Skip to content

v0.2.19

Latest

Choose a tag to compare

@eugenehp eugenehp released this 26 Oct 00:57

The first release since 0.2.11. The 0.2.14 and 0.2.16 releases were prepared but never published, so their changes are included here. Every crate in the workspace is versioned 0.2.19 and built against upstream rlx 0.2.19.

Full notes: CHANGELOG.md § 0.2.19

Upstream rlx 0.2.19

  • Every rlx* pin is ^0.2.19, including rlx-program in the standalone program/ manifests.
  • All 32 pinned upstream crates resolve from crates.io (including rlx-torch-ckpt and rlx-opscope). A fresh clone builds without a sibling ../rlx checkout.
  • scripts/publish.sh tier fixes: rlx-sanotts moved to tier 2, and rlx-models-plugin now has its own final tier after the facade.

New model families (highlights)

  • LLM / MoE: Recursive Flint (40B-A2.5B, experts paged from disk), DeepSeek-V4.1-Flash, Qwen3.8-Flash-Next / Qwen4 arch (rlx-qwen4exp), GLM-5.3-Flash (rlx-glm5next), Kimi-K3 (training, inspection, distributed), Ling 3.0, Motif-3, Jamba, Llama 4, Llama-3.2-Vision, MiniCPM-V / MiniCPM-SALA, Nanbeige4.2, TinyLlama, Laguna, Ternary-Bonsai-2-27B, Perplexity pplx-decider
  • TTS / audio: Fish S2-Pro + Drama 3, KittenTTS 2, Pocket TTS, Kokoro, Parler-TTS, Soprano, MOSS-TTS-Nano, TADA, sanoTTS, ChatterBox, Orpheus, and many more. TTS crates are now rlx-native (ort-free).
  • ASR / speech: Moonshine v2 streaming, Conformer-CTC, Parakeet-TDT, VibeVoice-ASR, Qwen3-ASR, TEN-VAD (bit-identical frontend, embedded/MCU targets), wake-word suite
  • Vision / image: FLUX.1, FLUX 3 Klein, Open Genmoji, DINOv3, SigLIP 2, UNI2-h, BiRefNet, rlx-upscale (15 SR architectures), TRELLIS.2, Unlimited-OCR / Jina-OCR, PP-OCRv6
  • Embeddings: EmbeddingGemma-300M, jina-embeddings-v5-text-nano
  • Time series: Mistral Timestral, TimesFM 3

Platform

  • One exporter for every family: rlx_core::export / rlx-export (.rlxp packs)
  • Number formats: --number-format f16|bf16|f8|f6|f4|fNeXmY|f64|f64x2 on every LM CLI, NVFP4-style blocks, f64 double-single on GPU backends. QAT and fine-tuning via rlx-run qat.
  • Serving: rlx-openai multi-model OpenAI HTTP server and the rlx-serve library. Host-driven decode now covers 8 families (up from 3), and lm_registry lets any crate register a model with auto_runner.
  • Browser: rlx-models-web (wasm + WebGPU). Qwen3.5 runs split by layer across Web Workers.
  • Decoding: decoders share rlx-search. Also one single-bucket decode cache, and stop-signal fixes in five families.
  • Memory: streamed packed-weight uploads (qwen3/qwen35/gemma/lfm/dflash: RSS reduced by 4.7–19.6 GB) and a lazy token_embd.
  • Backends: many correctness fixes across CPU, Metal, MLX, CUDA, ROCm, wgpu and Vulkan. Experimental tpu and coreml device features.
  • Tooling: just lint-features (non-default feature states) and just release-preflight.

Known issues

  • CPU arena reuse miscompiles a real-scale DeepSeek-V4.1 attention graph (see CHANGELOG).