backend/v3.1.0 v0.6.0
ggml-org/llama.cppOct 1, 2026 → Oct 5, 20264 days between releases
What's New
llama.cpp v0.6.0 adds extended batch APIs, new models, and Metal optimizations.
Breaking Changes
- Session formats bumped to `LLAMA_SESSION_VERSION` 11 and `LLAMA_STATE_SEQ_VERSION` 4.
- `mtmd_get_memory_usage()` signature updated to new struct.
- `/v1/embeddings` endpoint now requires typed content for vision/audio/video inputs.
New Features
- Extended batch API (`llama_batch_ext`) for mixed token/embedding inputs.
- Support for GLM-5.3-Flash (320B) hybrid model.
- Support for Clef decision model (text and vision).
- New `/v1/systemone` server API endpoint.
- Metal flash attention kernel for F16 KV.
Improvements
- Qwen4Exp support with MTP speculative decoding.
- Web UI overhaul with Hugging Face Hub integration.
- Vulkan sparse flash attention for quantized K/V.
- ggml library updated to v0.26.0.
Bug Fixes
- Various correctness fixes for Qwen4Exp.
Upgrade Notes
You must update session serialization logic to handle the new `LLAMA_SESSION_VERSION` 11 and `LLAMA_STATE_SEQ_VERSION` 4 formats. The `mtmd_get_memory_usage()` function now requires the new `mtmd_memory_usage` struct. The `/v1/embeddings` endpoint now expects typed content for vision, audio, or video inputs.
The Two Releases
Hanko v3.1
Oct 1, 2026· 21 days ago
Hanko v3.1 introduces Organizations and Roles for B2B applications, Custom Claims and Mapping to normalize identity data, and session lifecycle webhooks. It also addresses a security vulnerability by changing the default client IP extraction method.
Full release notesv0.6.0
Oct 5, 2026· 16 days ago
llama.cpp v0.6.0 introduces a new extended batch API for mixed token/embedding inputs, adds support for the GLM-5.3-Flash (320B) and Clef decision models, and implements a new `/v1/systemone` API. It also includes Metal flash attention optimizations and a Web UI overhaul.
Full release notes