Qwen3.8-Flash-Next
Alibaba · Vision-language · MoE 125B (A6B) · 262k context · Released 26 August 2026
Qwen3.8-Flash-Next is a sparse mixture-of-experts with 512 experts, eleven active per token (ten routed plus one shared), using a hybrid of Gated DeltaNet and Qwen Sparse Attention that Alibaba positions as a look at the Qwen4 architecture. It takes text, images and video, produces text, and carries a 262k context extensible towards a million tokens. The memory catch matters: although only about 6 billion parameters compute per token, the full checkpoint is roughly 180 billion parameters in BF16 (it includes a large n-gram embedding and multi-token-prediction components), and it is the total, not the active count, that sets the memory footprint. That makes it a 128GB unified-memory machine (Apple Silicon, AMD Strix Halo, NVIDIA DGX/RTX Spark) or a multi-GPU server, not a single consumer card.
Strengths
- The community's favourite for agentic, coding and design work on 128GB unified memory
- Very sparse: only about 6 billion of ~125 billion parameters compute per token, so it is fast
- Multimodal input across text, images and video, with a 262k context
- Genuinely open weights, usable commercially below a large user and revenue threshold
Weaknesses
- Server-scale footprint: the full checkpoint is ~180B in BF16, so memory is driven by the total, not the 6B active
- Needs a 128GB unified-memory machine or multi-GPU server, not a consumer card
- A bespoke Qwen Community Licence with conditions, not Apache like the smaller Qwen models
- A novel architecture, so local runtime support may be uneven
Hardware requirements
| Quantisation | Approx. VRAM | Notes |
|---|---|---|
| Q4 (approx) | ~95GB | Roughly 90 to 100GB at 4-bit, which fits a 128GB unified-memory machine. The figure is driven by the total parameters, not the 6B active, and is approximate. |
| BF16 (as released) | ~350GB | Roughly 350GB at full precision (about 180B parameters including auxiliary components). Multi-GPU server territory. Figure is approximate. |
Optimised builds available for Apple Silicon.
What you'd need to run this
Roughly what a machine to run this would need, at up to three levels of quality. Memory is the deciding factor.
Minimum to run it
Q4 (approx) · ~95GB needed
128GB of unified memory
Mac Studio M4 Maxor an 80GB-class data-centre card, which is usually rented by the hour, NVIDIA H200 .
Unified memory is shared with the model, so it is already counted above.
Best quality
BF16 (as released) · ~350GB needed
A multi-GPU server, roughly 6× 80GB-class GPUs
Server memory in the hundreds of gigabytes or more, plus fast interconnect.
tens of thousands of pounds, or rented by the hour
Usually rented — see the cost calculatorLicence
Qwen Community License 1.0 — read the licence
Qwen3.8-Flash-Next: common questions
- What hardware do I need to run Qwen3.8-Flash-Next?
- At its most compressed (Q4 (approx)) it needs roughly 95GB of VRAM. VRAM figures are approximate and depend on context length and settings.
- Is Qwen3.8-Flash-Next free for commercial use?
- Commercial use is permitted, but with conditions. Released under the Qwen Community License 1.0, not Apache. Read the licence before relying on it at scale.
- Can I run Qwen3.8-Flash-Next on Apple Silicon?
- Yes. Qwen3.8-Flash-Next has builds optimised for Apple Silicon, through MLX or GGUF on a Mac.
- What is Qwen3.8-Flash-Next's context window?
- Qwen3.8-Flash-Next has a context window of 262,144 tokens, about 262k.
Availability
Recommended for
- Agentic, coding and design work on a 128GB unified-memory machine
- High-volume inference where sparsity keeps a large model fast
- Long-context multimodal work kept in-house
Related models
Catalogue entry last verified 2 October 2026. Specifications change; verify anything you are about to spend money on.