New open models, tested and explained, with code you can copy and run.
A 3B-decoder streaming ASR model with 240–560 ms delay. With the authors' adapted vLLM build, a rolling KV cache lets it transcribe audio of any length at constant memory.
Edge0/Audio8-ASR-Infinite
Tencent's MIT-licensed AuK does zero-shot TTS, speech editing, enhancement, and separation from natural-language instructions.
tencent/AuK
Breeze TTS 2 is an open-weight English and Chinese TTS model with voice cloning, text-described voices and a streaming API. Weights are non-commercial.
BreezeBlue/Breeze-TTS-2
A contrastive scorer on frozen Qwen3-8B embeddings. It ranks the candidates you give it and answers typed questions about a state.
Contrastive-LM/CLM-v0.1-8B
NetEase Youdao's streaming ASR model built on Qwen3-ASR: transcript text never gets rewritten, with 80 ms to 2 s chunks. Setup, code and caveats.
netease-youdao/Confucius4-R2T2
DeepSeek's 552B MoE activates 8B params on prefill and stores 890 bytes of KV cache per token. What the card says and how to start running it.
deepseek-ai/DeepSeek-V4.1-Flash
Edge0-35B-A3B streams experts from SSD so a 4-bit Qwen3.6-35B-A3B derivative decodes at about 15 tok/s in under 3 GiB of active memory.
Edge0/Edge0-35B-A3B-preview
Google DeepMind's largest dense Gemma 4 model handles images, video frames and 256K-token context. Here's how to run it with Transformers.
google/gemma-4-31B-it
A LoRA adapter plus a small decision head on Gemma-4-26B-A4B that gives a calibrated probability for every option in about 45 ms, and can think when it is unsure.
autotrust/GEV-26B-Decide
A 340M classifier from Fastino that scores intent, urgency, sentiment and routing labels you choose at call time, in one forward pass, on CPU or GPU.
fastino/GLiNER2.5-Decide
Z.ai's first natively multimodal GLM-5 model: 320B total / 18B active parameters, hybrid sparse-linear attention, MIT license. What the card covers.
zai-org/GLM-5.3-Flash
A 27B vision decision model that returns a calibrated probability for every option in one forward pass. Here's how to serve it and use it.
autotrust/JEV-27B-VL
Jina-OCR-v1 is a 570M-active-parameter OCR model that turns page images into Markdown. Here is how to run it with Transformers, vLLM, or the hosted API.
jinaai/jina-ocr-v1
IFM's open MoE model runs 4B active parameters with a 512K context window. How to serve it, call it, and use it for agents and long documents.
IFM/K2-Horizon-MoVA-36B-A4B
Kimi K3 is a 2.8T-parameter open-weight multimodal MoE with a 1M-token context. What the model card says and how to call it.
moonshotai/Kimi-K3
Kokoro-82M is an 82M-parameter open-weight TTS model. How to install it, generate speech, and use it for narration, pronunciation fixes, and captions.
hexgrad/Kokoro-82M
Laya answers typed questions about text with probabilities instead of generated text. It covers 100+ languages and is Apache 2.0. Fit temperatures before trusting the probabilities.
convaiinnovations/laya
Xiaomi's 9B SFT model, fine-tuned from Qwen3.5-9B, scores higher than its base on SWE Pro, Terminal Bench and Toolathlon. How to serve it with SGLang.
XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B
Xiaomi's 1.02T-parameter MoE model has 1M context and omnimodal input, and was trained for agents. Here is how to serve it with vLLM or SGLang.
XiaomiMiMo/MiMo-V2.6-Pro-RL
MiniCPM5-2B is a 2.5B Llama-architecture model with 131K context and tool calling. Here's how to serve it and what it's good for.
openbmb/MiniCPM5-2B
MiniMax H3 is an open-weight model that generates video and stereo audio together. Here's how to serve it with SGLang and where the hosted APIs fit in.
MiniMaxAI/MiniMax-H3
NVIDIA's 100M-parameter diarizer labels up to 8 speakers, streaming with 0.32 s of buffer or offline with 30.4 s. Here's how to run it.
nvidia/Nemotron-3-Diarization
NeoHorse-1-4B is a text-only Qwen3.5-4B fine-tune for agent harnesses, tool use and coding. Here is how to serve it with SGLang or vLLM.
TokenRhythm/NeoHorse-1-4B
Nex-AGI's smallest Nex-N2.5 agent model, served with SGLang on 2 H100s, with tool calling and per-request thinking modes.
nex-agi/Nex-N2.5-mini
Qwen's driving model adds a BEV perception head and a flow-matching planner to Qwen3.5-4B. Install it and sample ego trajectories from demo scenes.
Qwen/Qwen-Drive-1.0-4B
Qwen-Image-2.1 combines text-to-image, image editing and native RGBA output in one diffusers pipeline. Here is how to run it.
Qwen/Qwen-Image-2.1
Qwen3.8-27B is an Apache-2.0 dense 27B vision-language model with adjustable reasoning. How to serve it and call it today.
Qwen/Qwen3.8-27B
Qwen's 125B MoE with 6B active params handles text, images and video. What it's good at, how to call it, and where it falls short.
Qwen/Qwen3.8-Flash-Next
TeleOCR is a 1.2B Apache-2.0 model that turns text, tables, formulas and layouts into structured output, including from photos of warped pages.
XingChen-AGI/TeleOCR
China Telecom's 29B MoE activates 4B parameters per token and has a 256K context. Here's how to call it through an OpenAI-compatible API.
XingChen-AGI/Xing4.0-29B-A4B
YuE2-3B turns lyrics and a style prompt into a full 48 kHz stereo song. You can edit the melody and chords as an ABC score and render it again.
m-a-p/YuE2-3B
Clef answers typed questions about any input (text, JSON, images) with a probability per option in one forward pass. Here's how to use it for routing, triage and classification.
Cloudflare/clef
Google's new 740M multimodal embedding model runs on a CPU. Here's how to build a working semantic search over your own documents, and the prefixes that make it accurate.
google/embeddinggemma-2
Loading EmbeddingGemma 2 in float16 silently breaks it. Here's how to detect it and pick the right dtype for your hardware.
google/embeddinggemma-2
Kolibri 1 is a German–English reasoning model with only 3.5B active parameters. How to serve it, control its thinking effort, and wire up tool calling.
Aleph-Alpha/Kolibri-1
A practical rule of thumb for truncating embeddings to cut vector-DB cost, with the numbers from EmbeddingGemma 2.