Qwen 3.8 Flash Next Benchmarks: The Real Numbers

Wondering what the qwen 3.8 flash next benchmarks actually show? The quick version is just below — I’ve been digging through the official model card since the weights dropped, and the numbers are genuinely interesting for anyone building AI agents or AI SEO systems on open models.

Qwen 3.8 flash next benchmarks: the headline numbers

First, the sourcing: every number here comes from the official Qwen/Qwen3.8-Flash-Next model card on Hugging Face, which I pulled directly on 28 August 2026. The repository went live on Hugging Face around 24 August, with the formal release landing on ModelScope on 26 August 2026. Alibaba’s Qwen team positions the model as an experimental preview of the architecture that will underpin Qwen4, and the production variant — Qwen3.8-Flash — is already serving on Qwen Cloud with 1M context by default.

Here are the headline scores from the model card:

Benchmark	What it measures	Qwen3.8-Flash-Next score
SWE-bench Pro	Real-world software engineering fixes	62.5%
Multilingual SWE	Software engineering across languages	81.0%
CoWorkBench	Agentic co-working tasks	73.9%
Toolathlon	Tool-use across long workflows	73.5%
AndroidWorld	Multimodal on-device agent control	84.5%
RealWorldQA	Real-world visual understanding	88.5%

Even with that caveat, 62.5% on SWE-bench Pro from a model with 6B active parameters is the story here. That is the kind of score we were seeing from far heavier closed models earlier this year, and it is exactly the class of model I plug into my AI SEO automations — the same systems I teach inside the AI Profit Boardroom.

What the qwen 3.8 flash next benchmarks actually mean

The scores only make sense once you understand the architecture, because this is not a normal dense model. The card describes 125B total parameters with just 6B activated per token, plus a 51B n-gram embedding table (20 million bigrams and trigrams injected at layer 2) and a 4B multi-token prediction module — about 180B on disk all-in.

The layer stack is genuinely new: 48 layers arranged as repeating blocks of Gated DeltaNet followed by Mixture-of-Experts, punctuated by Qwen Sparse Attention (QSA) layers. The MoE router picks 10 routed experts plus 1 shared expert out of 512 per token. In plain English: you get big-model knowledge with small-model inference cost. That is why the agentic numbers (CoWorkBench 73.9%, Toolathlon 73.5%) matter more than the raw coding score — cheap tokens are what make long agent runs affordable.

Context is 262,144 tokens natively, extensible to 1,000,000. It is also properly multimodal — text, images and video in — which explains the AndroidWorld 84.5% score: that benchmark has the model driving a phone UI from screenshots.

How to try it yourself (and what I’d test first)

Deployment support at launch covers Transformers, vLLM, SGLang and TokenSpeed, with quantised community builds for llama.cpp, Ollama and LM Studio — Unsloth already has a GGUF repo up. If you have run Qwen 3.8 locally before (my setup guide is at how to use Qwen 3.8), the workflow is identical: pull the quant that fits your VRAM, wire it into your agent harness, and test on your own tasks rather than trusting the card.

My suggested first tests, in order:

• Your own repo, one real bug. SWE-bench Pro at 62.5% is impressive on paper; your codebase is the benchmark that pays your invoices.
• A 30-minute agent run. The 6B-active design is built for long tool-use loops — pair it with an agent framework the way I pair models in my Hermes Agent + Qwen 3.8 setup.
• A screenshot task. RealWorldQA 88.5% suggests the vision side is not a bolt-on — test it against the vision models I covered in GLM-5.3-Flash.

For SEO specifically, the interesting play is cost-per-page: a model this cheap to run changes the economics of programmatic content QA, entity extraction and internal-link mapping. That workflow — picking the model, building the pipeline, shipping pages that rank — is what the daily tutorials inside the AI Profit Boardroom cover step by step, and if you want it mapped onto your site first, book a free SEO strategy session.

The bottom line on qwen 3.8 flash next benchmarks

The qwen 3.8 flash next benchmarks tell a consistent story even after you discount for vendor optimism: Alibaba has shipped an open-weight preview of the Qwen4 architecture that trades disk space for inference speed, and the agentic and multimodal scores are the ones to watch. If independent evals confirm even 90% of these numbers, this becomes the default open model for long-running agents on a budget.

My verdict: download it now, benchmark it on your own tasks this week, and do not rebuild your stack around it until the community numbers land. Released circa 24–26 August 2026, per the official Hugging Face model card and ModelScope release.

FAQ: qwen 3.8 flash next benchmarks

Is Qwen 3.8 Flash Next open weight?

Yes. The official Qwen/Qwen3.8-Flash-Next repository is live on Hugging Face under the qwen-community-1.0 licence, with an FP8 variant and community GGUF quantisations (including Unsloth’s) also available.

How big is Qwen 3.8 Flash Next?

The model card lists 125B total parameters with only 6B activated per token, plus a 51B n-gram embedding table and a 4B multi-token prediction module — roughly 180B on disk in full precision.

What context window does Qwen 3.8 Flash Next have?

262,144 tokens natively, extensible up to 1,000,000 tokens. The hosted production variant, Qwen3.8-Flash on Qwen Cloud, ships with 1M context by default.

Can I run Qwen 3.8 Flash Next locally?

Yes — the card lists Transformers, vLLM, SGLang and TokenSpeed support, with quantised builds for llama.cpp, Ollama and LM Studio. With 6B active parameters it is far lighter at inference than the disk size suggests, but you still need serious RAM/VRAM to hold the weights.

Is Qwen 3.8 Flash Next multimodal?

Yes. It is an image-text-to-text model that handles text, images and video input, which is why it posts scores on multimodal agent benchmarks like AndroidWorld and RealWorldQA.

Is Qwen 3.8 Flash Next the same as Qwen 4?

No — it is an experimental preview of the architecture that will underpin Qwen4. Think of it as a public test drive of the Qwen4 design: Gated DeltaNet, Qwen Sparse Attention and the MoE layout are the story, not the version number.

Related reading

• How to Use Qwen 3.8: Setup & Best Uses
• Hermes Agent + Qwen 3.8: Setup & Why It’s So Good
• GLM-5.3-Flash: Z.AI’s New Vision Model Explained

Last updated August 2026. This is the living guide to qwen 3.8 flash next benchmarks — it gets updated as the tools change.