LLM Models

Meta's natively multimodal safety classifier, a dense 12B model pruned from Llama 4 Scout and trained jointly on text and multiple images. It screens both inputs and outputs of a language model, emitting a safe/unsafe verdict plus the violated categories from a 14-class hazard taxonomy (S1-S14) aligned to the MLCommons hazards standard. It runs on a single 24 GB GPU.

Tencent's open-weight reasoning and agent model, a 295B-parameter mixture-of-experts with 21B active parameters (plus a 3.8B MTP layer) and a 256K-token context window, released under Apache 2.0. It follows the Hy3 Preview launch in late April 2026 with scaled-up post-training. Secondary coverage often calls it 'Hunyuan 3.0', but Tencent's own model card uses only 'Hy3'.

OpenAI's streaming speech-to-speech translation model, taking spoken input in one language and producing spoken output in another without an intermediate text round trip, for live interpretation and multilingual conversation.

The smaller, cheaper variant of OpenAI's GPT-Realtime-2.1, keeping realtime speech with reasoning and tool use for voice workloads that do not need the full model's capability.

OpenAI's realtime speech model with reasoning and tool use, driving low-latency voice applications where the model listens and speaks directly rather than routing through separate transcription and text-to-speech stages.

OpenAI's low-latency speech-to-text model for realtime transcription, trading some of GPT Transcribe's accuracy for the responsiveness that live captioning, meeting notes and voice interfaces need.

OpenAI's high-accuracy speech-to-text model, handling both uploaded audio files and Realtime input transcription. It is the accuracy-oriented half of OpenAI's transcription pair, with GPT Live Transcribe covering latency-sensitive streaming instead.

Moonshot AI's open-weight frontier model and the first open release in the 3-trillion-parameter class, at 2.8T total parameters with 104B activated. Built on Kimi Delta Attention and Attention Residuals with a Stable LatentMoE framework activating 16 of 896 experts, it has native vision and a 1M-token context window. Quantization-aware training from the SFT stage onward yields MXFP4 weights, cutting storage to roughly 1.4 TB against the ~5.6 TB FP16 would require.

Jina AI's 597M-parameter multilingual document reranker, built on Qwen3-0.6B with a 'last but not late' interaction architecture that scores up to 64 documents in a single forward pass. It reaches 61.94 nDCG@10 on BEIR while being roughly an order of magnitude smaller than generative listwise rerankers, and supports a 131K-token context.

The latency-optimised variant of Cohere's Rerank 4 family, targeting low-latency and high-throughput reranking where the Pro tier's peak relevance is not required. It shares the generation's 32,000-token context window and multilingual coverage, and is billed per search query.

Cohere's quality-tier reranking model, the newest generation of its foundational ranking family. Rerank 4 Pro reorders an initial candidate set against a query for state-of-the-art relevance on complex queries, over English and non-English documents as well as semi-structured JSON, with a 32,000-token context window. Reranking is the second stage of a retrieval pipeline, applied after a vector or keyword search has produced candidates.

Google DeepMind's fastest and most cost-effective Gemini 3.5 model, released July 2026 for high-throughput execution. Flash-Lite runs at roughly 350 output tokens per second on Artificial Analysis while posting large gains over the previous Lite generation on agentic and long-context benchmarks.