New and emerging models that could change the board. Everything here is tracked from the day it appears: unverified until independent numbers exist, early data once they do — and a formal challenger or title case when the evidence clears our bar.
Google DeepMind · could contend for best overall LLM. Google's new frontier model, announced on the DeepMind blog days after the Gemini 4 Argon launch. It has already taken the top spot on Artificial Analysis's headline index at 59, a point above Claude Opus 5.5. Whether that lead holds on coding, long-context and agentic work, and at what latency and price, is not yet clear.
OpenAI · could contend for best overall LLM. A GPT-6.1 variant aimed squarely at computer use, pitched as fast enough to put browser and desktop control in everyday hands. The striking part is positioning rather than scores: OpenAI frames it as general availability for computer use rather than a research preview. No independent numbers on task success, cost per completed task or safety under adversarial pages yet.
Liquid AI · could contend for best small / on-device LLM. The text model at the centre of Liquid's new LFM3 edge family, with GGUF builds published alongside the safetensors release and multilingual coverage spanning English, Arabic, Chinese and French. It arrives with sibling vision, audio and omni 2B variants, which is a broad edge line-up to ship at once. Quality against the established sub-3B open models, and real device throughput, remain unverified.
DeepSeek · could contend for best open-weight LLM. A multimodal MoE from DeepSeek shipped under MIT licence with FP8 weights in the box. Serving frameworks moved quickly: SGLang added a cookbook entry for it and vLLM pins an NVFP4 evaluation against it in CI, so the deployment path is real rather than theoretical. Independent quality comparisons against the current open-weight leaders are still missing.
Edge0 · could contend for best open-weight TTS. A compact 0.6B text-to-speech model published with safetensors weights and custom modelling code, advertising zero-shot voice cloning across multiple languages with a Chinese-language focus. The striking part is the size: voice cloning at well under a billion parameters, where most open cloning systems sit several times larger. Nothing independent yet on intelligibility, speaker similarity or how it holds up outside Chinese, and the publisher is unfamiliar, so treat the multilingual claim as unverified.