Qwen3.8-Flash-Next (2026): The Complete Guide to Qwen's Qwen4-Preview Architecture Model
🎯 Core Takeaways (TL;DR)
- Qwen3.8-Flash-Next is Alibaba Qwen's open-weight preview of the architecture that will underpin Qwen4 — a 125B-parameter MoE model that activates just 6B parameters per token (plus a 51B n-gram embedding), delivering frontier-adjacent intelligence at a fraction of the compute cost.
- The headline is efficiency: Qwen3.8-Flash-Next scored DeepSWE 1.1 at 58.7, SWE-bench Pro at 62.5, and 91.9 on LiveCodeBench v6 — beating Claude Opus 4.6 Max on most benchmarks despite being a "mid-tier Flash" release. Training cost came in at roughly one-ninth of Qwen3.7-Plus.
- Qwen3.8-Flash-Next introduces three genuinely new architectural building blocks — Qwen Sparse Attention (QSA), Gated Residual, and N-gram Embedding — rethinking how the core components of a modern LLM interact at scale.
- Qwen3.8-Flash-Next ships with a native 262,144-token (256K) context window, extensible to 1M tokens via YaRN, with Qwen reporting prefill up to 7.6× faster and decode up to 4.9× faster at the 1M-token mark. The official hosted version, Qwen3.8-Flash, adds 1M context by default and built-in tools.
Table of Contents
- What Is Qwen3.8-Flash-Next?
- Qwen3.8-Flash-Next Architecture & Specs
- Qwen3.8-Flash-Next Benchmarks: The Efficiency Leap
- Why "Flash" Matters: Qwen3.8-Flash-Next Efficiency
- Running Qwen3.8-Flash-Next (Deployment & API)
- Qwen3.8-Flash-Next vs the Competition
- FAQ
- Summary & Recommended Actions
What Is Qwen3.8-Flash-Next?
Qwen3.8-Flash-Next is the newest open-weight release from Alibaba's Qwen team, and it marks a pivotal shift in the company's roadmap. Unlike Qwen3.8-Max (the 2.4T-parameter flagship) or Qwen3.8-27B (the 27B dense model), Qwen3.8-Flash-Next is the first open-weight model built on a fundamentally reworked architecture — what the model card calls "an experimental preview of the architecture that will underpin Qwen4."
The name tells the story. "Flash" is Qwen's mid-tier, efficiency-first brand: a model that trades some raw ceiling for dramatically lower cost and faster inference. "Next" signals that this is the architectural blueprint for the coming generation. In Qwen3.8-Flash-Next, the old pairing of Gated DeltaNet with Gated Attention has been reworked into Gated DeltaNet plus Qwen Sparse Attention (QSA), alongside two other brand-new components: Gated Residual and N-gram Embedding.
The result is a model that Qwen positions as delivering "most of the reasoning and coding capability of its larger Qwen3.8 line while running at a fraction of the compute cost." As the official announcement frames it: the question is no longer how much we can scale, but how efficiently we can do so.
💡 Professional Tip: Qwen3.8-Flash-Next is available on Hugging Face at
Qwen/Qwen3.8-Flash-Nextunder the qwen-community-1.0 license (open weights, not fully permissive Apache 2.0). For managed inference, the official Qwen3.8-Flash on Qwen Cloud is the production-hardened version with 1M context by default and official built-in tools.
Architecture and Specs
Here's what's under the hood of Qwen3.8-Flash-Next:
| Spec | Qwen3.8-Flash-Next |
|---|---|
| Model type | Causal LM with vision encoder |
| Total parameters | 125B MoE + 51B n-gram embedding + 4B MTP |
| Activated parameters per token | 6B |
| Mixture of Experts | 512 experts (10 routed + 1 shared activated), expert intermediate dim 640 |
| Hidden dimension | 2,560 |
| Number of layers | 48 |
| Hidden layout | 12 × (3 × (Gated DeltaNet → MoE) → 1 × (Qwen Sparse Attention → MoE)) |
| Vision | Yes (image + text input, native multimodal) |
| Native context | 262,144 tokens (256K) |
| Extensible context | Up to 1,000,000 tokens (YaRN) |
| License | qwen-community-1.0 |
The defining feature of Qwen3.8-Flash-Next is that it's not just a bigger or smaller model — it's an architectural redesign. Three new mechanisms stand out:
1. Hybrid Attention with Qwen Sparse Attention (QSA)
The biggest change is in how Qwen3.8-Flash-Next handles attention. Instead of selecting individual tokens, QSA operates at the micro-block level — it picks entire blocks of tokens to attend to, using a lightweight indexer (MQA with 4 query heads and 1 shared key head, 128-dim). This dramatically cuts long-context latency, which matters as agentic workloads — tasks with huge tool-call histories and long codebases — dominate real usage. Budget: 512 blocks, or 2,048 tokens.
