What Is Qwen 3.8 Flash Next? 125B MoE
Elena Rossi
AI Adoption Analyst

TLDRQwen 3.8 Flash Next launched on August 27, 2026, with 125B main-model parameters, 6B active parameters, 262K native context, open weights, and an architecture preview for Qwen 4.
Qwen 3.8 Flash Next Is a 125B-A6B MoE — 262K Native Context
Qwen 3.8 Flash Next is Alibaba’s open-weight multimodal mixture-of-experts model, officially launched on August 27, 2026, and it previews the architecture used in the Qwen 4 family. Its main model contains 125B parameters, activates 6B parameters per token, and adds a 51B N-gram Embedding component. The model supports 262,144 native context tokens and can extend to 1,000,000 tokens with YaRN. Alibaba published the weights as an architecture preview rather than as the final production Qwen4 model. The official technical overview describes the release and its design goals.
Updated 2026-10-06: New community tests report faster low-memory and long-context local inference, alongside an isolated tool-use failure (see the Update below).
Key Takeaways
- Qwen 3.8 Flash Next is a launched multimodal causal language model with a vision encoder and sparse MoE routing.
- The main model has 125B parameters, with 6B activated for each token, plus 51B N-gram Embeddings.
- Its native context window is 262,144 tokens, with a stated 1,000,000-token extension through YaRN.
- Qwen Sparse Attention, or QSA, reduces long-context attention work by selecting relevant context at micro-block granularity.
- The weights are available on Hugging Face and ModelScope under the qwen-community-1.0 license.
- Qwen 3.8 Flash Next has no separately confirmed hosted API price; the production Qwen3.8-Flash service is listed at $0.16 per million input tokens and $0.47 per million output tokens.
What Is Qwen 3.8 Flash Next?
Qwen 3.8 Flash Next is an experimental open-weight model from Alibaba’s Qwen team, officially launched on August 27, 2026. It is designed to test architectural changes before a complete Qwen 4 family arrives. Alibaba describes it as a multimodal MoE model, meaning it combines text generation, visual input, sparse expert routing, and specialized long-context components.
The “Next” label matters. This is not simply a larger version of Qwen 3.8 Flash. It is an architecture preview that exposes new model components to developers and inference researchers. The production Qwen3.8-Flash version is a related service with 1M-token context enabled by default and official built-in tools.
The model has moved beyond a leak-only stage. Alibaba’s official August 27, 2026 technical post says the weights are open, and the Qwen model repository is available on Hugging Face. Availability of weights does not mean every inference engine is mature. Community reports still describe active work on kernels, quantization, speculative decoding, and long-context behavior.
Qwen 3.8 Flash Next is best understood as a public testbed for cheaper inference at high model capacity.
Qwen 3.8 Flash Next at a Glance
| Specification | Qwen 3.8 Flash Next |
|---|---|
| Developer | Alibaba Qwen team |
| Launch date | August 27, 2026 |
| Type | Multimodal causal language model with vision encoder; sparse MoE |
| Main model size | 125B parameters |
| Active parameters | 6B parameters per token |
| Additional component | 51B N-gram Embedding parameters |
| Modality | Text, image, and video input |
| Context window | 262,144 tokens native; extensible to 1,000,000 with YaRN |
| Pricing | Not separately confirmed for a hosted Qwen 3.8 Flash Next API |
| Related production pricing | Qwen3.8-Flash: $0.16 per million input tokens and $0.47 per million output tokens |
| Availability | Weights on Hugging Face and ModelScope; local engine support varies |
| License | qwen-community-1.0 |
The parameter figures require careful interpretation. The 125B figure describes the main model, while the 51B figure describes a separate N-gram Embedding component. Some community deployment recipes describe the combined deployment as approximately 176B parameters because they count both components.
How Qwen 3.8 Flash Next Works
The architecture combines several mechanisms that target different bottlenecks.
Gated DeltaNet (GDN) compresses historical information into a compact state. Alibaba says three out of every four layers use GDN, while the remaining layer provides global attention. This arrangement reduces the cost of carrying every previous token through full attention at long sequence lengths.
Qwen Sparse Attention (QSA) changes how the global-attention layers retrieve information. Instead of evaluating every token equally, QSA uses a lightweight indexer to select important context at micro-block granularity. Its purpose is to reduce long-sequence computation while preserving access to information that matters.
QSA is one of the main Sparse Context Retrieval anchors associated with this model. The approach is promising for long documents and agent histories, but independent tests have not yet established how reliably it preserves long-context understanding. Ivan Fioravanti described QSA as impressive while calling for evaluations of its effect on large-context comprehension. His technical discussion links to Alibaba’s architecture explanation.
Gated Residual (GR) widens the residual stream into four branches. Dynamic read and write gates control how information moves through those branches. This is intended to improve cross-layer information flow and training stability without adding a large inference burden.
