alexsjones--llmfit/CHANGELOG.md

Changelog

1.1.2 (2026-07-10)

Features

• share: ship the registered OAuth App client id — interactive login enabled (a94ffcb)
• share: ship the registered OAuth App client id — interactive login enabled (325f8a3)

1.1.1 (2026-07-10)

Features

• bench: ship merged community benchmarks to every user in the next release (f2fd476)
• bench: ship merged community benchmarks to every user in the next release (5c46f92)

1.1.0 (2026-07-10)

Features

• bench --share — contribute benchmarks as PRs without the gh CLI (RFC #710) (5114694)
• bench: add --share to contribute benchmarks via PR without gh CLI (721eb66)
• bench: detect llama-server as a first-class provider (6dc367b)
• bench: local benchmark store — save runs, share the backlog later (0e22a75)
• bench: local store follow-ups — share backlog, dedup PRs, local-measurement overrides + calibration (390cbf1)
• fit: local benchmark runs override estimated tok/s (274c048)
• fit: purge test-stub catalog entries; calibrate estimates from local benchmarks (fb049c2)
• share: reuse open benchmark PR and make submissions idempotent (52f3afb)
• tui: expose installed/GGUF availability filter in the Filter modal (dd940d5)
• tui: filter by installed models in the Filter modal (75f4f1e)
• tui: offer to benchmark the selected model on entering leaderboard (61f6a47)

Bug Fixes

• data: verify GGUF source provenance; unlink wrong-model quants (63b2d3a)
• providers: stop marking whole model families as installed for llama.cpp (f839633)
• tests: pin GitHub-credential state in bench-offer toggle test (618ba16)
• tests: pin GitHub-credential state in bench-offer toggle test (fb5caff)

1.0.1 (2026-07-09)

Bug Fixes

• bump internal llmfit-core version requirement to 1.0.0 (7274a45)

1.0.0 (2026-07-09)

Features

• add global --tui flag to force interactive TUI (749b22f)
• add global --tui flag to force interactive TUI (#547) (f0146b1)
• add model language metadata (784af4d)
• add TTS model discovery (40a69f6)
• auto-detect system RAM bandwidth for MoE-offload estimates (c10f567)
• auto-detect system RAM bandwidth for MoE-offload estimates (f469543)
• discover TTS language metadata in scraper (74cd130)
• estimate basis + measured tok/s provenance (#292) (93f6806)
• expose estimate basis and local verification commands (#292) (af54caf)
• pin Huihui-Qwen3.6-35B-A3B-abliterated and add GGUF source (d048900)
• pin Huihui-Qwen3.6-35B-A3B-abliterated and add GGUF source (#656) (e0a5654)
• surface community-measured tok/s when hardware matches benchmark data (34eca7e)
• surface community-measured tok/s when hardware matches benchmark data (730cc5d)

Bug Fixes

• harden AMD sysfs GPU detection against silent card drops (#303, #638) (84a6cfa)
• harden AMD sysfs GPU detection against silent card drops (#303, #638) (4dc94ca)
• keep text generation discovery budget (5b55e2c)
• mark TTS models as unsupported runtime (5ca524e)
• ollama: don't mark cloud models as locally installed (closes #619) (580afee)
• preserve popularity metadata through scraper merge and discovery (b2b35d6)
• resolve #171 — Include ONNX models. (9939e67)
• support TTS catalog metadata (c98b3d5)
• use cross-platform which crate for binary detection in TUI (55d77ec)
• use cross-platform which crate for binary detection in TUI (d450daa)
• wire onnx catalog into model database (f2e7dba)

Miscellaneous Chores

• cut 1.0.0 (21ebc84)

0.9.38 (2026-07-04)

Bug Fixes

• docker: drop COPY of removed root data/ directory (a4445c7)
• docker: drop COPY of removed root data/ directory (5bb514d)

0.9.37 (2026-07-03)

Features

• calibration: weekly benchmark refresh + estimate-accuracy audit (#112/#119) (5548723)
• calibration: weekly benchmark refresh + estimate-accuracy audit against measured data (6c40dcf)
• claim: JSON output for programmatic consumers (llmfit-dra M0) (f469c22)
• claim: machine-readable JSON output via the global --json flag (1636e51)

Bug Fixes

• calibration: exclude speculative-decoding/MTP runs from the accuracy audit (1c058aa)

0.9.36 (2026-07-03)

