Skip to content
Releases

Versions and compatibility

A released profile is immutable. Changing the model, revision, precision, runtime, flags, prompt set, token shape, assets, step count, seeds or metric definitions is a major version bump that creates a separate leaderboard.

Current profile versions

ProfileVersionStateModelRuntimeProfile hash
diffusion-enthusiast-base0.1.0Release candidatestabilityai/stable-diffusion-3.5-largediffusers 0.39.0142edc9e59f5b3…
diffusion-entry-base0.1.0Release candidatestabilityai/stable-diffusion-xl-base-1.0diffusers 0.39.0d1eed7cc6558ed…
diffusion-frontier-base0.1.0Release candidateblack-forest-labs/FLUX.1-schnelldiffusers 0.39.0e35d8f0880ddcb…
llm-enthusiast-base0.1.0Release candidateQwen/Qwen3-8Bvllm 0.26.061ec8da4e3a2e0…
llm-enthusiast-int40.1.0Release candidateQwen/Qwen3-32B-AWQvllm 0.26.04acbd6459badce…
llm-enthusiast-nvfp40.1.0Release candidatenvidia/Qwen3-32B-NVFP4vllm 0.26.04643cf31dded44…
llm-entry-base0.1.0Release candidateQwen/Qwen3-4B-Instructvllm 0.26.0c976c1bd573623…
llm-entry-int40.1.0Release candidateQwen/Qwen3-8B-AWQvllm 0.26.0786cba1eae562a…
llm-entry-nvfp40.1.0Release candidatenvidia/Qwen3-8B-NVFP4vllm 0.26.066f58d0e43ecc1…
llm-frontier-base0.1.0Release candidateQwen/Qwen3-32Bvllm 0.26.098cfb0979cbf81…
llm-frontier-int40.1.0Release candidateQwen/Qwen2.5-72B-Instruct-AWQvllm 0.26.05ac4055c2a6495…
llm-frontier-nvfp40.1.0Release candidatenvidia/Qwen2.5-72B-Instruct-NVFP4vllm 0.26.08c27b806d4f34a…
vision-enthusiast-base0.1.0Release candidateQwen/Qwen3-VL-8B-Instructvllm 0.26.0c5243aa080b5c3…
vision-entry-base0.1.0Release candidateQwen/Qwen3-VL-4B-Instructvllm 0.26.070b66e37e396f7…
vision-frontier-base0.1.0Release candidateQwen/Qwen3-VL-32B-Instructvllm 0.26.0405cabdb56c02c…

Version policy

Major
Anything that changes a measured number: model, revision, precision, runtime, runtime flags, workload shape, assets, seeds, metric definitions or ranking gates. A new leaderboard; old results keep theirs.
Minor
A new optional workload, or a new secondary metric that does not affect the rank. Existing results stay comparable and simply lack the new column.
Patch
Display names, summaries, comments and documentation. Nothing that could move a value. Results are unaffected.

Release candidates

Profiles at version 0.x are not frozen. Their model revision, runtime version and harness commit are still being pinned against the tier's minimum-VRAM reference system, and their leaderboards are provisional. Results collected against a 0.x profile are ranked only against that same version, and are not migrated when it freezes.

Freezing a profile requires conformance data from at least two GPU classes including the tier minimum, a repeatability coefficient of variation at or under 3%, and a completed licence review of every model and asset it references.

Container images

Every released image is multi-architecture (linux/amd64 andlinux/arm64), signed with Cosign, and published with an SBOM and build provenance. A result reaches Verified only if its recorded image digest matches a released one.

ghcr.io/pavel4ai/localmax-llm
ghcr.io/pavel4ai/localmax-vision
ghcr.io/pavel4ai/localmax-diffusion

Released digests are tracked incontainers/released-images.jsonand compiled into the validator, so the accepted list is itself public and versioned.