Qwen3.8 Flash Next in Ollama: 27 GiB less memory on Macs

Ollama 0.40.16 cuts Qwen3.8 Flash Next memory on Apple silicon by about 27 GiB, aimed at 128GB Macs. What changed, what is unmeasured, how to run it.

· 2 min read

Ollama 0.40.16, released on 9 November 2026, cuts the memory Qwen3.8 Flash Next needs on Apple silicon by about 27 GiB, according to Ollama's release notes. That matters because the change is aimed at making Qwen's 125B mixture-of-experts model practical on a 128GB Mac, leaving room for the long context a coding agent needs.

This brief covers what changed, where the saving comes from, what has not been measured yet, and what to do if you run Qwen3.8 Flash Next in Ollama on a Mac.

What changed in Ollama 0.40.16?

The saving comes from one pull request, "MLX: make Qwen3.8 Flash Next memory efficient", merged on 6 November 2026 and shipped in 0.40.16. Its description states the goal plainly: to make the model fit on 128 GiB Macs. It reports the model's wired memory falling from 97.6 GiB to 70.8 GiB (pull request #18078), with no measurable loss in decode speed.

The trick is in the model's unusual design. Qwen's model card describes 125B parameters with 6B active per token, plus 51B of n-gram embeddings: a very large lookup table of which each token needs only a few rows. Ollama now keeps that table, given as 26.8 GiB in its 4-bit form in the pull request, memory-mapped from the model files and copies only the rows each token selects into MLX, instead of holding the whole table resident on the GPU.

What has not been measured yet

The 97.6 to 70.8 GiB figure in the pull request does not state the hardware, context length or workload behind it. In the same discussion, a reviewer reported decode speed as good but prompt processing as poor, suggested that long prompts touch many more embedding rows, and noted that no before-and-after comparison had been run. As of 11 November 2026, we found no published measurement of prompt-processing speed with the change, so expect long first prompts from a coding agent to be the part to watch.

What it means if you run models yourself

Ollama's library lists the 4-bit MLX build, qwen3.8-flash-next:125b-a6b-nvfp4, as a 105GB download with a 256K-token context (Ollama library, checked 11 November 2026). The download is larger than the memory figure above because the embedding table stays on disk and is read as needed.

ollama --version   # needs 0.40.16 or later
ollama pull qwen3.8-flash-next:125b-a6b-nvfp4

If you are choosing hardware for this model, our guide to the best local LLM by memory puts Qwen3.8 Flash Next as the pick at 128GB, and the Mac Studio buying guide covers which configurations reach that memory. The model is published under the Qwen Community licence (model card), not Apache 2.0, so check the terms before building it into a product. For a 24GB graphics card instead, Edge0 has published a derivative of this model with its own engine: see our Edge0-145B brief.

Once it runs on the Mac, you can use it from your laptop or a coding agent on another computer: the Ollama guide shows how Corewire shares Ollama's models across your own devices.

More in Runtimes

Corewire shares the models on your computers with every device you own, and with the coding agents you already use. Download it free or read the guides.