heretic Releases

Maintenance release fixing bugs in export (ONNX/TFLite background logit slot, keypoint tensor identification) and augmentation (keypoint flip, Albumentations TimeReverse/SquareSymmetry). Fixes `BestModelCallback` scoring sanity checks.

Aug 24, 2026

Maintenance release improving inference speed (CUDA sync), fixing YOLO test-split evaluation, and ONNX/TFLite multi-label detection dropping. Improves auto-batch memory estimation.

Aug 17, 2026

Fix-and-performance release fixing hierarchical COCO label mapping and improving matcher performance. Restores per-callback state on training resume.

Aug 11, 2026

This patch release adds support for quantized DSpark Markov heads on top of the v0.27.0 baseline.

Aug 11, 2026

A major release featuring 561 commits, introducing Kimi K3 support, Qwen3.5 models, PyTorch 2.13 upgrade, and significant enhancements to FlashAttention and Model Runner V2.

Aug 10, 2026

This patch release exposes granular configuration options for the Phonic realtime plugin, allowing users to fine-tune behavior through settings like VAD tuning, MCP servers, and observability integrations.

Aug 6, 2026

A patch release focused on dependency updates to align the Trugen plugin with the core LiveKit agents library.

Aug 6, 2026

A patch release focused on dependency updates to align the Silero plugin with the core LiveKit agents library.

Aug 6, 2026

A patch release focusing on dependency updates to align the XAI plugin with the core LiveKit agents library and related plugins.

Aug 6, 2026

A patch release focusing on dependency updates to align the Baseten plugin with the core LiveKit agents library.

Aug 6, 2026

Maintenance release making inference and export faster and more correct. Fixes segmentation post-processing speed and export resize parity.

Aug 4, 2026

Major release adding configurable training, ExecuTorch & CoreML export, and multi-GPU training support. Removes deprecated import paths.

Jul 29, 2026

A major release introducing the new Inkling model family, DeepSeek-V4 performance optimizations, and fp32 lm_head support for improved generation accuracy. It also features flexible attention backends, matured KV offloading capabilities, and an updated Rust frontend with multimodal support.

Jul 27, 2026

A patch release containing two targeted bug fixes to prevent startup blocking and data corruption in quantization scenarios.

Jul 14, 2026

A major release that makes Model Runner V2 the default execution path, removes legacy PagedAttention, optimizes the Transformers backend, and introduces support for new models and advanced decoding techniques.

Jul 11, 2026

Introduces architectural locking for coordination placement, environment variable support for pack paths, and review-time regression gates. Fixes bugs related to stale assertions, prompt accumulation, and skill retirement.

Jul 8, 2026

This reliability-and-trust release fixes everyday mission lifecycle friction points, closes a bulk-edit safety gap, and improves SaaS sync reporting. It also reshapes the CI pipeline and decomposes internal command modules for faster feedback and lower regression risk.

Jul 5, 2026

A stabilization-and-foundations release that hardens behavior under non-trivial branch topologies, smooths the day-to-day planning loop through usability upgrades, and establishes a governed documentation foundation.

Jun 29, 2026

A patch release focused on stabilizing coordination/primary surface resolution (split-brain remediation), hardening CI test coverage, and decomposing the monolithic agent/tasks.py module.

Jun 24, 2026

A patch release stabilizing functionality from v3.2.0 by fixing orchestrator resolution, charter status coherence, and CI pipeline issues, alongside security dependency updates.

Jun 18, 2026