Features

• add 'llmfit doctor' hardware diagnostic report for bug reports (53663bc)
• add llmfit doctor hardware diagnostic report (bcd7dc0)
• add issue templates for bug reports (bab1fe9)
• request 'llmfit doctor' output in the bug report template (b4069d9)
• scoring: benchmark-informed task alignment per use case (ce6d37e)
• scoring: benchmark-informed task alignment per use case (issue #150) (c994615)
• serve over a Unix domain socket (--unix-socket) (de62c2b)
• serve over a Unix domain socket (--unix-socket) (03d6ca5)
• show usable context per model (native→achievable) in TUI, CLI, and JSON (e33a3a3), closes #621
• usable-context display — native→achievable (closes #621) (1706929)
• user-local custom model overlay (closes #451) (decef82)
• user-local custom model overlay (custom_models.json) (e55c8de), closes #451

Bug Fixes

• guard optional attribute lookups in generated claim CEL (c6f481f)
• guard optional attribute lookups in generated claim CEL (a6f3c6a)
• hardware: correct Intel GPU detection — iGPUs get the unified RAM pool, dGPUs get Vulkan VRAM (4834767)
• hardware: detect generic-named MI50-class accelerators (issue #638) (180f484)
• hardware: don't drop large-VRAM AMD accelerators with generic ROCm names (bf16eb8)
• hardware: Intel iGPU unified-memory pool + discrete Arc VRAM via Vulkan (issue #609) (a120aa3)
• schema: add 'audio' to capability enum (a812232)

0.9.35 (2026-07-02)

Features

• add claim subcommand generating Kubernetes DRA ResourceClaims (08547ef)
• add claim subcommand generating Kubernetes DRA ResourceClaims (3369a49)
• expose GPU memory bandwidth in system JSON and add Strix Halo entries (3bfd334)

Bug Fixes

• reset list cursor to top when the search query changes (5e4dc7b)
• reset list cursor to top when the search query changes (43884f5), closes #657

0.9.34 (2026-06-28)

Bug Fixes

• pin Docker builder to bookworm to match runtime glibc (f13a633)
• pin Docker builder to bookworm to match runtime glibc (48caf17), closes #655

0.9.33 (2026-06-23)

Bug Fixes

• docker: bump builder image to rust 1.95 (sysinfo MSRV) (b63a142)
• docker: bump builder image to rust 1.95 (sysinfo MSRV) (b943f3a)

0.9.32 (2026-06-23)

Features

• add NCAI VAETKI models to curated model database (b3c04a8)
• analysis: add InstalledIndex and build_model_fits to consolidate provider detection (33dce8b)
• data: add Whisper/ASR model entries + audio support roadmap (27bae48)
• python: add riscv64 wheel build (e12bb85)
• python: add riscv64 wheel build (cb377f4)

Bug Fixes

• bundle WebView2 bootstrapper for Windows desktop app (c72103b)
• fit: CPU-only models that fit with headroom reach Good, not Marginal (aad4c37)
• fit: CPU-only models that fit with headroom reach Good, not Marginal (0ced1b7)
• hardware: detect all GPUs across sysfs, ROCm tabular, and fit pool (0989ca2)
• hardware: ROCm tabular multi-GPU parsing + summed fit VRAM pool (4e27617)
• scoring: MoE active-param quality, recency bonus, smooth fit curve (c499c00)
• scoring: MoE active-param quality, recency bonus, smooth fit curve (2db3e55)
• scoring: treat headroom as a perfect fit (one-sided fit curve) (235be33)
• tui: make provider filter input cursor-aware and ASCII-only (933644e)

Performance Improvements

• analysis: parallelize provider detection in detect_all with thread::scope (d0660bc)

0.9.31 (2026-06-09)

Features

• lmstudio: Add LMSTUDIO_API_KEY support for authenticated instances (0dc2712)
• lmstudio: Add LMSTUDIO_API_KEY support for authenticated instances (302cb53)
• speed up provider navigation when holding shift (gh-571) (f0d4783)
• speed up provider navigation when holding shift (gh-571) (7d74d7d)
• upgrade MiniMax default model to M3 (d45a546)
• upgrade MiniMax default model to M3 (7407d10)

