The latest and greatest in open-source AI models of all kinds.
35B/3B active MoE from Alibaba Qwen. First Qwen with hybrid linear attention (Gated DeltaNet + MoE). Agentic coding rivals Qwen3.5-27B: SWE-bench 73.4%, Terminal-Bench 51.5%. Multimodal, 262K context (1M ext.), Thinking Preservation. vLLM, SGLang, KTransformers. Apache 2.0.
Lightricks ships distilled v1.1 for LTX-2.3 (8 steps, CFG=1). Across 3k+ A/B-tested video generations, users report: no more opening blur, sharper detail, better prompt adherence, improved character consistency, better camera transitions, and significantly improved audio. LoRA variant (2.74 GB) works with the full base model.
Q1_0 CUDA kernels merged in b8806, follow-up to the CPU-only backend (#21273). Bonsai 8B (1.07 GiB): 374 tok/s tg128; 4B (540 MB): 485 tok/s; 1.7B (231 MB): 626 tok/s — all on RTX 5090. Requires Turing MMA+. Works on some AMD GPUs. KLD vs FP16: 0.0005, 98.7% same top token.
First open-source SOTA 3D world model from Tencent Hunyuan. WorldMirror 2.0 (1.2B): video/images → point cloud, depth, normals, camera params, 3DGS in one pass. SOTA over Gen3C, SEVA, Lyra. Full text/image → navigable 3D world pipeline coming. Tencent HY-World Community License.
17B MoE DiT from Nucleus AI (2B active). First fully open-source MoE diffusion model at this quality. GenEval 0.87, DPG-Bench 88.79 (#1), OneIG-Bench 0.522 (beats Imagen4). Base model, no post-training. Full weights, code, dataset. diffusers. Apache 2.0.
8B DiT from Baidu. Top-1 open-weight T2I on GENEval (0.8856). Strongest at dense text rendering, structured layouts (posters, comics), and complex prompt following. Turbo: 8 steps (DMD+RL). 24GB VRAM. GGUF Q4 ~5GB (unsloth), ComfyUI (Comfy-Org), diffusers. Day-zero LoRA (ostris).
DDTree builds a draft tree from DFlash's block diffusion distributions, then verifies the whole tree in one target-model forward pass with tree attention. Lossless. Qwen3-30B-MoE HumanEval T=0: 8.22x over AR (+2.13x over DFlash). Uses existing DFlash drafters, no retraining. MIT.
Qwen3-Omni-30B-A3B (30B/3B active MoE) lands in llama.cpp: image + audio input, text output. Qwen3-ASR 1.7B (speech-to-text only) also supported. GGUFs from ggml-org: Q4_K_M 18.6 GB, Q8_0 32.5 GB. Apache 2.0. llama-server -hf ggml-org/Qwen3-Omni-30B-A3B-Instruct-GGUF
Lossless DFlash speculative decoding for Apple Silicon via stock MLX, no fork. Block diffusion drafts 16 tokens; target verifies in one pass. 89% acceptance. Qwen3.5-4B: 54→197 tok/s (3.7x). Qwen3.5-9B: 31→127 tok/s (4.1x). pip install dflash-mlx. MIT.
Gemma 4 audio conformer encoder merged into llama.cpp. 12-layer USM-style Conformer with 30s chunking; E2B 14/14 quants pass short audio, E4B 19/21. CUDA, Metal, CPU, Vulkan. BF16 mmproj required -- lower quants cause repetitions via ClippableLinear sensitivity. Unsloth E2B/E4B GGUFs ship with BF16 mmproj.