Dev ports MiniMax H3 to Swift, cuts 480p video gen to ~4 min on a 16GB M1 MacBook
Lokoto123 · reddit · 2026-11-19
Using Claude, a developer ported MiniMax H3 (21GB 8-bit video transformer + 27GB Qwen3-VL encoder) to a native Swift engine on a 16GB M1 Pro, cutting 3-second 480p generation from 7:46 to 3:54 and 5-second clips to 6 minutes.
Key techniques:
- Bit-for-bit memory-mapped weights without re-quantization; RoPE channel permutation for the attention kernel
- Full-precision witnesses for verification (Python pipeline's half-precision activations were 2–25% off)
- Best plan: GPU plus CPU AMX sharing output columns; 8-bit activations and Neural Engine were slower
- Shortcuts: 3-step LoRA with merged schedule, multi-core decoder precision conversion, continuing clips from 5 held latent frames
Next targets: 720p and faster upscaling (Topaz currently takes 30 minutes).
More from Infra
- Swapping two used RTX 3090s for two Radeon R9700s: is ROCm ready for local inference? — Pear_Virtual · 2026-11-19
- Hugging Face's open-source nudge: 100GB private storage vs ~7TB for public data — RisingSayak · 2026-11-19
- Databricks explains LTAP: the lake is the database's storage, killing the CDC pipeline — matei_zaharia · 2026-11-19
- Solo dev cuts sandbox cold-start to 0.45s, sleeping VMs cost 1¢/month — emilahlback · 2026-11-19
- Team cuts agent sandbox cost 88% by moving to GKE, open-sources gke-sandboxes — emilahlback · 2026-11-19
- Open-source prefill engine hits 3.2× vLLM throughput serving GLM-5.3-Flash on 4 B200s — bariskasikci · 2026-11-19