signerless/llm-checker v3.8.2

v3.8.2: Exact catalog metadata and hardware fixes

on GitHub

Published to npm as llm-checker@3.8.2.

Fixes merged since 3.8.1

• Strix Halo and other AMD APUs: unified memory is recognized, and 256 MB
  iGPU apertures are no longer read as 256 GB (#113, #134).
• Uncensored models are only recommended when you ask for them (#114).
• ai-check reuses the canonical GPU detection, and gpu-plan no longer
  counts the same GPU twice across backends (#116, #117).
• New --cpu-only mode forces CPU/RAM execution even when a GPU is
  detected (#119).
• Independent hardware probes run in parallel (#127).
• Local speed measurements survive catalog rebuilds (#128).
• All six benchmark sources feed model recommendations (#129).
• Non-language repositories are excluded from recommendations (#131).
• Evaluation counts are reported per category (#132).
• The registry uses native SQLite, with a WASM fallback on older Node
  versions (#133).
• llama.cpp, Transformers and explicit runtime choices are honored (#135).
• Ranking artifacts, benchmark identity and desktop compatibility (#136).
• Two hardware tests simulate Linux explicitly, so the suite also passes on
  macOS.

Exact catalog metadata (#138)

• Hugging Face ingestion lists repos per language task and sweeps official
  publishers, so the limit is spent on models the recommender can rank. The
  weekly snapshot build had failed since July because the global download
  ranking left too few language-model repos.
• Exact parameter counts, dtypes, GGUF context windows and lineage come from
  the Hub's expand[] metadata; registry-sync --deep adds observed file
  sizes, SHA-256 hashes and config.json context windows. Complete shard sets
  are one artifact; projectors, importance matrices and duplicate exports are
  not artifacts. Official Mistral releases (no pipeline tag) are included.
• The Ollama sync reads each tag's size, context window, input types and
  digest, excludes cloud-only tags, and no longer invents sizes, Q4_0 or
  4096-token defaults. sync --exact reads byte sizes, quantizations and
  parameter counts from registry.ollama.ai.
• Recommendations use each Ollama tag's own context window and inputs.
• Existing databases adopt a newer packaged snapshot; sync keeps the
  registry's Ollama rows instead of deleting them.
• MCP recommend and project_recommend honor the requested category.
• The packaged snapshot ships benchmark scores; existing databases adopt
  them unless quality-sync refreshed a source more recently. New
  hf_eval_results source (self-reported model-card results), and LiveBench
  releases newer than the curated list are discovered automatically.
• Ollama tags record their license from the manifest (sync --exact).
• New registry source: ModelScope (--sources modelscope), with exact file
  sizes, hashes, config.json context windows and modelscope download /
  VLLM_USE_MODELSCOPE commands.
• New registry source and runtime: Docker Model Runner (docker), from
  Docker Hub's official ai/ namespace, with pull/run commands and
  format-aware hardware checks (MLX tags on Apple Silicon, vLLM tags on
  NVIDIA).
• Base/pretrained checkpoints rank below instruct builds; Ollama aliases and
  quantizations of one base checkpoint collapse into one recommendation.
• check on Apple Silicon ranks the synced catalog like other platforms
  instead of only the 26 static definitions; check, ai-check and ai-run
  read each model's real parameter count, size and context window instead of
  guessing from its name. The curated fallback catalog was refreshed from
  registry data (wrong families and non-existent tags fixed), and the unused
  scripts/update-models.js, which injected invented models, was removed.