LLM Models

Unlimited-OCR is Baidu's open-weight document-parsing vision-language model, presented on 22 June 2026 under the MIT licence. It targets long-horizon parsing: one call can cover a multi-page image set or an entire PDF through an infer_multi path, alongside single-image parsing in either a tiled 'gundam' configuration or an untiled 'base' configuration. Parsed output carries <|det|> markers giving each block's element category and bounding box. Baidu ships the weights on Hugging Face and ModelScope, an official vLLM recipe with prebuilt Docker images, an SGLang server path, and a hosted document-parsing endpoint in the Baidu Cloud OCR product.

Kyutai Pocket TTS is a 100-million-parameter text-to-speech model that the vendor says was released in January 2026 and runs in real time on a CPU using only two cores, reaching roughly six times real time on a MacBook Air M4. It streams audio with about 200 ms to the first chunk, clones a voice from any audio sample locally, ships a catalogue of preset voices, and handles arbitrarily long text inputs. It speaks English, French, German, Spanish, Portuguese and Italian, and can run client-side in the browser.

CSM (Conversational Speech Model) is a speech generation model from Sesame that produces RVQ audio codes from text and audio inputs. Its architecture pairs a Llama backbone with a smaller audio decoder that emits Mimi audio codes, and it accepts prior conversation turns as context so speech is generated in the flow of a dialogue. Sesame released the 1B variant on 13 March 2025 under Apache-2.0; it has been supported natively in Hugging Face Transformers since version 4.52.1. The released checkpoint is a base model, not tuned to any specific voice.

Voxtral TTS is Mistral's open-weights text-to-speech model for voice agents, released with BF16 weights and 20 reference voices. It covers nine languages (English, French, Spanish, German, Italian, Portuguese, Dutch, Arabic and Hindi), does zero-shot voice cloning with no transcript required for the voice prompt, and streams with roughly 90 ms time-to-first-audio. Output is 24 kHz audio in WAV, PCM, FLAC, MP3, AAC or Opus. The weights inherit the CC BY-NC 4.0 licence of the reference voices, so self-hosting is non-commercial.

Voxtral Mini Transcribe Realtime is Mistral's 4B-parameter streaming speech-to-text model, built from a roughly 3.4B language backbone and a roughly 970M causal audio encoder trained from scratch. It supports 13 languages and a configurable transcription delay, which the model card describes as multiples of 80 ms between 80 and 1200 ms plus a standalone 2400 ms setting; Mistral recommends 480 ms. Weights are released in BF16 under Apache-2.0, and the hosted version is served through the /v1/audio/transcriptions endpoint.

Nanonets-OCR2-3B is an open-weights image-to-markdown OCR model fine-tuned from Qwen2.5-VL-3B-Instruct. It turns document images into structured markdown: LaTeX for equations, markdown and HTML for complex tables, mermaid for flow and organisational charts, and dedicated tags for embedded images, signatures, watermarks and page numbers. Checkboxes and radio buttons become Unicode symbols. It is trained on handwritten documents and on documents in English, Chinese, French, Spanish, Portuguese, German, Italian, Russian, Japanese, Korean and Arabic among others, and answers visual questions about a page, replying "Not mentioned" when the answer is absent.

The 9B member of Google's ShieldGemma series of safety content-moderation models, built on Gemma 2 and available in English with open weights alongside 2B and 27B sizes. It is a text-to-text, decoder-only classifier: the caller supplies a preamble establishing the model as a policy expert, the text to be classified wrapped in turn tokens, a written description of the safety guideline, and an epilogue, and the model answers with a string starting 'Yes' or 'No' for whether that policy was violated. Google publishes guideline wording for four harm types - dangerous content, harassment, hate speech and sexually explicit information - in both prompt-only and prompt-response phrasings. Although generative, it is designed to be run in scoring mode over the Yes/No token probability.

An open one-stop moderation model from the Allen Institute for AI that achieves three goals in a single pass: detecting harm in user prompts, detecting harm in LLM responses, and evaluating whether a response is a refusal. The refusal head makes it usable for measuring how often a model over-refuses, for example as the evaluation module for the XSTest benchmark. WildGuard is a 7B model fine-tuned from mistralai/Mistral-7B-v0.3 and trained on WildGuardTrain, and its risk taxonomy covers four high-level categories - Privacy, Misinformation, Harmful language and Malicious uses - split into thirteen subcategories. English only.

A generative safety-moderation model from the Qwen team, fine-tuned from Qwen3-4B, which frames safety classification as an instruction-following task. It labels prompts and responses across three severity tiers - Safe, Controversial and Unsafe - and names the violated categories from a taxonomy of Violent, Non-violent Illegal Acts, Sexual Content or Sexual Acts, Personally Identifiable Information, Suicide & Self-Harm, Unethical Acts, Politically Sensitive Topics, Copyright Violation, and Jailbreak (input only). For responses it additionally emits a Refusal label. The Qwen3Guard series was trained on 1.19 million prompts and responses labelled for safety and supports 119 languages and dialects.

NVIDIA's multimodal content-safety classifier: a 4-billion-parameter fine-tune of Google's Gemma-3-4B-it with a SigLIP vision encoder at 896x896, trained as a LoRA adapter merged back into the base. It moderates inputs to and responses from both LLMs and VLMs, taking a prompt, an optional image, an optional response and an optional user-defined safety policy, and returning User Safety and Response Safety labels plus the violated Safety Categories from the Aegis Content Safety Dataset V2 taxonomy. In custom-policy mode it can emit a short reasoning trace before classifying. It unifies NVIDIA's earlier Nemotron 3 Content Safety multimodal model with the custom-policy capabilities of Nemotron Content Safety Reasoning 4B.

Mistral AI's text moderation model, served through a classifier endpoint rather than as a chat model. It scores text against the policy categories Sexual, Hate and Discrimination, Violence and Threats, Dangerous, Criminal, Self-Harm, Health, Financial, Law, PII and Jailbreaking, returning per-category scores that the caller thresholds itself. The service exposes one endpoint for raw text and one for conversational content, supports batching, and backs Mistral's Custom Guardrails config (moderation_llm_v2). It replaces mistral-moderation-2411, which was deprecated on March 31, 2026.

Gemini Embedding 2 is Google's multimodal embedding model, which maps text, image, video, audio and PDF inputs into a single unified embedding space. It takes up to 8,192 input tokens and returns Matryoshka-style flexible output dimensions from 128 to 3072 (768, 1536 and 3072 recommended), covers over 100 languages for text, and expresses task types as instructions written into the prompt rather than as an API parameter.