We have been experimenting with whether strong multi-step reasoning actually requires a full-precision scale.
Alpie Core is a 32B parameter model trained and served entirely at 4-bit precision. Instead of training in FP16 and compressing later, we optimized the model end-to-end for low-precision reasoning.
Despite the constraint, it performs competitively on reasoning and coding benchmarks like GSM8K, BIG-Bench Hard, and SWE-Bench Verified, while using far less memory and infrastructure.
We are especially interested in feedback on long-context behaviour, reasoning stability across retries, and how it performs inside agent or tool-using workflows.
ChiragArya•1d ago
I’m one of the builders behind Alpie Core. Happy to answer questions, clarify benchmarks, or share more technical details about the model, training setup, or deployment choices.
ChiragArya•1d ago
Alpie Core is a 32B parameter model trained and served entirely at 4-bit precision. Instead of training in FP16 and compressing later, we optimized the model end-to-end for low-precision reasoning.
Despite the constraint, it performs competitively on reasoning and coding benchmarks like GSM8K, BIG-Bench Hard, and SWE-Bench Verified, while using far less memory and infrastructure.
We are especially interested in feedback on long-context behaviour, reasoning stability across retries, and how it performs inside agent or tool-using workflows.