lemonade-sdk/lemonade candidate-v2026.44.1 on GitHub

Headline

Model loading on shared-memory (iGPU/APU) systems fixed, enabling larger models by using total physical RAM instead of RAM-in-use for the context-size auto-tune pool cap.
Stable diffusion ROCm backend restored on Windows with full HIP runtime staging, preventing silent CPU fallback.
LFM2.5-24B-A2B-GGUF added to the server model registry with chat, reasoning, and tool-calling labels.
Deleted models now remove their files from disk instead of leaving orphaned checkpoints.

The /v1/models/{model}/options endpoint's resolved_ctx_size field can now return null instead of a numeric value when automatic GPU sizing cannot fit a valid context — clients must handle null gracefully.

Operating System	Downloads
Windows	lemonade.msi
Ubuntu 24.04+	Launchpad PPA
Debian 13 (x86_64)	lemonade-server_2026.44.1-debian13_amd64.deb
Debian 13 (ARM64)	lemonade-server_2026.44.1-debian13_arm64.deb
Fedora 43 (x86_64)	lemonade-server-2026.44.1-fc43.x86_64.rpm
Fedora 43 (ARM64)	lemonade-server-2026.44.1-fc43.aarch64.rpm
Fedora 44 (x86_64)	lemonade-server-2026.44.1-fc44.x86_64.rpm
Fedora 44 (ARM64)	lemonade-server-2026.44.1-fc44.aarch64.rpm
macOS	Lemonade-2026.44.1-Darwin.pkg

Other platforms? See our Installation Options for Docker, Snap, Arch, Debian, and more.

Portable binaries for bundling into your own installer. Run lemond ./ as a subprocess.

Platform	Download
Ubuntu x64	lemonade-embeddable-2026.44.1-ubuntu-x64.tar.gz
Ubuntu arm64	lemonade-embeddable-2026.44.1-ubuntu-arm64.tar.gz
Windows x64	lemonade-embeddable-2026.44.1-windows-x64.zip
macOS arm64	lemonade-embeddable-2026.44.1-macos-arm64.tar.gz

Click to expand changelog

Fixes 3267: Model delete keeps files on the disk by @meghsat in #3765
fix(ci): avoid SIGPIPE race in launchpad upgrade assertions by @sreeram-11 in #3860
Add conformance tests (phase 2: model-backed) by @QuTosh in #2865
Add LFM2.5-24B-A2B by @sawansri in #3866
chore(flm): bump FLM to v1.0.9 by @zaneni6 in #3873
ci: group inference suites by runner class to reduce handoffs by @sreeram-11 in #3861
fix(auto-tune): improve context sizing to avoid GPU memory pressure by @fl0rianr in #3862
Roll back FLM to 1.0.7 release by @superm1 in #3903
Restore sdcpp rocm support on Windows by @sreeram-11 in #3718
docs: extra_models_dir under /home needs a drop-in on deb/rpm by @bernardogv in #3895
Improve lemonade launch pi integration by @sawansri in #3786
fix(backends): resolve a *_bin directory to the binary inside it by @bernardogv in #3893
ci: publish news posts to the website as soon as they merge by @jeremyfowers in #3884
fix(server): report AMD GPU usage and VRAM on Windows by @iswaryaalex in #3872
fix(ci): stop Hugging Face rate limits from failing model pulls by @sreeram-11 in #3859
(ci) : Adding halo box backend benchmarking support by @iswaryaalex in #3715
News: auto-BKC vision blog post by @jeremyfowers in #3846
News: edit auto-opt blog based on feedback by @jeremyfowers in #3921
[v2026.44] fix(auto-tune): never refuse a load over a context estimate (backport #3986) by @jeremyfowers in #3994

Full Changelog: v2026.43.3...candidate-v2026.44.1

The Windows installers in this release candidate are not signed; Windows SmartScreen will warn before installing. Stable releases are signed.