2. Gated Residual
Deep LLM training is only manageable because of residual streams with normalization. Qwen3.8-Flash-Next refines this with a Gated Residual that modulates information flowing through widened residual streams via an element-wise, data-dependent read gate and a per-branch scalar write gate (4 branches, bottleneck rank 320). The result is finer-grained expressiveness across layers while preserving training stability and keeping inference overhead low.
3. N-gram Embedding
Qwen3.8-Flash-Next scales parameters along a new axis: embeddings rather than expert weights. By indexing with short n-grams (20M bigrams/trigrams at layer 2, with 51B of embedding parameters), the model achieves parameter scaling that requires less computation and is far more amenable to memory offloading than a pure MoE — highly efficient for memory-constrained accelerators. This is why Qwen3.8-Flash-Next can hold 125B+ parameters while activating only 6B.
4. A Tailored Training Recipe
Training memory and cost were optimized with a tailored recipe that applies Muon and AdamW optimizers to specific weight categories. Guided by refitted scaling laws, Qwen eliminated traditional batch-size warmups and started directly at the target batch size — substantially reducing optimizer steps while safely supporting larger learning rates. This is a large part of how training cost dropped to roughly one-ninth of Qwen3.7-Plus.
✅ Best Practice: For agentic work with Qwen3.8-Flash-Next, use thinking mode (on by default) with
reasoning_effort="xhigh"for complex planning. But remember: higherreasoning_effortisn't always slower overall — lower effort can cause insufficient analysis, more failures, and retries that inflate total latency and token spend.
Benchmarks
The reason momentum around Qwen3.8-Flash-Next built so quickly is the shape of its benchmark curve: a "Flash" (mid-tier) model beating a frontier flagship on most tests. Here are the official language benchmarks, with comparison to prior Qwen models, DeepSeek-V4-Flash-0731, and Claude Opus 4.6 Max:
| Benchmark | Qwen3.8-Flash-Next | Qwen3.8-27B | Qwen3.7-Plus | DeepSeek V4 Flash 0731 | Opus 4.6 Max |
|---|---|---|---|---|---|
| Activated params | 6B | 27B | 17B | 13B | – |
| DeepSWE 1.1 | 58.7 | 42.2 | 16.5 | 54.4 | – |
| SWE-bench Pro | 62.5 | 61.7 | 55.8 | 56.0 | 53.4 |
| SWE-bench Multilingual | 81.0 | 73.8 | 75.8 | – | 77.5 |
| NL2Repo-Bench | 48.1 | 42.3 | 41.1 | 54.2 | 47.6 |
| CoWorkBench | 73.9 | 70.7 | 65.1 | 45.1 | 68.2 |
| JobBench | 55.7 | 33.4 | 27.6 | 41.3 | 36.6 |
| Agents' Last Exam (score) | 51.2 | 42.9 | 33.6 | – | – |
| Toolathlon Verified | 73.5 | 67.1 | 50.6 | 70.3 | – |
| IFBench | 81.3 | 79.5 | 79.1 | 79.2 | 62.5 |
| GPQA Diamond | 91.7 | 89.2 | 90.3 | 90.8 | 91.3 |
| HLE | 35.9 | 30.8 | 34.7 | 33.8 | 40.0 |
| LiveCodeBench v6 | 91.9 | 90.3 | 89.6 | 90.6 | 88.8 |
The message is unmistakable. Qwen3.8-Flash-Next leads on DeepSWE 1.1 (58.7), SWE-bench Pro (62.5), SWE-bench Multilingual (81.0), CoWorkBench (73.9), JobBench (55.7), Toolathlon Verified (73.5), and LiveCodeBench v6 (91.9) — all with only 6B activated parameters. Claude Opus 4.6 Max maintains its edge on GPQA Diamond (91.3) and Humanity's Last Exam (40.0), and DeepSeek-V4-Flash wins NL2Repo-Bench (54.2).
