1.1.19 (2026-11-03)
Features
- cluster mode (preview): find models that fit a Ray or multi-node GPU cluster — see docs/cluster.md (#1224, #1225, #1228; takes over #279 by @michaeljabbour)
- distinguish exact and legacy PCIe benchmark matches (#1211) (52a6f49)
Bug Fixes
- exclude NVFP4 on ROCm GPUs (#1203) (e3d5fe5)
- hardware: avoid collecting unrelated process information (#1200) (a560b35)
- match benchmark calibration by PCIe link width (#1202) (ad2859b)
- models: keep the most-downloaded repo as a deduped model's identity (#1205) (242407f)
- models:
supports_tpnow allows GQA KV-head replication like vLLM (TP larger than the KV-head count), which changes the TUI tensor-parallel filter and API results for all users (f01092f) - models: stop flagging non-chat models as tool-capable (#1207) (012eba4)
- providers: prevent false MLX installed matches (#1206) (6a4f070)
- scraper: read the HF migration createdAt as no release date (#1194) (1b3c707), closes #176
- scraper: use the GGUF header count when the config estimate is lower (#1166) (7c88807)
- tui: align simulation popup cursor with the wider VRAM/GPU label (855c75c)
- web: restore jsdom localStorage in tests on Node 25+ (#1222) (1068476)
What's Changed
- bench: community results for amd-radeon-rx-9060-xt by @MacherG7 in #1187
- [Automated] Weekly benchmark cache refresh by @github-actions[bot] in #1158
- fix: exclude NVFP4 on ROCm GPUs by @luojiyin1987 in #1203
- fix(scraper): read the HF migration createdAt as no release date by @Akciali in #1194
- fix(scraper): use the GGUF header count when the config estimate is lower by @jayhemnani9910 in #1166
- chore(deps): bump the github-actions-minor-patch group across 1 directory with 4 updates by @dependabot[bot] in #1157
- fix(hardware): avoid collecting unrelated process information by @sidalit in #1200
- bench: community results for amd-radeon-rx-5700-xt by @Sadeeed in #1201
- fix: match benchmark calibration by PCIe link width by @luojiyin1987 in #1202
- fix(models): keep the most-downloaded repo as a deduped model's identity by @AlexsJones in #1205
- [Automated] Weekly model data refresh by @github-actions[bot] in #1159
- fix(providers): prevent false MLX installed matches by @AlexsJones in #1206
- fix(models): stop flagging non-chat models as tool-capable by @AlexsJones in #1207
- ci(release): hold unsigned Windows assets and make signing explicit by @AlexsJones in #1208
- feat: distinguish exact and legacy PCIe benchmark matches by @luojiyin1987 in #1211
- chore(deps): bump the cargo-minor-patch group with 4 updates by @dependabot[bot] in #1215
- chore(deps): bump jsonschema from 0.58.6 to 0.59.0 by @dependabot[bot] in #1216
- fix(web): restore jsdom localStorage in tests on Node 25+ by @AlexsJones in #1222
- feat: cluster mode with Ray discovery (takes over #279) by @AlexsJones in #1224
- fix(cluster): structured refusal reason, tp_size, KV replication, TP=1 and Ray discovery follow-ups by @AlexsJones in #1225
- fix(cluster): consistent quant, no phantom KV on refused splits, no speed for refused layouts by @AlexsJones in #1228
- docs: mark cluster mode as a preview by @AlexsJones in #1229
- chore(main): release 1.1.19 by @AlexsJones in #1204
New Contributors
- @MacherG7 made their first contribution in #1187
- @sidalit made their first contribution in #1200
- @Sadeeed made their first contribution in #1201
Full Changelog: v1.1.18...v1.1.19