PlatformsThe briefing
AI-generated illustration
Inception launched Mercury 3, a diffusion language model it says runs above 1,000 tokens per second with a 1 million token context window.
Most chat models write one token at a time. Diffusion language models start with a rough draft of the whole answer and refine it in parallel. On October 19, Inception launched Mercury 3, the newest model in its diffusion line.
Inception says Mercury 3 runs at more than 1,000 tokens per second on widely available NVIDIA GPUs, has a 1 million token context window, and matches or beats speed-focused models such as GPT-6 Luna and Claude Haiku 5.5 on its benchmarks. Listed pricing is $0.20 per million input tokens and $0.70 per million output tokens, with a launch discount.
It is available through Inception's API, Baseten, OpenRouter and Vercel, and Inception says it is compatible with the OpenAI API format. The speed and quality figures are the company's own; independent testing will tell how they hold on real workloads.
Speed matters most where a person or another program is waiting, such as autocomplete, browser agents or support chat. Because the API format is OpenAI-compatible, a trial can be a configuration change. Keep your evaluation set fixed and compare answer quality as well as latency.
Faster and cheaper tokens lower the cost of each task only if quality holds. Ask for cost per completed task at the accuracy you need, including retries.
If you run a latency-sensitive feature, route a small share of test traffic to a faster model and compare completion rates and user wait times over a week.
Source material behind this briefing. Our practical implications are Ashlr’s interpretation.
This briefing was prepared with AI assistance from the linked sources. The analysis may include interpretation; consult the original sources for complete context. Cover artwork is illustrative.
AWS released general availability of training sessions in SageMaker AI serverless model customization. Builders can now write their own post-training loops, including custom loss functions and multi-turn RL, while SageMaker handles GPU allocation, distributed execution, and fault recovery.
PlatformsThe briefing
AI-generated illustration
Google and Hugging Face announced native TPU support for PyTorch. Compatibility, warm-up time and cost still need workload-specific checks.
PlatformsThe briefing
AI-generated illustration
MiniMax says M3.1 Flash can run agent workloads for less. Its posted token prices are low; the useful comparison is what you pay for a completed job, including retries.