Bug Fixes

• add Gemma 4 QAT models and fallback HF search (e40d3b8)
• add Gemma 4 QAT models and fallback HF search (9b4582f), closes #613
• lmstudio: empty string guard, README docs, and unit test for LMSTUDIO_API_KEY (46f70b7)
• replace unsafe env mutation in tests with pure logic tests (33047d8)
• tui: handle search input cursor and overflow (66f4c88)
• tui: handle search input cursor and overflow (8596005)
• tui: handle unicode download directory input (b34d164)
• tui: handle unicode download directory input (da120f8)
• tui: ignore modified key input in search (6c0b697)
• tui: ignore modified key input in search (8de0a1b)
• use GGUF metadata for param counts in GGUF-only repos (3f9b22b)
• use GGUF metadata for parameter counts in GGUF-only repos (ff28cdd), closes #622
• wrap env::set_var/remove_var in unsafe blocks for Rust 2024 edition (55b7979)

0.9.30 (2026-06-01)

Features

• tui: fuzzy search in provider filter popup (#600) (b055a04)

0.9.29 (2026-05-27)

Features

• open TUI with best-fit model preselected (#582) (834ee20)
• rescrape model catalog and expand Ollama mappings (#573) (c3e3f0d)

Bug Fixes

• align SortColumn::next with display order of columns in TUI (#576) (9e3477e)
• core: scan XDG cache path for HF models on macOS (#568) (#575) (9ec7dd3)
• release: unblock downstream publish jobs when sign-windows fails (18b581f), closes #581
• release: use always() to actually override transitive skip (0d6aeb3)

0.9.28 (2026-05-22)

Features

• tui: add ollama_name field to JSON output (#566) (36b3181)

Bug Fixes

• core: detect GGUF models in HuggingFace cache for llama.cpp (#568) (6093bcc)
• tui: include GGUF source providers in search and provider filter (#569) (8698b99)

0.9.27 (2026-05-20)

Bug Fixes

• tui: correct live-bench keybinding label from B to I (7aae532)

0.9.26 (2026-05-19)

Bug Fixes

• hardware: detect distinct AMD GPU models in multi-GPU ROCm systems (#559) (a100552), closes #550

0.9.25 (2026-05-17)

Features

• scoring: add generation-aware quality scoring (#553) (a3e0eb2)
• tui: move provider detection to background threads for faster startup (#554) (9e8308d)

Bug Fixes

• docker: skip HTTP probe on Linux when Docker Desktop is not running (#548) (ce735c9)

0.9.24 (2026-05-12)

Features

• add MCP server mode and NATS event publishing to llmfit serve (#545) (c4958f7)

Bug Fixes

• hardware: fix iGPU filtering on Ryzen 9 9950X where 2 GB iGPU passes the >= 2 GB discrete GPU threshold (#543) (c655bc1)
• MoE offload speed overestimate for Qwen3-Coder-Next (#544) (5bdc9df)

0.9.23 (2026-05-10)

Features

• populate architecture metadata for precise KV cache and MoE speed estimation (#539) (bab8c01)

Bug Fixes

• bench: implement 5 follow-up improvements to PR #278 (#523) (91fee77)

0.9.22 (2026-05-06)

Features

• overhaul model discovery with cursor pagination, trending, and additive DB (#537) (93e9d67)

0.9.21 (2026-05-05)

Bug Fixes

• add AutoRound format detection to prevent misclassification as GGUF (#532) (8214b1f)

0.9.20 (2026-05-04)

Bug Fixes

• lmstudio: resolve HF repo IDs to direct GGUF file URLs for download (#530) (115a01d)

0.9.19 (2026-05-03)

Features

• add llmfit bench for live inference benchmarking with quality metrics (#278) (8896fdb)

Bug Fixes

• lmstudio: wrap HuggingFace repo IDs in canonical URL for download (#520) (3399c70), closes #519
• ollama: treat OLLAMA_HOST=0.0.0.0 as a bind address and fall back to localhost (#521) (3a7bcf5), closes #516

0.9.18 (2026-04-30)

Features

• recommend: add --output-llamacpp flag (#508) (7c5925c)

Bug Fixes

• filter CPU Vulkan devices and suppress ROCm+Vulkan duplicates (fixes #271) (#328) (fe8d3f0)

0.9.17 (2026-04-29)

Features

• add benchmark cache with offline fallback (928390b)
• add RWKV v7 G1f models (1.5B, 2.9B, 7.2B, 13.3B) (#517) (fab2bec), closes #515

0.9.16 (2026-04-27)

