A developer going by the handle synchopate has demonstrated DeepSeek-V4-Flash, a 284-billion-parameter mixture-of-experts model, running on a machine with only 64GB of RAM and no graphics card. The project, hosted on GitHub under the name llama-modchip, claims the model uses fewer than 59GiB of RAM at 256K context length and can generate around 9.1 to 14.7 tokens per second on a pure CPU setup. Higher prompt-processing speeds of 44.4 tokens per second are also cited. Smaller 30B mixture-of-experts models achieve 25 to 111 tokens per second under the same constraints.
The repository has quickly drawn attention on Hacker News and Reddit, where developers are calling it a major flex for local AI. Running a model of this size without accelerators challenges the assumption that frontier-class inference requires expensive GPUs, and it arrives amid growing interest in DeepSeek and agent-based workflows.
DeepSeek-V4-Flash 284B specs and performance
The numbers come from the project's own README and issue tracker. The repository mentions additional components: SwiGLU-H2, a Stockfish-PVS L3 integration, and something called a 13ms Claude Code workflow. A 13ms latency figure for a Claude Code interaction is striking, though the repository does not provide full context or benchmark methodology in the excerpt available.
As with any open-source performance claim, independent verification is still pending. The project is public on GitHub and has attracted ten issues, suggesting active community scrutiny. Those issues may contain questions about reproducibility, which is typical for viral AI efficiency projects.
- Model: DeepSeek-V4-Flash, 284B parameters (mixture-of-experts)
- Hardware: pure 64GB CPU, zero GPUs
- RAM usage: under 59GiB at 256K context length
- Inference speed: 9.1 to 14.7 tokens per second; 44.4 tokens per second prompt processing
- Smaller models: 30B MoEs running at 25 to 111 tokens per second
