philipjohnbasile/DeepSeek-V4-Flash-0731-MLX-M5Max-TargetOnly
Text Generation • 284B • Updated • 1.75k • 2
Apple Silicon — Experimental Models
Research artifacts with documented limits: target-only DeepSeek, AR-only Akka, and small-model FIM/MTP experiments with Wisp.
Note Superseded experimental reference derived from DeepSeek's official model. Tested on a 128 GB M5 Max with a companion runtime; roughly 100 GB peak memory in the recorded smoke. No MTP. Published results include quality and response-format regressions.
philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX
Text Generation • 27B • Updated • 303 • 1Note My 6-bit MLX conversion of nightmedia's Akka merge. Text-only, autoregressive generation; the MTP head was omitted after 0% draft agreement in calibration. About 20 GB of weights. Includes the calibration failure and same-harness perplexity comparison.
philipjohnbasile/wisp-coder-110m
Text Generation • 0.1B • Updated • 839 • 2Note Research on fill-in-the-middle and native multi-token prediction at small scale. The registered short-hole study found no superiority over the compared FIM models. Standard inference uses the trunk; MTP reference decoding is a correctness path, not a speed claim.