Qwen3.8-Flash-Next: 6B active, 173 GiB on disk

Somebody on the team read "6B activated" and asked which of our boxes could run it. Fair question, wrong number: the FP8 checkpoint for Qwen3.8-Flash-Next is about 173 GiB of safetensors, and the BF16 one is 335 GiB. Six billion is what fires per token, not what you have to keep somewhere. Alibaba put the model out on 26 August 2026 as an open-weight preview of the architecture behind Qwen4, and honestly the engineering in it is the most interesting thing we have read all month. Two details keep getting flattened in the coverage, though. The licence is not Apache 2.0, whatever a few launch write-ups said. And the model you call on the API is named Qwen3.8-Flash, which is not the same artefact as the one on Hugging Face.

Two names, one release

This trips people up within about thirty seconds of the announcement, so let’s clear it first.

The weights on Hugging Face are Qwen/Qwen3.8-Flash-Next, with an FP8 variant beside them. The thing you hit on Qwen Cloud is Qwen3.8-Flash. Qwen’s own model card puts it plainly:

Qwen3.8-Flash is the official version based on Qwen3.8-Flash-Next with more production features, e.g., 1M context length by default, official built-in tools.

So the API model is a superset, not a mirror. If you benchmark the open weights locally and then ship against the hosted endpoint, you are not testing the same system. That matters more than usual here, because the context handling is one of the things that differs: the open release is 262,144 tokens natively and stretches to a million with YaRN, while the hosted one starts at a million.

Pricing on Qwen Cloud is $0.16 per million input tokens and $0.47 per million output. OpenRouter’s listing reads $0.15 in, same $0.47 out. Nobody has explained the cent.

The architecture is the actual story

Image: Qwen (Alibaba Group), from the Qwen3.8-Flash-Next model card

Forty-eight layers, arranged as twelve repetitions of three Gated DeltaNet blocks followed by one Qwen Sparse Attention block. Every one of those blocks feeds a mixture of experts with 512 experts, of which ten routed plus one shared fire per token. Expert intermediate dimension is a slim 640. Hidden dimension across the model is 2560, which is small for something with 125B parameters in it, and that’s the point: the width lives in the expert count, not the tensor shapes.

Then there’s the n-gram table. Twenty million bigram and trigram entries, 51B parameters, consulted at layer 2 only. Qwen’s framing is that embeddings scale parameters without scaling compute and are easier to offload than MoE weights are. Which reads as a polite way of saying they found somewhere cheaper than HBM to keep half the model.

I might be wrong about how well that holds up under a real serving load. The idea is sound, the memory hierarchy is not always as forgiving as an architecture diagram makes it look.

The 6B figure, read properly

Six billion parameters activate per token. That’s a compute claim, and it’s a good one. It is not a memory claim, and the gap between the two is where the disappointment usually lands.

We pulled the file sizes off the Hugging Face API on 30 August: 144 files summing to 172.8 GiB for the FP8 repo, 335.3 GiB for BF16. Add the KV cache for whatever context you actually intend to use, add framework overhead, and this is a small cluster or a very well equipped single node. Not a workstation.

Compare that to GLM-5.3-Flash, which Z.ai shipped the same day at 320B total and 18B active. Different tradeoff, same conclusion for anyone hoping to self-host on a desk.

What the benchmarks say, losses included

Credit where it’s due: Qwen published a table that includes the places it comes second.

It beats Qwen3.7-Plus, a 397B model with 17B active, almost everywhere. SWE-bench Pro 62.5 against 55.8. CoWorkBench 73.9 against 65.1. JobBench 55.7 against 27.6, which is a gap wide enough to make us suspicious of the harness rather than impressed by the model. Same instinct on DeepSWE 1.1, where Flash-Next posts 58.7 and Qwen3.7-Plus posts 16.5. Numbers that far apart usually mean the older model was failing to complete the task format, not that it’s fifteen times worse at software engineering.

The losses are more informative. On repo-level generation, NL2Repo-Bench, DeepSeek-V4-Flash-0731 takes it 54.2 to 48.1. On HLE, Claude Opus 4.6 Max wins 40.0 to 35.9. And on Agents’ Last Exam the Pass@1 goes to DeepSeek, 25.2 to 24.3.

One thing the table doesn’t do is compare against anything newer than Opus 4.6. Opus 4.8 and Opus 5 both shipped before 26 August. Their absence is a choice.

The licence is the part to actually read

The repo metadata says license: other, license_name: qwen-community-1.0. Not Apache 2.0. Several outlets reported Apache 2.0 on launch day and it isn’t a small distinction.

You can use it commercially, modify it, redistribute it, run it inside your own company. What you cannot do without asking Qwen first is run a Model as a Service or an AI Work Assistant business on it. The licence defines MaaS as giving third parties access to inference or fine-tuning via API, and explicitly carves out relaying requests to a third-party hosted copy. So reselling an endpoint you host yourself needs a separate licence. Pointing your users at somebody else’s hosted Qwen doesn’t.

Above 100,000,000 monthly active users or US$20,000,000 monthly revenue, you also have to display the model name prominently in the product.

So when would we reach for it

If you’re paying for a hosted frontier model to do long agentic runs, $0.16 per million in is worth an afternoon of evaluation, particularly on the office-work and tool-use side where CoWorkBench and Toolathlon flatter it. If you want to self-host, budget for a node and not a card, and take the FP8 checkpoint.

And if you’re building a product that sells inference to other people, read the licence before you read the benchmarks. That order saves time.

Sources

Model card and licence: Qwen/Qwen3.8-Flash-Next on Hugging Face and the Qwen Community License 1.0. Official announcement: Qwen3.8-Flash-Next, A New Architecture, Towards Ultimate Cost-Efficiency and the QwenLM repository. API listing: qwen/qwen3.8-flash on OpenRouter. Coverage: TechNode and MarkTechPost. Checkpoint sizes are our own sum of the repository files via the Hugging Face API on 30 August 2026.

Frequently asked questions