What's new in 3.2.0 (2026-08-14)
These are the changes in inference v3.2.0.
New features
- feat(llm): surface logprobs in /v1/chat/completions by @SuperMarioYL in #5252
- feat: support deploying R3-embedding-0.6b and R3-rerank-0.6b models by @xiaoyesoso in #5272
- feat(rerank): support jina-reranker-v3.5 with its native listwise API by @xiaoyesoso in #5269
- feat: PDF input support for the OCR API by @OliverBryant in #5280
- feat(sglang): surface logprobs in /v1/completions and /v1/chat/completions by @SuperMarioYL in #5270
- feat: add DeepDoc (ragflow) OCR model as builtin image model by @kejhz653-stack in #5230
- feat(model):support minimax-m3 model by @llyycchhee in #5273
- feat: support per-replica worker and GPU placement by @m199369309 in #5286
- feat(webui): add per-replica placement controls by @m199369309 in #5287
- feat: support scaling running models by replica by @m199369309 in #5288
- feat(webui): add running-model replica scaling by @m199369309 in #5289
- feat: expose and display per-replica runtime resources by @m199369309 in #5291
- feat(ocr): expose DeepDoc's document-parsing pipeline as task="parse" by @OliverBryant in #5299
- feat(audio): add CAMPPlus speaker embedding support by @leslie2046 in #5298
- feat(image): add FireRed Image Edit 1.1 support by @Minamiyama in #5306
- feat: add Wan2.2 Animate 2 video models by @Minamiyama in #5309
- feat: auto-detect model download hub between Hugging Face and ModelScope by @OliverBryant in #5265
- feat(video): add MiniMax-H3 support by @qinxuye in #5321
- feat(llm): add Ling 3.0 model support by @Minamiyama in #5311
- Feat: enhance download progress details by @Minamiyama in #5310
- feat: [New Model] indextts 2.5 by @Minamiyama in #5319
- feat: [New Model] Add OvisOCR2 model support by @Minamiyama in #5322
Enhancements
- ENH: update models JSON [embedding, rerank] by @XprobeBot in #5276
- ENH: update models JSON [rerank] by @XprobeBot in #5279
- ENH: update models JSON [image, llm] by @XprobeBot in #5292
- ENH: update models JSON [image] by @XprobeBot in #5294
- ENH: add searchable dropdowns to Audit Center text filters by @leslie2046 in #5282
- ENH: add page navigation and deep pagination to log and audit centers by @leslie2046 in #5304
- ENH: update models JSON [image, llm, video] by @XprobeBot in #5316
- ENH: update models JSON [audio, llm] by @XprobeBot in #5324
- perf(supervisor): add whole-result debounce cache for list_models by @m199369309 in #5264
Bug fixes
- fix(model): avoid Set-changed-during-iteration crash and gc.get_objects() overhead under concurrent downloads by @SuperMarioYL in #5257
- fix: add anthropic SDK to core dependencies for route registration by @m199369309 in #5281
- FIX: use Xorbits/deepdoc mirror for the DeepDoc huggingface source by @OliverBryant in #5293
- fix(webui): ignore stale launch model requests by @m199369309 in #5285
- BUG: fix audit center text filters returning no results by @OliverBryant in #5277
- fix(worker): preserve Jina v3 allocator env for recovery by @m199369309 in #5300
- fix: preserve replicas across worker recovery by @m199369309 in #5290
- fix(ocr): stop task="parse" rejecting ordinary multi-page PDFs by @OliverBryant in #5308
- fix(vllm): keep WorkerActor importable without vllm for multi-GPU launch by @bluefish-08 in #5312
- fix(deploy): stop polling requests flooding the access log by @bluefish-08 in #5317
- fix(llm): declare audioread for qwen_omni_utils virtualenvs by @bluefish-08 in #5313
- fix: make cancelling a model launch actually cancel it by @bluefish-08 in #5315
- fix: normalize replica during autostart recovery by @medisean in #5318
Documentation
Others
- docs: add v3.1.0 to release notes index by @martinma51 in #5274
- docs: fix broken star history chart links by @SimonFoobar648 in #5323
- docs: update README new model highlights by @qinxuye in #5325
New Contributors
- @SimonFoobar648 made their first contribution in #5323
- @medisean made their first contribution in #5318
Full Changelog: v3.1.0...v3.2.0