v0.72.0

v0.72.0

Highlights

• Added the public meshllm.cloud website, catalog viewer, onboarding docs, canonical docs URLs, and refreshed install documentation.
• Hardened release packaging around versioned native runtimes, explicit runtime version selection, release-matrix validation, runtime install smoke checks, relocatable shared libraries, CUDA profile detection, and Windows runtime packaging.
• Improved Skippy split serving with latency-aware stage planning, faster decode return paths, topology-aware prefill chunking, repeated-prompt cache replay, KV pressure fixes, cache epoch telemetry, stage role metadata, and native runtime event visibility.
• Expanded model packaging and conversion flows with native skippy-quantize conversion/quantization CLI support, direct quantized layer package workflows, inline package-writer/download progress, native MTP metadata, GLM DSA IndexShare support, and GLM-DSA crash fixes.
• Strengthened mesh routing and reliability with LAN-only mDNS hardening, bounded direct-path repair, iroh 1.0, stale split topology withdrawal, local target reputation routing, split-startup diagnostics, and safer model reconciliation.
• Improved MoA and OpenAI-compatible behavior for agent/tool loops, including committed Responses streaming, usable-contributor context advertisement, small-model consensus safeguards, low-quality response retries, and chat/thinking fixes.
• Revamped configuration with schema-driven controls, clearer runtime/model/network settings, validation diagnostics, UI-backed config behavior, and more realistic control surfaces.
• Improved SDK and embedding surfaces, including the native Rust SDK, native runtime install APIs, SDK packaging updates, and refreshed Node/Swift/Kotlin bindings.
• Fixed Windows installer and binary compatibility issues, including PowerShell 5.1 irm | iex, missing checksum sidecars, and crashes on CPUs without AVX-512.
• Tightened CI and release workflows with restored PR gates, sccache build-script support, native runtime matrix checks, nightly stability fixes, HF model cache/smoke hardening, and faster release builds.

Notable Fixes

• Fixed CUDA runtime profile detection for release/native-runtime selection.
• Fixed release publish credential preservation and Windows CUDA runtime DLL handling.
• Fixed chat system prompt feature flag gating.
• Fixed model deletion cleanup for broken hf_hub cache symlinks.
• Fixed /api/models repeatedly re-downloading catalog data.
• Fixed GPU CLI stderr behavior and improved TUI panic/debug capture.
• Fixed insecure-origin chat ID behavior.
• Fixed Metal Skippy shutdown race.
• Fixed release version stamping and runtime startup initialization.

Contributors

Thanks to @IvGolovach, @i386, @michaelneale, @ndizazzo, and first-time contributor @wesbillman.

Full Changelog

• Current run compare: v0.71.0...100f548
• Final tag compare, once published: v0.71.0...v0.72.0