lemonade-sdk/lemonade candidate-v2026.43.0 on GitHub

Headline

Multi-reference image editing now accepts one or more reference images with automatic dimension sizing, preserving original image bytes without re-encoding.
TheNoise image backend moves from Experimental to Core tier for supported AMD GPUs on Linux and Windows.
Live inference tracking reports active request count and estimated tokens per second on GET /v1/stats.
Prefill throughput (TTFT-based prompt tokens per second) added to the lemonade bench table and JSON output.

GET /api/v1/system-stats now reports gpu_percent and vram_gb for the primary GPU by total memory instead of the most utilized GPU — on multi-GPU systems the values change and existing dashboards or alerting that depends on these fields will see different numbers.
Guest-tier models are now hidden from GET /v1/models?show_all=true and return model_not_supported on pull until the guest backend is installed — setups that relied on the model being visible without installing the backend will see them disappear.

Operating System	Downloads
Windows	lemonade.msi
Ubuntu 24.04+	Launchpad PPA
Debian 13 (x86_64)	lemonade-server_2026.43.0-debian13_amd64.deb
Debian 13 (ARM64)	lemonade-server_2026.43.0-debian13_arm64.deb
Fedora 43 (x86_64)	lemonade-server-2026.43.0-fc43.x86_64.rpm
Fedora 43 (ARM64)	lemonade-server-2026.43.0-fc43.aarch64.rpm
Fedora 44 (x86_64)	lemonade-server-2026.43.0-fc44.x86_64.rpm
Fedora 44 (ARM64)	lemonade-server-2026.43.0-fc44.aarch64.rpm
macOS	Lemonade-2026.43.0-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-2026.43.0-ubuntu-x64.tar.gz
Ubuntu arm64	lemonade-embeddable-2026.43.0-ubuntu-arm64.tar.gz
Windows x64	lemonade-embeddable-2026.43.0-windows-x64.zip
macOS arm64	lemonade-embeddable-2026.43.0-macos-arm64.tar.gz

Thanks @bitgamma, @fl0rianr, @jeremyfowers, @joshuacox, @sreeram-11, @zaneni6 for your awesome contributions to this release!

Click to expand changelog

refactor(server): run backend servers through ServerProcess and NativeProcess by @jeremyfowers in #3705
fix(server): support multi-reference image editing by @fl0rianr in #3801
fix(config): read the distro defaults override from lemond's own install prefix by @jeremyfowers in #3775
feat(images): add per-model edit size caps and Qwen-Image-2.1 Turbo INT8 by @fl0rianr in #3803
fix(models): preserve catalog identities in extra_models_dir by @fl0rianr in #3805
fix(models): prevent duplicate discovery of incomplete shards by @fl0rianr in #3811
docs: add systemd start step to the Arch install page by @joshuacox in #3784
ci: bump auto-labeler to Claude Haiku 5.5 by @fl0rianr in #3774
feat(backends): label each backend with a tier and a format by @jeremyfowers in #3706
Update Lemonade to ROCm 10.1.0 by @sreeram-11 in #3821
chore(flm): bump FLM to v1.0.8 by @zaneni6 in #3810
Build cpp-httplib once instead of in every file that includes it by @jeremyfowers in #3837
fix(server): report NVIDIA GPU and VRAM usage by @fl0rianr in #3282
mark TheNoise as Core-tier by @bitgamma in #3826
test: pre-pull server_llm.py's models and cover implicit pull on inference by @jeremyfowers in #3839
lemonade bench: add prefill TPS by @bitgamma in #3759
Add AnythingLLM and Phoronix Lemonade 11.9 articles to news page by @jeremyfowers in #3836
fix(server): expose live inference throughput across backends by @fl0rianr in #3791
test: isolate every C++ test from the real config dir by @jeremyfowers in #3612

Full Changelog: v2026.42.1...candidate-v2026.43.0

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