Features

• add community benchmarks view from localmaxxing.com (4d0ffa7)
• add hardware picker to benchmarks view (H key) (21661b2)
• add vLLM inference provider (bfb0bb2), closes #501
• download: add --output-dir flag to override cache path (#499) (f70409f)

Bug Fixes

• add debug warning when util > 1.0 in GPU-mode VRAM pressure calculation (#513) (92d0f63), closes #496
• deduplicate AMD GPUs detected by both ROCm and Vulkan (fixes #139) (#314) (43dd9fc)
• match leaderboard API nested response schema (4282eb4)

0.9.15 (2026-04-25)

Features

• add DeepSeek V4 model family (#502) (0fd800f), closes #500
• make key 'g' cycle top/bottom (#454) (2ccb9bc)

Bug Fixes

• MoE min_vram regression — all experts must be in memory (#497) (b8f37fd)
• resolve a build warning for an unused variable (#331) (1e68d80)

0.9.14 (2026-04-23)

Features

• add Qwen3.6 models and refresh model database with discovery (5397a8a)

Bug Fixes

• MoE GPU-mode uses active params + UD-quant support (#475) (2768c0f)
• populate fit.installed in llmfit fit CLI command (#491) (2b67d2e)

0.9.13 (2026-04-22)

Bug Fixes

• installed flag always false for models without explicit mapping (#481) (#492) (7d4937f)

0.9.12 (2026-04-22)

Features

• add --tool-use filter flag to fit command (#465) (0ffd8cd)
• add Filter Popup (F key) with range filters, sort direction, and fit (#458) (d182cc3)

Bug Fixes

• ci: pass crates.io token via env, not argv (#487) (7be5edb)
• fall back to 127.0.0.1 when localhost Ollama detection fails (fixes #57) (#319) (126f81a)
• parse GPU name from rocm-smi output correctly (#479) (fca1984)
• providers: prevent argument injection in hf and docker subprocess calls (#329) (319c0fe)
• synchronize Cargo.lock and update release workflow to keep it in sync (#477) (9a99e69), closes #476
• tui: bind auto-spawned dashboard to loopback by default (#330) (6513a69)
• tui: fix filter popup Esc-revert, cursor alignment, and code cleanup (5591b90)
• tui: use theme background for popups and wrap GGUF source URLs (#407) (eb5c9bf), closes #344

0.9.11 (2026-04-19)

Bug Fixes

• MoE GPU-mode TPS uses full model size + expert-count-scaled overhead (#464) (e1f3245)
• read LM Studio download progress from streaming POST response (#467) (3b90e2c), closes #463
• replace unreachable!() with graceful fallback in estimate_tps_with_gpu (#459) (fd1b233)

0.9.10 (2026-04-18)

Features

• add --csv flag for spreadsheet-compatible model fit export (#456) (48227e6)
• tui: add Download Manager view (D) (c86dbb4)
• tui: show VRAM estimate at common context sizes in detail view (#320) (3e6f1ab)

Bug Fixes

• use Win32_PhysicalMemory to detect AMD APU VRAM on Windows (fixes #194) (#452) (1da9aac)

0.9.9 (2026-04-17)

Bug Fixes

• correct Llama 4 Maverick metadata (expert count, context length) (#446) (ac5b115)
• improve CPU-only TPS accuracy & add cross-platform Advanced Config modal (#455) (7e0918b)

0.9.8 (2026-04-13)

Bug Fixes

• ci: use RELEASE_TOKEN in release-please to trigger downstream workflows (f216779)
• prefer discrete GPU over integrated on hybrid systems (#303) (2707b54)

0.9.7 (2026-04-13)

