Qwen's New Model Qwen3.8-Flash-Next Runs Locally on 75GB of Memory

The first preview of the Qwen4 architecture

Qwen3.8-Flash-Next is the first preview model of Alibaba's next-generation Qwen4 architecture, which has been in the works for a while. It's a multimodal model that reads text and images together, and it supports a 262K-token context window, so long documents or entire code repositories can be fed in and processed at once. Back on August 14, Alibaba Qwen released a 27B dense model, Qwen3.8-27B, under Apache 2.0, and that one ran on a single 24GB-class graphics card. Qwen3.8-Flash-Next is a different animal — at 125B scale, it's nearly five times larger.

How it runs on 75GB of memory

Unsloth released Qwen3.8-Flash-Next in several quantization tiers, from 1-bit up to 8-bit. The original BF16 weights come in at 355GB, but the lightest 1-bit version shrinks that down to 75GB.

Quantization level	Memory required
1-bit	75GB
2-bit	79GB
3-bit	90GB
4-bit	112GB
6-bit	128GB
8-bit	156GB
BF16 (original)	355GB

The figures in the table refer to combined RAM and VRAM, or unified memory. Unsloth says that even at 1-bit — 79% smaller than the original — the model holds onto 79% top-1 accuracy. The company explains that the model's architecture is unusual enough that inference on CPU RAM or unified memory can approach GPU VRAM speeds, which makes it well suited to high-memory machines like Macs or NVIDIA's DGX Spark. In its own post, Unsloth thanked Qwen, saying the release "made day-zero access possible."

Benchmarks: how it compares to Claude Opus 4.6 Max

Benchmark	Qwen3.8-Flash-Next	Claude Opus 4.6 Max
SWE-bench Pro (agentic coding)	62.5	53.4
SWE-bench Multilingual	81.0	77.5
NL2Repo-Bench	48.1	47.6
CoWorkBench (long-horizon work)	73.9	68.2
JobBench (professional tasks)	55.7	36.6

In Unsloth's comparison table, Qwen3.8-Flash-Next beat Claude Opus 4.6 Max on every benchmark where both models had scores. The gap was widest on JobBench, which measures professional task performance: Qwen3.8-Flash-Next scored 55.7 versus Claude's 36.6, a nearly 20-point lead. That said, these numbers come from Alibaba Qwen's and Unsloth's own releases, and no independent third-party reproduction has been confirmed yet.

How to try it

To run Qwen3.8-Flash-Next locally, start with Unsloth's guide and its Hugging Face repository. The Unsloth guide walks through installation steps and recommended settings for different hardware setups.

1. First, check how much RAM or unified memory your machine has. Based on the table above, 75GB or more gets you the 1-bit quantized version; if you have more headroom, you can go with a 4-bit version or higher.
2. Download the GGUF file at your preferred quantization level from the huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF repository.
3. Install Unsloth's dedicated llama.cpp PR build or the Unsloth Desktop app.
4. Load the downloaded GGUF file, and the model runs right away — no server GPU needed.

Any machine with 75GB or more of RAM or unified memory can run it, even without GPU VRAM. Unsloth points to Macs and NVIDIA's DGX Spark as examples of high-memory systems that fit the bill. In practice, that means people can work with a large model on a personal workstation without renting a cloud GPU.

The 262K-token context window means you can feed in an entire code repository or a long document and hand off refactoring work in one pass. That makes it applicable to agentic coding tasks like the ones SWE-bench Pro measures, as well as office work such as report writing or scheduling, which is what JobBench evaluates.

Editor's take

What stands out most about this release isn't just the numbers — it's the strategy behind it. Alibaba chose to release its next-generation Qwen4 architecture not as a finished product but as a preview, testing market reaction with a single 125B-scale MoE model. Considering it's been barely ten days since the August 14 release of the 27B dense model, and this new model is nearly five times larger, it reads like Alibaba doubling down on a strategy of densely filling out its open-weight lineup to go head-to-head with closed-model players like Anthropic and Google.

Qwen3.8-27B was a practical, workhorse model that ran on a single 24GB graphics card. Qwen3.8-Flash-Next feels different — it's far larger, yet it runs on memory alone with no GPU at all. Having worked with open models at this scale before, the shift is notable: tasks that used to require stitching together multiple GPUs are now moving onto a single workstation with enough RAM.

For companies in Korea, rather than rushing to add more server GPUs, it may be more cost-effective to first attach a model like this to a single memory-rich workstation or a DGX Spark-class machine and run a pilot on coding or document tasks. That said, keep in mind these benchmarks are self-reported by Alibaba and Unsloth — it's worth validating against your own internal data rather than skipping that step.

There's a good chance Unsloth or other parts of the community will publish independent benchmarks or fine-tuning guides for this model within the next few weeks. And how the market responds to this preview will likely shape the timeline for a full Qwen4 release.