Mercury 2 enters production on Baseten

// 2h agoMODEL RELEASE

Mercury 2 enters production on Baseten

Mercury 2 is a proprietary diffusion large language model (LLM) developed by Inception Labs that generates tokens in parallel to achieve inference speeds exceeding 1,000 tokens per second on NVIDIA GPUs. Designed for high-speed reasoning and agentic workflows, the model is now available in production on Baseten, with early adopters like Augment Code reporting a 90% reduction in inference costs.

// ANALYSIS

Diffusion-based text generation is a promising paradigm shift that bypasses the sequential bottlenecks of traditional autoregressive models, enabling real-time agentic reasoning at scale.

* Generating 1,000+ tokens per second on standard GPUs dramatically reduces the latency floor for complex multi-agent reasoning loops.

* A 90% cost reduction makes high-frequency model calls economically viable for enterprise coding and reasoning applications.

* Deployment on Baseten simplifies the production serving and scaling process for developer integrations.

// TAGS

llmdiffusion-modelreasoningartificial-intelligencebaseteninception-ai

DISCOVERED

2h ago

2026-06-11

PUBLISHED

2h ago

2026-06-11

RELEVANCE

8/ 10

AUTHOR

_inception_ai

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

NEWS11m ago

Pieter Levels ports Wolfenstein with Claude Fable 5

Continuing his trend of reviving classic games, developer Pieter Levels (@levelsio) utilized Anthropic's new AI model, Claude Fable 5, to port the 2001 classic "Return to Castle Wolfenstein" to the web. The AI seamlessly handled the porting process, enabling the game to be fully playable in a browser complete with multiplayer capabilities, following his recent successful ports of Quake 1 and Quake 2.

NEWS16m ago

Carmack proposes AI-optimized coding styles for weaker models

A recent thought shared by John Carmack proposes that large language models could be used to optimize coding styles and code structures specifically to make codebases more accessible and manageable for weaker models. By identifying stylistic quirks that uniquely impact transformer-based models, developers could potentially adopt new "AI-friendly" coding standards that allow smaller, more efficient models to successfully perform tasks within a codebase.

UPDATE16m ago

Perplexity integrates Deep Research into Computer harness

Arav Srinivas announced that Deep Research is now a built-in, native skill within Perplexity Computer. Users no longer need to explicitly switch into a standalone mode for Deep Research; it simply works seamlessly as part of the agent harness, marking a significant step forward for state-of-the-art AI automation.