Features

• add --force-runtime flag to override automatic runtime selection (c31099f)
• add --memory flag to override GPU VRAM autodetection (9a35a76)
• add --memory flag to override GPU VRAM autodetection (b6e323d)
• add "Sort by Release Date" to TUI (f371cd8)
• add "Sort by Release Date" to TUI (d55670a)
• add 15 popular models from HuggingFace (eb886d6)
• Add 15 popular models from HuggingFace (33→48 models) (e0c0b52)
• add 18 new models to database (Feb 2026) (8612c0e)
• add 18 new models to database (Feb 2026) (1c040f2)
• add 18 new models to database (Feb 2026) — conflict-resolved (0970845)
• add AWQ/GPTQ support with vLLM inference runtime (e398490)
• add AWQ/GPTQ support with vLLM inference runtime (d7f2f11)
• add Catppuccin color themes to TUI (f05d299)
• add Catppuccin color themes to TUI (104b281)
• add CLI/TUI model diff compare workflow (3cab068)
• add context-length cap for memory estimation (--max-context) (0d9a0ef)
• Add Docker containerization support (42ff7d3)
• Add Docker containerization support (fbb483c)
• add Docker Model Runner as a runtime provider (2a84138)
• add fuzzy search with space-separated terms (89bbe3d)
• add fuzzy search with space-separated terms (3d967f4), closes #65
• add GGUF download source enrichment for models (967253b)
• add Google Gemma 4 models and fix Gemma 3 capabilities (2cb21c3)
• add homebrew tap support and update release workflow (67b6fcf)
• add InferenceRuntime enum and MLX quantization support (186b95e)
• add license filter for models (5bb4b1e)
• add license filter for models (#186) (33bd708)
• add Liquid AI LFM2/LFM2.5 models (83d038a)
• add max-context cap for memory estimation (afba24a)
• add MiniMax-M2.7 to curated model database (a5cda9b)
• add MiniMax-M2.7 to curated model database (7c16b9d)
• add MlxProvider for MLX model detection and downloads (7f0ee24)
• add model comparison workflow (llmfit diff) and TUI compare mode (76a8a9c)
• add Ollama mappings for new models (5c97e75)
• add Qwen 3.5 series (397B/122B/35B/27B) to model database (0d550f3)
• add Qwen3-Coder-Next (80B MoE) and Qwen 3.5 Ollama mappings (5f8a640)
• add Qwen3.5 small model series (0.8B, 2B, 4B, 9B) (21cc8ac)
• add Qwen3.5 small model series (0.8B, 2B, 4B, 9B) (226c5c2)
• add runtime model database updates from HuggingFace (d465fb9)
• add runtime model database updates from HuggingFace (1e20ec3)
• add tensor parallelism awareness for multi-GPU model fitting (e8a6001)
• add tensor parallelism awareness for multi-GPU model fitting (5d99ec2)
• add theme switcher with 6 color schemes for TUI (317c44a)
• add theme switcher with 6 color schemes for TUI (89bd63f)
• add tok/s column sorting in TUI (41b7358)
• add tok/s sort option to TUI sort cycle and table header highlighting (82bb526)
• add use-case popup filter and use-case search matching to TUI (4b1aeab)
• add use-case popup filter to TUI (4b0f452)
• added arc support (0d5e991)
• added in vim like bindings (a3df62f)
• added LM studio (70e60cc)
• added logo (86dd737)
• added moe (4723782)
• added Rest API (2adb593)
• adding ollama as supported provider (e198c17)
• adding release please (c68914a)
• adding serve capabilities (503229e)
• api: add capabilities, license, supports_tp to model response and new endpoints (c02e189)
• append (WSL) to RAM label in tui when running under WSL (aa30f57)
• availability filter [a] in TUI (All / GGUF Avail / Installed) (9356150)
• bandwidth-based tok/s estimation for known GPUs (0c552b4)
• bandwidth-based tok/s estimation for known GPUs (1c34600)
• cargo fmt (cde77e3)
• caught some unavailable models on ollama (1d68bba)
