github AlexsJones/llmfit v1.1.16

7 hours ago

1.1.16 (2026-09-19)

Features

  • concurrency: concurrent-session capacity estimator (#140) (#999) (1936acc)
  • models: Sept 2026 model refresh (GLM-5.3, Qwen3.8-Flash-Next, DeepSeek-V4.1, Kimi-K3 and more) (#1055) (6dd1f3c)
  • recognize native ternary (1.58-bit) models (#886) (569a9ac)
  • storage: add disk planning for model libraries (#1023) (42c2641)

Bug Fixes

  • bench: error when the requested model is not available (#1041) (f66aad9)
  • bench: ignore implausible tok/s from degenerate Ollama timings (#1042) (86a0d08)
  • bench: stop wall timer after body read in OpenAI and Ollama paths (#1037) (63753e9), closes #1028
  • fit: size and price MXFP4-native models at MXFP4 (#1059) (650cd1e)
  • hardware: recognize A-series Apple Silicon unified memory (#1044) (e138964)
  • plan: grade GPU paths against free VRAM, not total capacity (#1058) (0d48118), closes #835
  • providers: find HF models in HF_HUB_CACHE and XDG_CACHE_HOME (#1051) (83d4963)
  • providers: map Gemma 4 instruct models to their Ollama tags (#1043) (a172f2a), closes #1024
  • providers: map gemma4:12b and other Sept 2026 catalog additions to Ollama tags (#1056) (01c805c), closes #1024
  • scraper: budget GGUF source probing so the weekly run can finish (#1061) (44de365)
  • scraper: coerce list-valued expert counts; a failed estimate cannot abort a scrape (#1064) (84007d5)
  • scraper: never size hybrid SSM models by the estimator; cap by declared size (#1063) (2dd0628)
  • scraper: pace HuggingFace requests on the ratelimit headers (#1047) (c0fd811)
  • scraper: revalidate retained catalog entries (fixes packed AWQ/GPTQ/INT4 parameter counts) (#1057) (a4b669c)
  • update: read the context window from config.json (#1054) (9db4a4c), closes #1021
  • update: request every mapped field in the HuggingFace list expand[] (#1048) (b662f4d), closes #1021

What's Changed

  • bench: community results for intel-raptor-lake-p-iris-xe-graphics-integrated by @FabioLeitao in #1004
  • chore(deps): bump the cargo-minor-patch group with 2 updates by @dependabot[bot] in #1005
  • feat(concurrency): concurrent-session capacity estimator (#140) by @Akciali in #999
  • bench: community results for apple-m2-pro by @perrwa in #1031
  • bench: community results for nvidia-geforce-gtx-1650 by @Land5hark in #1030
  • bench: community results for nvidia-geforce-rtx-4060-ti by @Ingyro in #1014
  • feat: recognize native ternary (1.58-bit) models by @BlueVelvetSackOfGoldPotatoes in #886
  • bench: community results for apple-m4 by @PascalMasny in #1013
  • bench: community results for nvidia-geforce-rtx-2080 by @azumukupoe in #1029
  • bench: community results for apple-m4 by @mwave1239 in #1026
  • chore(deps): bump softprops/action-gh-release from 3.0.2 to 3.0.3 in the github-actions-minor-patch group by @dependabot[bot] in #1035
  • bench: community results for apple-m3-pro by @nikolalek in #1019
  • fix(bench): stop wall timer after body read by @AlexsJones in #1037
  • bench: community results for radeon-ai-pro-r9700 by @lucasschirm in #1033
  • docs: clarify Windows code-signing status by @dajiaohuang in #1020
  • fix(hardware): recognize A-series Apple Silicon unified memory by @lukatao in #1044
  • feat(storage): add disk planning for model libraries by @eyenpi in #1023
  • fix(bench): ignore implausible tok/s from degenerate Ollama timings by @kaanisthatyou in #1042
  • fix(scraper): pace HuggingFace requests on the ratelimit headers by @Akciali in #1047
  • fix(update): request every mapped field in the HuggingFace list expand[] by @SulimanAbdulrazzaq in #1048
  • fix(update): read the context window from config.json by @AlexsJones in #1054
  • feat(models): Sept 2026 model refresh (GLM-5.3, Qwen3.8-Flash-Next, DeepSeek-V4.1, Kimi-K3 and more) by @AlexsJones in #1055
  • fix(providers): map Gemma 4 instruct models to their Ollama tags by @kaanisthatyou in #1043
  • fix(bench): error when the requested model is not available by @kaanisthatyou in #1041
  • fix(providers): find HF models in HF_HUB_CACHE and XDG_CACHE_HOME by @SulimanAbdulrazzaq in #1051
  • test: add per-chip-class estimator accuracy gate by @texasich in #1052
  • bench: community results for apple-m5-pro by @dizzysound in #1053
  • bench: community results for nvidia-geforce-rtx-4090 by @imharshita in #1046
  • fix(providers): map gemma4:12b and other Sept 2026 catalog additions to Ollama tags by @AlexsJones in #1056
  • fix(scraper): revalidate retained catalog entries (fixes packed AWQ/GPTQ/INT4 parameter counts) by @AlexsJones in #1057
  • fix(fit): size and price MXFP4-native models at MXFP4 by @AlexsJones in #1059
  • fix(plan): grade GPU paths against free VRAM, not total capacity by @AlexsJones in #1058
  • fix(scraper): budget GGUF source probing so the weekly run can finish by @AlexsJones in #1061
  • fix(scraper): never size hybrid SSM models by the estimator; cap by declared size by @AlexsJones in #1063
  • fix(scraper): coerce list-valued expert counts; a failed estimate cannot abort a scrape by @AlexsJones in #1064
  • [Automated] Weekly model data refresh by @github-actions[bot] in #1062
  • chore(main): release 1.1.16 by @AlexsJones in #1032

New Contributors

Full Changelog: v1.1.15...v1.1.16

Don't miss a new llmfit release

NewReleases is sending notifications on new releases.