v1.6.1 v0.76.0
silvia-odwyer/photonSep 2, 2026 → Sep 10, 20268 days between releases
What's New
Major update introduces durable KV cache, SafeTensors support, and removes legacy KV persistence.
Breaking Changes
- Removed persistence of KV cache state to disk
- Removed runtime-slice architecture allowlist
- Consolidated model=auto and model=mesh routing into a single directive
New Features
- Durable KV prefix cache that survives eviction and restart
- Skippy-quantize for GGUF composition
- Load SafeTensors checkpoints directly
- Iteration-level scheduler for concurrent staged serving
- Stream tool calls as they are generated
Improvements
- Unified radix for prefix caches and iteration-level scheduler
- Expose model-bound tokenizer capability
- Route repeated prompts to workers with verified radix cache reuse
- Pass reasoning effort through Skippy templates
Upgrade Notes
Remove any configuration related to legacy KV persistence and runtime-slice allowlists. Update model routing directives to use the consolidated `model` flag. Migrate checkpoints to SafeTensors format if applicable.
The Two Releases
v1.6.1
Sep 2, 2026· 2 months ago
This release focuses on bug fixes for connection handling and the TUI, updates model recommendations, and cleans up documentation.
Full release notesv0.76.0
Sep 10, 2026· 1 month ago
A major release introducing durable KV prefix cache, unified radix caching, and SafeTensors checkpoint support. It also enforces configuration schema parity and wires significant hardware placement and speculative decoding settings.
Full release notes