Colibri streams a 744B MoE off disk to run on a 24GB laptop
The pure-C engine treats storage as RAM, so models that never fit now run — plus a local ElevenLabs, Alibaba's AI code reviewer, and an Ollama config break.
Copy markdown
744B parameters on the RAM you already own
Colibri is a pure-C engine (zero runtime deps, GPU optional) that streams MoE experts from disk on demand, so a model's weights never have to fit in memory. Eight families run today — GLM-5.2 (744B), Kimi K3 (2.8T), DeepSeek V4 Flash — behind an OpenAI-compatible API, from a 16GB floor. The honest catch: a bare 24GB box decodes at ~0.05–0.1 tok/s, so it makes the impossible possible, not fast.
VoiceStudio: a fully-local ElevenLabs
VoiceStudio wires 16 TTS engines (OmniVoice, CosyVoice 3, GPT-SoVITS) and 11 STT engines (WhisperX) behind OpenAI-compatible /v1/audio endpoints, a WebSocket stream, and an MCP server for Claude and Cursor. Voice cloning, dubbing, and audiobook rendering span 646 languages, stay on-disk by default, and run CPU-only from 8GB RAM — build a private dubbing pipeline this weekend with no per-character bill.
Alibaba open-sources its code reviewer
open-code-review pairs deterministic pipelines with an LLM agent to post precise line-level comments, shipping a built-in ruleset for NPE, thread-safety, XSS, and SQL injection across languages. It's OpenAI- and Anthropic-compatible, so you point it at your own key and drop it into CI — the same reviewer Alibaba runs at scale.
Heads-up: Ollama 0.34.1 drops typical_p
Ollama's latest stops honoring the typical_p sampling parameter when you create new models (existing GGUFs still work), so audit any Modelfiles that set it before you upgrade. The release also cuts /api/tags lookups from 3.1s to 294ms and improves Apple Silicon MLX memory handling.