FUNDING · MODELS · #2362
Inception releases Mercury 3 diffusion LLM, claims much higher throughput than GPT-6 Luna and Claude Haiku 5.5
AI startup Inception announced Mercury 3, a diffusion-based language model that refines full answers over multiple passes rather than token-by-token generation. Inception's internal benchmarks claim Mercury 3 generates about 1,068 tokens/sec on Nvidia GPUs, handles up to 1 million-token context windows, outperforms GPT-6 Luna and Claude Haiku 5.5 on eight of 12 tests, and is available via an OpenAI-compatible API on Inception's platform and partners (OpenRouter, Baseten, Vercel); independent evaluations were not reported.
KEY POINTS
- AI startup Inception announced Mercury 3, a diffusion-based language model that refines full answers over multiple passes rather than token-by-token generation.
- Inception's internal benchmarks claim Mercury 3 generates about 1,068 tokens/sec on Nvidia GPUs, handles up to 1 million-token context windows, outperforms GPT-6 Luna and Claude Haiku 5.5 on eight of 12 tests, and is available via an OpenAI-compatible API on Inception's platform and partners (OpenRouter, Baseten, Vercel); independent evaluations were not reported.
- If upheld in independent testing, Mercury 3's diffusion approach could cut latency for real-time apps and agents by generating many tokens in parallel rather than sequentially, shifting trade-offs between speed, cost, and quality in LLM deployment.
WHY IT MATTERS
If upheld in independent testing, Mercury 3's diffusion approach could cut latency for real-time apps and agents by generating many tokens in parallel rather than sequentially, shifting trade-offs between speed, cost, and quality in LLM deployment.