• caught some unavailable models on ollama (6aaba9f)
• cli: add --sort for fit output (8114446)
• cli: add --sort option for fit output (3a06877)
• crate version yank skip rebuild (b1b0628)
• dashboard: ship embedded web UI and auto-start from CLI (96bd53c)
• dashboard: ship embedded web UI and auto-start it from CLI (37dd0db)
• detect installed Ollama models and support pulling from TUI (dc05d5a)
• expand mlx-community model mappings (7a9e917)
• expand mlx-community model mappings and improve fallback heuristic (5e8a676), closes #260
• fix for qwen3_5moe (455709f)
• fixed regression in LM studio model list (cbeccfa)
• fixed regression in LM studio model list (734f556)
• fixed up skill (38c8350)
• fixed up skill (368fa47)
• fixing vram on apple bug (bd59767)
• fixing vram on apple bug (db666b6)
• fixing vram on apple bug (c26d13b)
• fixing vram on apple bug (085cd25)
• GGUF download source enrichment (unsloth & bartowski) (6e3c828)
• hardware simulation mode and CLI hardware overrides (002eeff), closes #322
• hardware: detect NVIDIA Tegra/Grace Blackwell unified memory via ATS (rebased #93) (b454697)
• hardware: detect NVIDIA Tegra/Grace Blackwell unified memory via ATS addressing mode (dd1caa6)
• improvements based on #12 (7997525)
• increased model count (5be634f)
• increment version (d88bd3f)
• installed status for lm studio (a1cad30)
• mlx disablement on non-apple hw (c26bb49)
• MLX-native inference support for Apple Silicon (993599e)
• model detail modal + Ollama download in desktop app (c92053f)
• model detail modal + Ollama download in desktop app (dacae77)
• overall to the scoring system (e689f95)
• overall to the scoring system (d411a8d)
• overall to the scoring system (5ca2467)
• persist filter state across sessions (#430) (c3962e4)
• persist theme selection to ~/.config/llmfit/theme (765c87b)
• plan: KV cache fidelity + KV quant flag + TurboQuant gating (1f1b6fc)
• plumbing 2 (209f9d4)
• plumbing 2 (ef18920)
• pull functionality (e193d56)
• rebased (ac6639d)
• release plumbing (648a8c1)
• release plumbing (8215ba8)
• removed exact model count as it increases so often (1a19c42)
• reworked available models for download (6d56b03)
• show approximate disk space usage (#134) (#304) (32c835b)
• split-pane TUI detail view for GGUF downloads (de5c53b)
• support for windows vulkan (5a76a8f)
• supporting 94 models (c68d7aa)
• surface runtime info in TUI, CLI, and JSON output (4bdb4c4)
• This PR is to add the ability for capabilities to be described (d09b8fe)
• This PR is to add the ability for capabilities to be described (05768af)
• TUI GGUF downloads section, default enrichment, caching (e7dc0b2)
• tui: add runtime/backend filter and help popup (v0.9.2) (ad28138)
• update key bindings and add Catppuccin themes to README files (c41567b)
• updated build actions (59a8af4)
• updated demo (6c32985)
• updated demo (42fbeeb)
• updated images (20205bd)
• updated models (dfb9afb)
• updated sorting and new models (744465d)
• updated tui to support multiple providers better and also multiple GPU suppor (df338e8)
• updated urls (ce14772)
• updated version (4d1542b)
• updating models (d4f0a0e)
• version bump llama.cpp added (1643214)
• version bump llama.cpp added (319fb90)
• web: add 10 color themes from TUI theme.rs (ab14a06)
• web: add side-by-side model comparison view (ccb56e9)
• web: replace freeform limit input with preset options (96981ec)
• web: split App.jsx into Header, SystemPanel, FilterBar, ModelTable, DetailPanel (950e620)
• working on v0.8.1 (696107d)
• workspace restructure + Tauri desktop app (83ded27)
• workspace restructure + Tauri desktop app (3fb10d0)