Vision-Language Benchmarks
Qwen3.8-Flash-Next is also a native vision-language model. On multimodal agentic and general intelligence tests:
| Benchmark | Qwen3.8-Flash-Next | Qwen3.8-27B | Qwen3.7-Plus | Opus 4.6 Max |
|---|---|---|---|---|
| ClawEval-MM (Pass@3 / avg) | 64.4 / 60.4 | 57.4 / 56.9 | 57.4 / 60.1 | 52.5 / 54.7 |
| RecreationBench | 49.9 | 47.1 | 30.2 | – |
| AndroidWorld | 84.5 | 81.9 | 81.0 | 62.0 |
| OSWorld 2.0 (binary / partial) | 19.4 / 52.3 | 19.4 / 48.0 | 2.8 / 21.5 | – |
| Vision2Web | 64.0 | 62.9 | 42.1 | – |
| ERQA | 72.3 | 65.5 | 69.8 | 40.8 |
| LVBench (long video) | 76.6 | 72.4 | 76.2 | 63.0 |
| RealWorldQA | 88.5 | 85.9 | 86.9 | 73.9 |
| MathVision (with CI) | 95.7 | 94.6 | 88.7 | 65.5 |
| CharXiv RQ (with CI) | 90.6 | 90.2 | 85.9 | 66.0 |
Qwen3.8-Flash-Next beats Claude Opus 4.6 Max across computer use, mobile use, long-video understanding, embodied perception, and visual math — all from a model that activates just 6B parameters.
⚠️ Attention: These are Qwen's own reported numbers, not yet independently verified. Treat the deepest benchmark claims as directional until third-party leaderboards (like a DeepSWE or SWE-bench Pro leaderboard) confirm them.
Why "Flash" Matters
Positioning matters as much as raw scores. Qwen3.8-Flash-Next is deliberately priced and sized as a mid-tier, efficiency-first model — not a flagship. That framing makes its benchmark results genuinely surprising: a cheap, low-activation model taking the lead on the agentic-coding and long-horizon benchmarks that enterprises care most about.
The economics are the story. Qwen claims training Qwen3.8-Flash-Next cost roughly one-ninth of Qwen3.7-Plus, while beating it on nearly every benchmark. This is the same efficiency direction Qwen took with Qwen3.8-27B a few weeks earlier — position a smaller/cheaper model against far larger rivals and still come out competitive. Flash-Next takes the argument further by redesigning the architecture itself, so the savings come from how the model computes, not just how many parameters it has.
For anyone paying for API tokens or running their own GPUs, Qwen3.8-Flash-Next's activation ratio (6B active out of 125B) is the practical headline: most of the intelligence, a fraction of the compute.
💡 Professional Tip: If your workload is dominated by long-context agent tasks — large tool-call histories, chat logs, multi-file codebases — Qwen3.8-Flash-Next's QSA and 256K→1M context extension are purpose-built for you. Qwen reports prefill up to 7.6× faster and decode up to 4.9× faster at the 1M-token mark than the prior architecture.
Running Qwen3.8-Flash-Next
Getting Qwen3.8-Flash-Next running is straightforward. Qwen recommends API-based integration for most users, and for self-hosting it ships as Transformers-compatible weights usable with SGLang, vLLM, and TokenSpeed.
