Slotstream Streams 125B Qwen From SSD

Slotstream’s breakthrough is architectural: it turns SSD storage into a practical second-tier memory hierarchy for sparse-MoE inference, though latency and disk requirements remain substantial.

• –Qwen3.8-Flash-Next activates roughly 6B parameters per token, but its full 125B model plus 51B n-gram embeddings still require about 104 GB of storage. [Qwen model repository](https://github.com/QwenLM/Qwen3.8-Flash-Next/)
• –Performance is usable rather than spectacular: about 12 tok/s on a measured 48 GB M5 Pro Mac, with lower-memory configurations estimated around 4–9 tok/s.
• –The Ollama/OpenAI API compatibility makes Slotstream easy to connect to existing local AI tooling, but tools, images, JSON-schema output, logprobs, and embeddings are unsupported.
• –The project is tightly specialized around one model and allows only one model process at a time, limiting its role as a general-purpose local inference server.
• –Developers gain an unusually accessible way to experiment with frontier-scale open weights locally, provided they can spare roughly 110 GB of SSD space and tolerate slow prompt processing.