Bug Fixes

• accept OLLAMA_HOST values without URL scheme (#166) (c647b13)
• add --local flag to install script and improve fallback logic (1e5cda1)
• add --local flag to install script and improve fallback logic (23b500e), closes #56
• add AMD RX 9060 series to VRAM estimation database (1c02da8)
• add AMD RX 9060 series to VRAM estimation database (9305810), closes #55
• add download confirmation to prevent accidental pulls (9e4e881)
• add download confirmation to prevent accidental pulls (#95) (5b56d74)
• address AlexsJones review comments (587252b)
• address Copilot review comments (74384c9)
• AMD Ryzen AI unified memory detection and --memory override (174fc10)
• AMD unified memory detection and --memory override (c763717), closes #89 #91
• bump Dockerfile Rust version to 1.88 for dependency compatibility (c25c02e)
• bump Dockerfile Rust version to 1.88 for dependency compatibility (6304ba0)
• cap default estimation context at 8192 tokens (#311) (4699ecb)
• ci: switch release-please to simple type for workspace version inheritance (48be41a)
• cli: default auto dashboard host to 0.0.0.0 (4553fbf)
• cli: make dashboard auto-start controllable and self-cleaning (0c51ad6)
• correct crates.io metadata and prepare for publishing (f8b08a0)
• correct crates.io metadata and prepare for publishing (2ce5858), closes #58
• correct PCI ID reading on Linux and improve AMD GPU identification (dcfde57)
• correctly estimate VRAM for APU integrated GPUs (2df3298)
• correctly estimate VRAM for APU integrated GPUs (Radeon Graphics) (5797f10), closes #25
• Default theme uses terminal colors for light/dark compat (72a6d8d), closes #67
• desktop: prevent XSS via inline onclick handler in modal (#323) (d20dbee)
• detect installed Ollama tag variants with suffixes (#165) (7092fa6)
• detect NVIDIA GB10/GB20 as unified memory GPUs (a77cf36)
• detect NVIDIA GB10/GB20 as unified memory GPUs (#83, #17) (c0a584a)
• discover GGUF files in HuggingFace subdirectories (#291) (4e8a596)
• docker action version (7a38a25)
• docker action version pin (da96503)
• download: fetch all shards of multi-part GGUF models (b7494d9)
• filter AWQ/GPTQ models by GPU compute capability (e720a0e), closes #257
• find Tauri bundle in correct target directory (19c0ac3)
• fmt: align providers formatting and enrich TUI compare metrics (a65e4c5)
• hide MLX models on non-Apple-Silicon hardware (#113) (1e3cf8a)
• hide MLX models on non-Apple-Silicon hardware (#113) (1052819)
• hide MLX-only models on non-Apple Silicon systems (ab54fa9)
• ignore GGUF tests that require network access (#306) (8fd8f4d)
• iGPU inflating GPU count and force-runtime being ignored (#271) (f584d7e)
• improve Android CPU and Vulkan GPU detection (0ebb92a)
• improve download error messages for incompatible model formats (0e2f9cc), closes #123 #198
• improve MoE tok/s estimate and clarify baseline speed labels (81f88c4)
• incorrect installed model count in TUI status bar (307c74f)
• invoke hf instead of huggingface-cli (752680f)
• invoke hf instead of huggingface-cli (78f8268)
• pr review fixes (25952da)
• prefer discrete GPU over integrated on Windows (#303) (3312a40)
• prefer exact matches in info selection (68272d4)
• prefer exact matches in info selection (883ea1f)
• query local providers in recommend CLI command (2477d1d)
• regression in json only mode (36e8503)
• remove unused function to eliminate build-time warning (4406582)
• remove unused function to eliminate build-time warning (289c8b8)
• Runtimes are installed but no downloadable artifact exist (9424204)
• support owner-scoped MLX pulls and robust tag normalization (7a9e476)
• support owner-scoped MLX repos and normalize MLX pull tags (f0e516c)
• support sudo in piped install script (9fa1960)
• support sudo in piped install script (cf61557), closes #158
• surface MoE offloaded RAM in JSON output (96b73ff)
• surface MoE offloaded RAM in JSON output (929eaad), closes #230
• tui: avoid borrow conflict when marking compare model (ae60aa5)
• tui: write dashboard pid file under ~/.llmfit, not /tmp (#324) (00967f1)
• typo in CHANGELOG.md (suppor -> support) (1e26ec2)
• typo in CHANGELOG.md (suppor -> support) (aa6cd00)
• update (2b6767d)
• update OpenClaw skill to match actual CLI output (df585a7)
• update OpenClaw skill to match actual CLI output (d53c192)
• use accurate model counts instead of len()/2 for installed display (7a273cc), closes #189
• use aggregated VRAM for multi-GPU fit scoring (dc3db39)
• use aggregated VRAM for multi-GPU fit scoring (f55882d), closes #68
• use MoE active parameters for tok/s estimation and clarify baseline speed labels (47e70a1)
• use per-card VRAM instead of summed for multi-GPU systems (2f1f9f4)
• use per-card VRAM instead of summed for multi-GPU systems (f80b4b0)
• use quantization in path selection for both dense and MoE models (#49) (8703d43)
• web: align dashboard hero with project tagline (d4e77c1)

Performance Improvements

• api: optimize embedded asset serving and cache policy (4d8c0cc)

0.9.6 (2026-04-13)

Features

• persist filter state across sessions (#430) (c3962e4)

0.9.5 (2026-04-11)

Features

• plan: KV cache fidelity + KV quant flag + TurboQuant gating (1f1b6fc)

0.9.4 (2026-04-11)

Bug Fixes

• download: fetch all shards of multi-part GGUF models (b7494d9)

0.9.3 (2026-04-09)

Features

• hardware simulation mode and CLI hardware overrides (002eeff), closes #322
• show approximate disk space usage (#134) (#304) (32c835b)

Bug Fixes

• ci: switch release-please to simple type for workspace version inheritance (48be41a)
• desktop: prevent XSS via inline onclick handler in modal (#323) (d20dbee)
• tui: write dashboard pid file under ~/.llmfit, not /tmp (#324) (00967f1)

