TensorFold 1.2.3 ships faster Flash Next: +21% GB10 decode, M3 Ultra 64K-context reply in 8s
HankYeomans · x · 2026-11-06
Open-source local inference tool TensorFold released v1.2.3 with measured Flash Next engine improvements:
- Two GB10s with NVFP4 at TP=2: plain decode 21–23% faster, saving 2.62 GiB of model memory per rank.
- M3 Ultra parallel-serving test: a 1,024-token reply beside a 64K prompt finished in 8 seconds instead of 48.
Also new: PDF input, Ornith-1.5-122B on Metal, native dots.mocr OCR, Qwen embeddings/reranking on HIP, LFM3-1.7B on Intel Iris Xe; fixes for prompt caching, memory use, streamed tool calls and serving. Apache-2.0, Metal + CUDA, OpenAI-compatible API, speculative drafted decoding.
More from Infra
- sshhole Bug Analysis: Alpine/MUSL Systems Hit Hard, ~100-500M OpenWRT Routers Were at Risk — esrtweet · 2026-11-06
- Smarter Compact: MCP server trims agent context, 72% cheaper long sessions with full recall — goost95 · 2026-11-06
- TensorFold 1.2.3 boosts Flash Next decode 21–23%; Apache-2.0 Ripple-65B Preview lands — AIFlow_ML · 2026-11-06
- "What am I supposed to do if the data center bubble pops" — AI infra circle jokes — max_paperclips · 2026-11-06
- Inkling 2 Small pretrained on 8.4T tokens in 68 hours, natively in NVFP4 — clarejtbirch · 2026-11-06
- Stable reinforce handles 8-minute rollouts with FP4 weights in open RL training — simonguozirui · 2026-11-06