github nace-ai/kapitsa v0.2.13
Kapitsa 0.2.13

latest releases: v0.2.31, v0.2.30, v0.2.29...
13 days ago

What's Changed

  • Benchmark: Mac mini (2024) · M4 Pro · 64 GB · Qwen3.8 27B 4-bit by @chrisbensch in #397
  • Benchmark: MacBook Air (15-inch, 2026) · M5 · 32 GB · Qwen3.8 27B 4-bit by @chrisbensch in #398
  • Flash-Next holds a run of its layers per process, the rest of each pass on a Remote: the model side of splitting it across Macs (unsplit logits bitwise the same) by @fearampgreed in #385
  • Benchmark: Mac mini (2024) · M4 Pro · 64 GB · Recursive Flint 40B-A2.5B 4-bit by @bobosola in #403
  • Several Macs, 1 of 4: what a Kapitsa server tells the others (MacInfo, Bonjour TXT, GET /cluster JSON), where a Mac too small for a model sends its requests, kapitsa cluster's table; cluster_test, CPU only by @fearampgreed in #399
  • Where a split model's layers go: each Mac's run of layers in its memory with the fastest pass (src/cluster/place) by @fearampgreed in #402
  • Several Macs, 2 of 4: finding the Macs that run Kapitsa over Bonjour (_kapitsa._tcp through the system's dns_sd.h, no new dependency): Advertiser, Browser with each server's round trip and link; cluster_test --live by @fearampgreed in #404
  • Split Flash-Next across Macs, the wire and the head's link: each pass around the Macs once, rows as fp32, the layers placed by memory and speed (src/cluster/pipe) by @fearampgreed in #406
  • Several Macs: kapitsa serve on the network by default (no password), routing to a Mac that runs a model whole, kapitsa cluster, kapitsa stage / serve --split, and two Flash-Next Macs pairing on their own by @fearampgreed in #407
  • Split: the ring's return port takes only the connection with the head's token; a stage that closes during setup ends the wait at once; stage answers read without exceptions by @fearampgreed in #408
  • Benchmark: Mac mini (2024) · M4 · 24 GB · Qwen3.6 35B-A3B 2-bit experts by @willeyh-git in #409
  • POST /v1/draft: read the user's draft ahead while they type (M4 Pro 48 GB, 2,000-token paste's first token: 4B 1.27 -> 0.03 s, 35B-A3B 2.0 -> 0.03 s) by @fearampgreed in #384
  • Several Macs: servers named by address (--peer, KAPITSA_PEERS, ~/.cache/kapitsa/peers) for networks that block Bonjour between Macs by @fearampgreed in #410
  • Kapitsa 0.2.13 by @fearampgreed in #411

New Contributors

Full Changelog: v0.2.12...v0.2.13

Don't miss a new kapitsa release

NewReleases is sending notifications on new releases.