Alpie Seed
by 169Pi · Model · Language & code models
weak
Benchmark claims don't hold up in hands-on tests; the base Qwen 9B seems better.
9B vision-language model post-trained with SFT and GRPO; text and images, thinking mode, 262K context, Apache 2.0, GGUF released
AccessOpen weightsNeeds~9 GBDownloadQ4 GGUF needs about 9 GB of VRAMSize9B parametersLicenseApache-2.0
Why we say so
169Pi's own table shows it beating every rival, even 20B and 30B models. Fahd Mirza's hands-on tests disagree: it missed a planted bug and invented three false ones, failed at tool calling at first, botched a sketch-to-code task, and got a physics constant wrong by about 100x. He rates it worse than the original Qwen 9B. This is one reviewer, but the failures were consistent across tasks.
Evidence
- Nov 14Vendor benchmark table shows it topping every row against popular models including larger 20B and 30B ones; reviewer doubts the claims
- Nov 14Failed to find the single planted bug in a Docker app via Hermes agent, reported three false issues, and its first patch could break the dashboard
- Nov 14Reviewer judges it worse than the original Qwen 9B in practice, despite benchmark tables placing it on top of every row