N-gram Embedding supplies another capacity channel. The model looks up local bigrams and trigrams in a table containing 20,000,000 entries. Alibaba says the embedding table can be moved to host memory and prefetched asynchronously while model computation continues. This creates the model’s Host-Offloaded Memory Scaling approach: more stored capacity without making every parameter part of every transformer calculation.
The architecture also uses a tailored Muon and AdamW optimization recipe. The training system assigns those optimizers to different weight categories and refits scaling laws for the new design. Alibaba reports that training required approximately one-ninth the cost of Qwen3.7-Plus while producing stronger coding and office-task results.
What You Can Do With Qwen 3.8 Flash Next
Qwen 3.8 Flash Next is suited to workloads that need a mixture of coding, tool use, visual input, and long conversation history.
- Agentic coding: The model can serve as a coding agent or sub-agent, particularly where tool calls and repository context matter.
- Long-document analysis: A native 262,144-token window can hold large codebases, research collections, or extended task histories.
- Computer-use workflows: Alibaba positions the architecture for complex API interaction, calculators, external databases, and visual prompts.
- Multimodal assistance: The model accepts image and video input, enabling visual analysis alongside text reasoning.
- Local AI experimentation: Developers can test quantization, attention kernels, speculative decoding, and memory offloading on their own hardware.
- Cost-sensitive inference research: The sparse active-parameter count makes it useful for studying the relationship between model capacity and per-token compute.
Community testing shows the model running on hardware ranging from a 24GB RTX 4090 to Apple Silicon systems and DGX Spark machines. These results are not interchangeable benchmarks. A reported RTX 4090 test reached approximately 21 tokens per second in decode and 364 tokens per second in prefill at a 250,000-token context, but that result came from an individual setup rather than a standardized evaluation.
On Apple Silicon, community tests have reached approximately 65 tokens per second on an M3 Ultra using a Q2 build with DwarfStar. Another M4 Max test reached 224,000 tokens of context on a 64GB machine. These figures demonstrate implementation potential, not guaranteed performance for every Apple Silicon configuration.
How Qwen 3.8 Flash Next Compares
The closest comparison depends on whether the reader wants open weights, a hosted endpoint, or a smaller local model.
| Model | Positioning | Context and access | What is confirmed |
|---|---|---|---|
| Qwen 3.8 Flash Next | Open-weight architecture preview | 262K native; 1M with YaRN; local deployment | 125B main model, 6B active, multimodal |
| Qwen3.8-Flash | Related production service | 1M context by default; QwenCloud; built-in tools | $0.16 input and $0.47 output per million tokens |
| Qwen 3.8 27B | Smaller dense multimodal model | Local deployment | 27B parameters |
| DeepSeek V4 Flash Vision Exp | Community comparison target | Availability and pricing not confirmed here | A community post claimed Flash Next beat it in prefill speed and time to first token on DGX Spark |
The key trade-off is capacity versus deployment complexity. Qwen 3.8 Flash Next activates only 6B parameters per token, but its full weights and N-gram table still require substantial memory or offloading. Readers comparing local models can also review the earlier analysis of Qwen 3.8 27B.
A community Arena observation gave Qwen 3.8 Flash Next a score of 56, but the evaluation method and full comparison set were not supplied in the evidence. That number should be treated as an early signal, not a definitive ranking.

Source: @ivanfioravanti
Availability: How to Access Qwen 3.8 Flash Next
The primary access path is self-hosting the launched weights from Hugging Face or ModelScope. The Hugging Face page lists Transformers, vLLM, SGLang, and Docker-oriented usage paths. Exact engine versions and hardware requirements can change quickly because support remains under active development.
Community implementations also exist for MLX-Serve, llama.cpp, oMLX, MTPLX, and related projects. A benchmark repository documents tests across Apple Silicon, local GPU systems, and multiple serving engines. Users should verify that the selected engine supports QSA, the desired quantization format, vision input, and the intended context length.
For multi-GPU deployment, a community-maintained recipe serves an approximately 176B combined NVFP4 deployment across two DGX Spark systems. It reports 64 tokens per second on a single stream and 117 aggregate tokens per second at two-way concurrency, with vision input working. Those numbers are recipe-specific and should not be generalized to every DGX Spark configuration. The deployment code and measurements are available in the dual-DGX Spark repository.
Hosted access is a separate question. The available materials identify Qwen3.8-Flash, not Qwen 3.8 Flash Next, as the production QwenCloud service. Qwen 3.8 Flash Next’s own hosted API terms, endpoint naming, quotas, and service-level guarantees are not separately confirmed.
For a hosted model with a similar general chat and multimodal workflow, Gemini 3.8 Flash is a separate option to evaluate; it is not Qwen 3.8 Flash Next.
What We Don't Know Yet
Several practical questions remain open:
- The final release schedule and complete specification of Qwen 4 are not confirmed by the available materials.
