v1.4.1 v0.27.0

Name: immich
Author: immich-app

high risk

immich-app/immichAug 7, 2026 → Aug 10, 20263 days between releases

What's New

Adds Kimi K3 support, Rust gRPC frontend, and FlashAttention 4 integration alongside PyTorch 2.13 upgrade.

Breaking Changes

PyTorch 2.13.0 upgrade is a breaking environment change.
XPU and CPU backends updated to PyTorch 2.13

New Features

Kimi K3 support with full stack landing including Python and Rust frontends.
Rust frontend gRPC control plane introduced.
FlashAttention 4 integration with FP8 KV cache support.
Model Runner V2 expansion for encoder-only and pooling workloads.
DeepSeek-V4 performance optimizations including sequence parallelism.

Improvements

DeepSeek-V4 performance push reduces TTFT and improves kernel efficiency.
FlashAttention 4 integration adds headdim-256 support.
JIT warmup infrastructure removes first-request compilation stalls.
Support for new models including Qwen3.5, K-EXAONE-2.0-750B-A37B, and VaultGemma.
DeepGEMM support and compressed-tensors quantized checkpoints added.

Bug Fixes

Removes first-request compilation stalls via JIT warmup infrastructure.
Fixes performance by skipping empty c128 kernel launches.
Improves TTFT by skipping unneeded topk/router operations.
Improves TTFT via workspace reuse optimization.

Upgrade Notes

Requires updating to PyTorch 2.13.0, a breaking environment change. XPU and CPU backends are also updated to this version.

The Two Releases

v1.4.1

Aug 7, 2026· 3 months ago

A maintenance release focusing on self-hosted administrator tools, including a workspace member reactivation command, and significant security hardening via dependency upgrades.

Full release notes

v0.27.0

Aug 10, 2026· 2 months ago

This release introduces full Kimi K3 support, upgrades to PyTorch 2.13.0, and expands Model Runner V2 capabilities to non-generative workloads while optimizing DeepSeek-V4 performance.

Full release notes