169Pi ships Alpie Seed 9B: open-weight model for offline laptops
169Pi released Alpie Seed, a 9-billion-parameter Apache-2.0 vision-language model meant to run locally. Weights are on Hugging Face with quantized GGUF builds, plus Ollama and Kaggle, and a native 262K-token context window.
Why it matters
The Hugging Face model card says Alpie Seed keeps the Qwen3.5-9B architecture and adds post-training with supervised fine-tuning plus GRPO for reasoning, instruction following, coding, and visual understanding. It accepts text, image, and video. FP16 weights plus 2-bit, 3-bit, and 4-bit GGUF builds target laptops, Apple Silicon, phones, and even CPU-only machines.
Vendor-reported FP16 scores with thinking enabled include 61.20% on SWE-Bench Verified and 72.99% on LiveCodeBench v6. 169Pi notes quantized variants differ and comparison scores come from other model cards, so cross-model rankings are indicative. Example serving uses vLLM with an OpenAI-compatible localhost endpoint.
Alongside the weights, 169Pi opened invite-only Alpie Local, a coding agent on Seed that plans work, edits files, runs tests, and self-corrects on the developer’s machine. For teams that need private offline inference, Seed is a self-host option—re-run benchmarks in your own harness before trusting the headline numbers.