Supported Inference Frameworks
| Framework | Notes |
|---|---|
| SGLang | Qwen3.8-Flash-Next Cookbook — recommended for production/high throughput |
| vLLM | Qwen3.8-Flash-Next Recipe |
| TokenSpeed | Qwen3.8-Flash-Next Recipe |
| KTransformers | Recommended for high-throughput serving |
| Qwen Cloud | Official hosted Qwen3.8-Flash — 1M context by default + built-in tools |
[!Important]
Inference efficiency and throughput vary significantly across frameworks. Qwen strongly recommends using the latest framework versions, and for production or high-throughput workloads, dedicated serving engines such as SGLang, KTransformers, or vLLM.
API Usage
Qwen3.8-Flash-Next operates in thinking mode by default, generating reasoning content (signified by \n...\n\n) before the final answer. You can control thinking behavior with enable_thinking, preserve_thinking, and reasoning_effort (levels: xhigh, medium, low).
Recommended sampling parameters:
-
Thinking Mode:
temperature=1.0,top_p=0.95,top_k=20,min_p=0.0,presence_penalty=0.0,repetition_penalty=1.0 -
Instruct (non-thinking) mode:
temperature=0.7,top_p=0.80,top_k=20,min_p=0.0,presence_penalty=1.5,repetition_penalty=1.0
Text-only example (OpenAI-compatible API):
from openai import OpenAI
client = OpenAI() # set OPENAI_BASE_URL and OPENAI_API_KEY
completion = client.chat.completions.create(
model="Qwen/Qwen3.8-Flash-Next",
messages=[{"role": "user", "content": "Write a Python function to merge two sorted linked lists."}],
extra_body={
"chat_template_kwargs": {
"enable_thinking": True, # on by default
"preserve_thinking": True, # on by default
},
},
reasoning_effort="xhigh", # xhigh by default; levels xhigh, medium, low
stream=True,
stream_options={"include_usage": True},
)
Qwen3.8-Flash-Next also accepts image input through the standard OpenAI image_url content type, making it a drop-in native vision-language model for the Chat Completions API.
Comparison
How does Qwen3.8-Flash-Next stack up against the models it's being compared to?
| Model | Total params | Activated | DeepSWE | SWE-bench Pro | Local-friendly | Vision |
|---|---|---|---|---|---|---|
| Qwen3.8-Flash-Next | 125B | 6B | 58.7 | 62.5 | ✅ (efficient) | ✅ native |
| Qwen3.8-27B | 27B dense | 27B | 42.2 | 61.7 | ✅ (consumer GPU) | ✅ |
| DeepSeek V4 Flash 0731 | 284B | 13B | 54.4 | 56.0 | ⚠️ (larger) | ❌ |
| Qwen3.7-Plus | 397B | 17B | 16.5 | 55.8 | ⚠️ | ✅ |
| Claude Opus 4.6 Max | frontier | – | – | 53.4 | ❌ | ✅ |
The clearest takeaway: Qwen3.8-Flash-Next beats not only its bigger, more expensive sibling Qwen3.7-Plus, but also the frontier Claude Opus 4.6 Max on the majority of shared benchmarks — while activating a mere 6B parameters. Its main losses are narrow: DeepSeek-V4-Flash on NL2Repo-Bench, and Claude on the hardest frontier-reasoning tests (GPQA Diamond and HLE). For agentic coding and long-horizon work, Qwen3.8-Flash-Next is currently the efficiency king.
FAQ
Q: What is Qwen3.8-Flash-Next?
A: Qwen3.8-Flash-Next is Alibaba Qwen's open-weight preview of the Qwen4 architecture — a 125B-parameter Mixture-of-Experts vision-language model that activates only 6B parameters per token, plus a 51B n-gram embedding. It introduces Qwen Sparse Attention, Gated Residual, and N-gram Embedding, and targets coding, agent, and long-context workloads with high efficiency.
Q: How many parameters does Qwen3.8-Flash-Next activate?
A: Only 6B parameters are active per token out of 125B total, with 51B additional n-gram embedding parameters and 4B MTP. This low activation ratio is the core efficiency design of Qwen3.8-Flash-Next.
Q: How good is Qwen3.8-Flash-Next on benchmarks?
A: Qwen3.8-Flash-Next scores DeepSWE 1.1 at 58.7, SWE-bench Pro at 62.5, SWE-bench Multilingual at 81.0, Toolathlon Verified at 73.5, and LiveCodeBench v6 at 91.9 — beating Claude Opus 4.6 Max on most shared benchmarks with only 6B activated parameters.
