The supplied public source provides two concrete operating facts: its command downloads a GGUF file described as approximately 5.95 GB, and it requires Prism’s llama.cpp fork to run it. The author reports roughly 20 tokens per second on an M5 Pro, then 44 tokens per second after restarting the server. Startup also produced a message that the Metal tensor API was unsupported in that environment and had been disabled.
The mechanism is packaging compression and runtime support together: the smaller model artifact lowers storage and local-serving requirements, while a modified inference stack interprets its format. This matters now because the release is accompanied by a runnable local path instead of only an abstract compression claim. The implication is that teams can test 27B-scale capability under local-data and local-cost constraints, but must treat runtime compatibility as part of the deployment decision.
Evidence boundary: the source is a single practitioner report, not an independent quality evaluation, a controlled throughput benchmark, or proof of the release title’s near-lossless and 9x claims. Watch whether Prism support reaches mainstream runtimes and whether independent evaluations reproduce quality, throughput, and stability across hardware. This starts a distinct analytical record because no prior New Runtime record was supplied for this model-and-runtime release.