What's new in 3.5.0 (2026-09-25)
These are the changes in inference v3.5.0.
New features
- feat(audio): add AuK and AuK-Flash TTS support by @Minamiyama in #5525
- feat(audio): add Irodori TTS v4.1 support by @Minamiyama in #5527
- feat(virtualenv): add package details by @Minamiyama in #5529
- feat(ui): sort package details by @Minamiyama in #5539
- feat(world): add LingBot-World-V2 support by @Minamiyama in #5536
- feat(audio): add YuE2 music generation by @Minamiyama in #5526
- feat(llm): add Spark-X2.5 model support by @Minamiyama in #5538
- feat(api): add opt-in model request logging by @m199369309 in #5545
- feat(core): propagate correlation metadata across actor calls by @m199369309 in #5547
- feat(audio): support FishAudio-S2-Pro with MLX by @qinxuye in #5551
- feat(audio): support Irodori TTS Small with MLX by @qinxuye in #5553
- feat(model):support MinerU2.5 model by @llyycchhee in #5550
- feat(audio): support Breeze TTS 2 with MLX by @qinxuye in #5555
- feat(audio): support MiniMax Music3 with MLX by @qinxuye in #5558
- feat: recommend launch configurations for LLM, embedding, rerank and audio by @qinxuye in #5563
- feat(logging): add model request source identity by @m199369309 in #5559
- feat(api): add protected model request log access by @m199369309 in #5560
- feat(monitoring): index model request logs separately by @m199369309 in #5561
- feat(webui): add model request log observability by @m199369309 in #5562
- feat(UI): enhance running model capability experience by @maoyuehui in #5569
- feat(image): add Qwen-Image-2.1 support by @Minamiyama in #5571
- feat(ui): add audit time range options by @leslie2046 in #5577
- feat: recommend model sizes using offline memory metadata by @qinxuye in #5576
- feat(diffusion): expose native concurrency for image and Wan video engines by @qinxuye in #5579
- FEAT: show per-replica model launch progress by @leslie2046 in #5581
- feat(image): add Ming-Image support by @Minamiyama in #5582
- feat(image): add TeleOCR OCR and GGUF support by @Minamiyama in #5583
Enhancements
- ENH: Report disk usage in cache management by @Minamiyama in #5528
- ENH(llm): extend Ling-3.0-flash model support by @Minamiyama in #5556
- perf: modify model_ability to suit frontend pages by @llyycchhee in #5552
- perf(pooling): use native vLLM batching for embedding and rerank by @qinxuye in #5574
- perf(api): freeze startup heap and combine request middleware by @qinxuye in #5575
Bug fixes
- fix(core): reset client-supplied is_builtin on custom model registration by @kah-ja in #5530
- fix(metrics): start cluster updater from API lifespan by @m199369309 in #5534
- fix(llm): materialize client-supplied media before vendor readers by @bluefish-08 in #5533
- fix(audio): handle silent inputs for FunASR and MegaTTS3 by @m199369309 in #5537
- fix: improve ModelScope and offline model loading by @amumu96 in #5541
- fix(llm): normalize streamed tool calls and usage by @amumu96 in #5544
- fix(worker): run supervisor reporting RPCs on actor loop by @m199369309 in #5554
- fix(vllm): abort requests on client disconnect by @amumu96 in #5549
- fix: avoid full-sequence Qwen3 reranker logits by @Solaris-star in #5557
- fix: correct embedding benchmark concurrency and throughput by @qinxuye in #5566
- fix(logging): keep uvicorn access records out of application logs by @m199369309 in #5565
- fix(logging): normalize model request IDs and event types by @m199369309 in #5564
- fix(cluster): improve node version and resource display by @m199369309 in #5570
- fix(flux): remove duplicate quantizations by @Minamiyama in #5573
- fix(ui): improve launch progress layout by @leslie2046 in #5578
- fix(docker): include RPC context in pypiserver build by @qinxuye in #5584
- fix(docker): parse marked wheel references in mirror selfcheck by @qinxuye in #5586
Documentation
- doc: update README model highlights for v3.4.0 by @qinxuye in #5531
- doc: add v3.4.0 to release notes index and locale catalogs by @martinma51 in #5535
- doc: organize user guide and complete Chinese translations by @qinxuye in #5567
Others
- refactor: split built-in model catalogs into per-model files by @qinxuye in #5568
- refactor(image): remove custom Diffusers step scheduler by @qinxuye in #5580
- chore(catalog): add model spec schemas by @Minamiyama in #5572
New Contributors
Full Changelog: v3.4.0...v3.5.0