Q: What is the context length of Qwen3.8-Flash-Next?
A: Qwen3.8-Flash-Next natively supports 262,144 tokens (256K) and can be extended up to 1,000,000 tokens via YaRN scaling. With Qwen's reworked attention kernels, prefill is up to 7.6× faster and decode up to 4.9× faster at the 1M-token mark.
Q: Is Qwen3.8-Flash-Next open source?
A: Qwen3.8-Flash-Next is released as open weights under the qwen-community-1.0 license (not Apache 2.0), available on Hugging Face at Qwen/Qwen3.8-Flash-Next and ModelScope. Commercial use is subject to the community license terms.
Q: Can I run Qwen3.8-Flash-Next locally?
A: Yes. Qwen3.8-Flash-Next is Transformers-compatible and officially supported by SGLang, vLLM, TokenSpeed, and KTransformers. Because it activates only 6B parameters, it is far more memory-friendly than its parameter count suggests.
Q: Does Qwen3.8-Flash-Next support images?
A: Yes. Qwen3.8-Flash-Next is a native vision-language model with image + text input through the OpenAI-compatible Chat Completions API. It scores 88.5 on RealWorldQA, 95.7 on MathVision (with CI), and 90.6 on CharXiv RQ (with CI).
Q: Does Qwen3.8-Flash-Next support tool calling and thinking mode?
A: Yes. Qwen3.8-Flash-Next runs in thinking mode by default, with enable_thinking, preserve_thinking, and reasoning_effort controls (xhigh/medium/low). It supports tool calling and returns reasoning content alongside the final answer.
Q: What is Qwen3.8-Flash (the hosted version)?
A: Qwen3.8-Flash is the official Qwen Cloud production version built from Qwen3.8-Flash-Next, with a 1M context window by default and official built-in tools — recommended if you want managed inference rather than self-hosting.
Q: When was Qwen3.8-Flash-Next released?
A: Qwen3.8-Flash-Next open weights were released on August 26, 2026, alongside the technical report and Qwen Cloud availability of the production Qwen3.8-Flash version.
Summary & Recommended Actions
Qwen3.8-Flash-Next is one of the most important efficiency releases of 2026. It is the first open-weight glimpse of the Qwen4 architecture, and it proves that a "mid-tier Flash" model — just 6B activated parameters — can top a frontier flagship on the benchmarks that matter most for real agentic work: DeepSWE 1.1 (58.7), SWE-bench Pro (62.5), CoWorkBench (73.9), and LiveCodeBench v6 (91.9). With native vision, a 256K→1M context window, QSA-driven long-context speedups, and roughly one-ninth the training cost of Qwen3.7-Plus, it redefines the compute-to-intelligence ratio.
The caveats: these are Qwen's own benchmarks pending independent verification; the license is qwen-community-1.0, not fully permissive; and on the hardest frontier-reasoning tests (GPQA Diamond 91.3, HLE 40.0) Claude Opus 4.6 Max still leads. But for cost-sensitive, long-context, agentic and multimodal workloads, Qwen3.8-Flash-Next sets a new benchmark.
Your next steps:
- Try the API — test Qwen3.8-Flash-Next through the OpenAI-compatible Chat Completions API (via a local SGLang/vLLM serve or Qwen Cloud) on a real agentic coding task.
- Benchmark it against your stack — compare Qwen3.8-Flash-Next side-by-side with your current model on DeepSWE-style and long-context tasks to see the cost/perf win.
- Watch the architecture — since Qwen3.8-Flash-Next previews Qwen4, follow the Qwen3.8-Flash-Next blog and technical report for the full-generation rollout.
Sources: Qwen3.8-Flash-Next on Hugging Face · Qwen3.8-Flash-Next blog · Qwen3.8-Flash-Next technical report · OfficeChai: Alibaba Releases Qwen 3.8 Flash-Next · PC Watch (Japanese)
Originally published at: Qwen3.8-Flash-Next (2026): The Complete Guide to Qwen's Qwen4-Preview Architecture Model
