Allosaurus v0.2: Qwen3.8-Flash-Next on a 48 GB Mac
Qwen's 125B Qwen3.8-Flash-Next, compressed to fit a 48 GB Mac. About 27 GiB in memory, 93% correct on everyday tasks, and tool-calling turns up to 36% faster with speculative decoding.
Today we're releasing Allosaurus v0.2: Qwen's Qwen3.8-Flash-Next, a 125-billion-parameter model, compressed to fit a Mac with 48 GB of memory. We didn't retrain it. Every weight comes from Qwen's release. What we changed is how those weights are stored and which parts stay in memory.
About 27 GiB of it sits in memory. Our earlier build, v0.1, needs 35 GiB, which is more than a 48 GB Mac can set aside for the GPU, so it slows down sharply there. v0.2 fits, with conversations of up to 32,000 tokens. A huge 54 GiB lookup table stays on your SSD, and Osaurus reads only the few rows each word needs.
To get there, the model's experts use between 1 and 2 bits per weight, chosen layer by layer from measured error. On our everyday tasks (finance, reasoning, coding, terminal work and tool calls) 93% of its answers were correct, with no loops. Speculative decoding made tool-calling turns 27 to 36% faster.
It's the smaller model, not the better one. It agrees with the original model's next word 75% of the time, against 81% for v0.1, and it reads text only. If your Mac has 64 GB or more, v0.1 is still the better choice.
You need Osaurus 0.25.24 or newer and about 86 GB of free disk space. The model page explains how it works in plain language first, then in full detail: what stays in memory, the new expert format, how close it gets to the original, the sampling settings we measured, and speed.