Inception’s Mercury 3 gives latency-sensitive apps a diffusion LLM that runs at 1,068 tokens per second on NVIDIA GPUs, with a 1M-token context window and launch pricing of $0.10 per million input tokens and $0.35 per million output tokens. The model generates text through diffusion rather than one token at a time, and Inception says it is more than 40x larger than Mercury 1 while reaching comparable or better quality than GPT-6 Luna and Claude Haiku 5.5. In a 50-task web-navigation test, it completed 50% of tasks.
MiniMax’s M3.1 Flash gives developers a faster, cheaper model for long agent workflows, with usage costs 70% below its predecessor and pricing of $0.15 per 1M input tokens and $0.60 per 1M output tokens. It reaches 200+ tokens per second in everyday use, while scoring 73.7% on DeepSWE v1.1 and 61.7 on OSWorld V2. The company also identifies it as Space Bunny, an anonymous model that processed more than 200 trillion tokens.
Teams can run and customize Beam for coding and agentic work under Apache 2.0, with a 1M-token context window and support in Transformers, vLLM, and SGLang. It is a sparse mixture-of-experts model: although the network has 501B parameters, only 23B are active for each token, so each input uses a smaller slice of the system. A Rust-based harness, fine-tuning recipes, and reinforcement-learning support are included.
Developers can run a 0.5B-parameter Ming-3.0 checkpoint for multilingual speech synthesis, including zero-shot voice cloning and control over emotion, speed, pitch, volume, and delivery style. It takes text plus a short reference clip or speaker description and produces speech in Chinese, English, German, Spanish, French, Korean, or Japanese, with Sichuanese and Cantonese dialect control. Inference runs through the companion GitHub repository, not this checkpoint page alone.
Cloudflare’s Clairic gives applications typed values from unstructured text or images in a single pass, so recurring extraction and estimation tasks can avoid slower, more expensive reasoning calls. It can copy names and dates, compute values such as invoice totals, and forecast numbers, returning numbers with an 80% range. The hosted preview runs on Workers AI, with open weights planned.