0.3.7 (2026-02-21)

Features

• add --memory flag to override GPU VRAM autodetection (9a02f6e)
• add --memory flag to override GPU VRAM autodetection (39c5486)
• add 15 popular models from HuggingFace (128a020)
• Add 15 popular models from HuggingFace (33→48 models) (c45606b)
• add homebrew tap support and update release workflow (db09473)
• added arc support (b5892fc)
• added logo (c21d416)
• added moe (ac7ffe4)
• adding release please (f2bfc7f)
• append (WSL) to RAM label in tui when running under WSL (e0397cf)
• caught some unavailable models on ollama (b9f38da)
• caught some unavailable models on ollama (c0f7c20)
• detect installed Ollama models and support pulling from TUI (4159aaf)
• first pass (855ad3d)
• fixed up skill (fcb712a)
• fixed up skill (1f7d1de)
• fixing vram on apple bug (5e08754)
• fixing vram on apple bug (b3deca1)
• fixing vram on apple bug (92ddb0e)
• fixing vram on apple bug (42b2081)
• improvements based on #12 (5428ef8)
• increased model count (156b29d)
• increment version (283bebb)
• overall to the scoring system (f475938)
• overall to the scoring system (b0696cf)
• overall to the scoring system (37e2e10)
• plumbing 2 (1c615bb)
• plumbing 2 (dd6a3ec)
• pull functionality (923e7e7)
• release plumbing (7d21719)
• release plumbing (3accbb4)
• reworked available models for download (9adc84f)
• support for windows vulkan (cc0fd61)
• supporting 94 models (a652be3)
• updated build actions (1e65fdd)
• updated images (9141109)
• updated models (2d6c1d6)
• updated tui to support multiple providers better and also multiple GPU support (a3ca0bd)
• updated urls (f75ec27)
• updated version (2cfc73e)

Bug Fixes

• correctly estimate VRAM for APU integrated GPUs (72c8cb0)
• correctly estimate VRAM for APU integrated GPUs (Radeon Graphics) (8da5c2a), closes #25
• update OpenClaw skill to match actual CLI output (f38a0e5)
• update OpenClaw skill to match actual CLI output (e1adbfd)

0.3.6 (2026-02-21)

Features

• release plumbing (7d21719)
• release plumbing (3accbb4)

0.3.5 (2026-02-21)

Features

• add --memory flag to override GPU VRAM autodetection (9a02f6e)
• add --memory flag to override GPU VRAM autodetection (39c5486)
• add 15 popular models from HuggingFace (128a020)
• Add 15 popular models from HuggingFace (33→48 models) (c45606b)
• add homebrew tap support and update release workflow (db09473)
• added arc support (b5892fc)
• added logo (c21d416)
• added moe (ac7ffe4)
• adding release please (f2bfc7f)
• append (WSL) to RAM label in tui when running under WSL (e0397cf)
• caught some unavailable models on ollama (b9f38da)
• caught some unavailable models on ollama (c0f7c20)
• detect installed Ollama models and support pulling from TUI (4159aaf)
• first pass (855ad3d)
• fixed up skill (fcb712a)
• fixed up skill (1f7d1de)
• fixing vram on apple bug (5e08754)
• fixing vram on apple bug (b3deca1)
• fixing vram on apple bug (92ddb0e)
• fixing vram on apple bug (42b2081)
• improvements based on #12 (5428ef8)
• increased model count (156b29d)
• increment version (283bebb)
• overall to the scoring system (f475938)
• overall to the scoring system (b0696cf)
• overall to the scoring system (37e2e10)
• pull functionality (923e7e7)
• reworked available models for download (9adc84f)
• support for windows vulkan (cc0fd61)
• supporting 94 models (a652be3)
• updated build actions (1e65fdd)
• updated images (9141109)
• updated models (2d6c1d6)
• updated tui to support multiple providers better and also multiple GPU support (a3ca0bd)
• updated urls (f75ec27)
• updated version (2cfc73e)

Bug Fixes

• correctly estimate VRAM for APU integrated GPUs (72c8cb0)
• correctly estimate VRAM for APU integrated GPUs (Radeon Graphics) (8da5c2a), closes #25
• update OpenClaw skill to match actual CLI output (f38a0e5)
• update OpenClaw skill to match actual CLI output (e1adbfd)