion
Qwen4 is now more than a rumor: Qwen has explicitly identified Qwen3.8-Flash-Next as an experimental preview of the architecture that will underpin it. The preview confirms a multimodal, ultra-sparse MoE direction built around Gated DeltaNet, Qwen Sparse Attention, gated residuals, and n-gram embeddings. Qwen3.8-Max remains the stronger current capability baseline.
The real test for Qwen4 will not be whether its parameter count is larger. It will be whether the final model can combine Flash-Next's efficiency with Max-level capability, complete longer work with fewer failures, and use multimodal evidence more reliably. The architecture direction is now public, but the final specifications, benchmarks, license, price, API ID, and release date remain unknown.