- Independent evaluation has not established whether QSA maintains reliable retrieval and comprehension at 250,000- to 1,000,000-token contexts.
- Quantization effects remain significant. Community tests range from 1-bit to 4-bit and mixed-precision builds, making quality comparisons difficult.
- Official hardware guidance for a standard local deployment is not published in the supplied sources.
- Qwen 3.8 Flash Next’s own hosted API price, rate limits, data-retention terms, and regional availability are not confirmed.
- Some inference engines still require model-specific kernel work, particularly on DGX Spark and other newer accelerators.
- Reports of Arena scores and frontier-model parity come from individual or community posts rather than a complete independent evaluation suite.
The practical verdict is conditional: Qwen 3.8 Flash Next is a compelling launched architecture preview for developers willing to manage experimental inference, but its long-context quality and production readiness still need broader validation.
Frequently Asked Questions
What is Qwen 3.8 Flash Next?
Qwen 3.8 Flash Next is Alibaba’s open-weight multimodal mixture-of-experts model, launched on August 27, 2026, and an early preview of the architecture used in Qwen 4. It has 125B main-model parameters, activates 6B parameters per token, and supports 262,144 native context tokens.
Is Qwen 3.8 Flash Next open source?
Qwen 3.8 Flash Next is open weight, with weights published on Hugging Face and ModelScope. Its listed license is qwen-community-1.0, so open-weight availability should not be treated as an MIT or Apache open-source license.
What is the context window of Qwen 3.8 Flash Next?
Qwen 3.8 Flash Next has a native context window of 262,144 tokens. Alibaba says the window can be extended to 1,000,000 tokens with YaRN, but long-context understanding still needs broader independent evaluation.
How much does Qwen 3.8 Flash Next cost?
Qwen 3.8 Flash Next has no separately confirmed hosted API price. The related production Qwen3.8-Flash service is listed at $0.16 per million input tokens and $0.47 per million output tokens.
Can Qwen 3.8 Flash Next run locally?
Qwen 3.8 Flash Next can run locally through community-supported formats and inference engines, including vLLM, SGLang, MLX-Serve, llama.cpp, oMLX, and MTPLX. Hardware requirements and speed depend heavily on quantization, memory, engine support, and context length.
Qwen 3.8 Flash Next vs Qwen 3.8 Flash: what is the difference?
Qwen 3.8 Flash Next is the open-weight architecture preview, while Qwen3.8-Flash is the related production version with 1M-token context by default and official built-in tools. The production version is served through QwenCloud at $0.16 per million input tokens and $0.47 per million output tokens.
What to watch next is the publication of standardized QSA long-context evaluations, stable support across vLLM, SGLang, and local engines, and clearer official separation between Flash Next and the production Qwen3.8-Flash service.
Update — 2026-09-08
Community deployment work has progressed from dual-Spark recipes to a single-DGX-Spark implementation, with an accompanying NVIDIA Developer Forum discussion. A September 7 report also claims that an uncensored configuration runs on one Spark using a 99GB NVFP4 checkpoint, retains vision, video, and 262K context, and makes refusals optional through a flag; these capabilities remain community-reported rather than standardized results (post by @0x0SojalSec).
Apple Silicon support is moving toward 64GB systems through the DwarfStar project. Its maintainer shared Q2 weights and a development branch for testing on real 64GB machines, reporting approximately 65 tokens per second on an M3 Ultra in a video (testing request and performance report; Q2 model page; DwarfStar branch). The figure is an early implementation result, not a guaranteed speed for every 64GB Apple Silicon configuration.
Update — 2026-10-06
Quantized deployments have reached smaller consumer systems. A Q2_0 Strata test reports 94 tokens per second on an RTX 5070 12GB with 64GB of system RAM, while an IQ_3S test reports 150 tokens per second on an RTX 5090 during a Claude Code workflow. Another user says a 3-bit GSQ-RCQ GGUF fits fully on an RTX 3090 plus 64GB system (deployment report). These results use different quantizations, runtimes, and prompts, so they are not directly comparable—and sub-4-bit quality remains a concern.
Runtime changes are also producing measurable gains. A cafe-llama.cpp MTP implementation reportedly increased decode speed from 25 to more than 50 tokens per second. On one RTX PRO 6000, a llama.cpp build comparison reduced processing time for a 256K-token prompt from 184.9 to 107.9 seconds, while showing little improvement at 64K.
Quality and tool-use behavior still require scrutiny. One tester documented a reproducible failure in which the model fabricated file-proxy URLs, received 403 responses, and left an estimated 200–450 tokens of context debris per attempt. This is an isolated report rather than evidence of a model-wide defect.
Try related models on Kie.ai: Gemini 3.8 Flash, Claude Opus 5.5 and DeepSeek-V4.1-Flash.
About Elena Rossi
Elena watches developer chatter and early adoption signals to gauge which releases gain real traction.
View all posts by Elena Rossi