AgentionAI releases Qwen3.8-Flash-Next-Gyro-Coder for 24 GB GPUs
AgentionAI has published Qwen3.8-Flash-Next-Gyro-Coder, a quantized version of the Qwen3.8-Flash-Next model designed to run on a single consumer GPU with 24 GB of memory, such as an RTX 3090, 4090, or 7900 XTX. The model utilizes a mixture-of-experts architecture, retaining only the 256 experts out of 512 that are relevant for programming tasks. This selective pruning, combined with a custom…
Key points
- Runs on 24 GB GPUs like RTX 3090 using 19.9 GiB of memory.
- Retains 256 of 512 experts selected via evolutionary search for coding.
- Passes 13 of 16 coding tests; 93% top-1 agreement on Python.
The model achieves 93% top-1 agreement with the full model on held-out Python data and passed 13 of 16 coding tests in the author's evaluation. A larger variant, Gyro-Coder-L, is available for 32 GB cards, using 27.5 GiB of memory and passing 15 of 16 tests. The release includes instructions for running the model via a custom llama.cpp Docker build, supporting both NVIDIA and AMD/Intel GPUs. The file size is 70.6 GiB on disk due to an n-gram table stored on NVMe storage.
AgentionAI notes that the model is best suited for clear, well-specified coding instructions like implementing or refactoring code, rather than open-ended general knowledge queries. The project is currently experimental, with potential changes to file formats and kernels expected in future updates.
The story so far
3 episodes →- AgentionAI releases Qwen3.8-Flash-Next-Gyro-Coder for 24 GB GPUsthis story
agentionai/Qwen3.8-Flash-Next-Gyro-Coder-GGUF · Hugging Face
huggingface.co · 23 October 2026
Loading the full article…
This text was published by huggingface.co. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Moonshot AI releases Kimi K3.1 Pro with 1M-token context and open weights by Nov 4 · 1 src
- OpenAI makes its Decisions API widely available using GPT-6 Luna · 6 src
- OpenAI brings GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna and Codex to Google Cloud · 22 src
- TorchNeuron Space lists which Hub models run unmodified on AWS Trainium · 1 src
- Deepseek launches V4.1 Flash 1017 model with 1 M token context · 1 src
Comments
via GitHub Discussions