Moonshot AI has officially launched Kimi K3.1 Pro, its first 3T-class model featuring a 1 million token context window and substantial improvements in coding, 3D creation, and financial modeling. The accompanying K3.1 Pro Highspeed variant delivers identical capabilities at approximately 2–3x the inference speed, aiming to provide more responsive interactions.
Inference Speed
推理速度
The rate at which a model generates output tokens, typically measured in tokens per second.
Curated Intelligence
Developer releases open-source tool Llmspeed to simulate local LLM inference speeds on specific hardware before purchase. Integrating community benchmarks and spec-based projections, it covers major platforms like NVIDIA, AMD, and Apple Silicon, offering intuitive guidance for local AI deployment.
OpenAI has officially launched the Ultrafast version of GPT-6.1 Sol, offering up to 8x inference speed at 6x the standard price across API, Codex, and ChatGPT Work. The feature is currently restricted to Pro 500, eligible enterprise, and education users, requiring manual admin activation.
OpenAI has officially rolled out the 'Ultrafast' variant of GPT-6.1 Sol across its API, Codex, and ChatGPT Work platforms, claiming near-Astra intelligence with up to 8x faster inference speeds compared to the standard version. The model is now accessible via OpenRouter, targeting developers who require immediate high-speed responses.
OpenAI's Codex lead launched a 28-day intensive update sprint, delivering a ~50% inference speed boost for GPT-6 Astra and Sol on day one. Despite the performance optimization, community reaction was largely sarcastic, viewing it as PR damage control following prior latency issues and commercialization controversies.
OpenAI has introduced a new API service tier called Ultrafast, boosting GPT-5.6 Sol inference speed by up to 14 times. Powered by Cerebras hardware, it delivers output rates of up to 750 tokens per second. Currently in preview, this feature targets real-time applications requiring ultra-low latency.
Google DeepMind has launched Nano Banana 2, a new image generation model that combines the advanced world knowledge of Nano Banana Pro with the lightning-fast speed of Gemini Flash. The model offers production-ready specs, subject consistency, and rapid editing capabilities across Google products like the Gemini app and Search. Additionally, Google is integrating C2PA Content Credentials into its SynthID technology to improve the verifiability of AI-generated content.
Hugging Face has launched the SmolLM series of lightweight large language models, emphasizing a balance between blazing-fast inference and remarkable power. The series aims to provide efficient solutions for resource-constrained environments; however, the current material only contains the title, lacking specific parameters, benchmark data, and architectural details.