Inception Releases Mercury 3 Diffusion Model
Inception has launched Mercury 3, a diffusion language model reaching 1,068 tokens per second while outpacing competing models across major benchmark tests.
AI startup Inception has introduced Mercury 3, its fourth-generation diffusion language model following Mercury 1, Mercury 2, and Mercury 2.5. Unlike traditional autoregressive models that generate output sequentially, Mercury 3 refines complete responses in parallel to hit 1,068 tokens per second on Inception-hosted NVIDIA GPUs. Featuring a 1M-token context window and a parameter footprint over 40 times larger than Mercury 1, the closed-weights API system combines high-throughput generation with a 3x quality improvement over its early predecessor.
In self-reported benchmark evaluations at medium settings, Mercury 3 led on 8 of 12 tests against OpenAI's GPT-6 Luna (~130 tok/s, 1.05M context) and Anthropic's Claude Haiku 5.5 (~240 tok/s, 1M context). Mercury 3 scored 89.23 on GPQA Diamond (versus Luna's 86.36 and Haiku 5.5's 84.51), 76.84 on IFBench (versus 61.82 and 54.65), and 51.08 on SWE-Bench Atlas (versus 33.33 and 32.26). However, Haiku 5.5 outperformed Mercury 3 on DeepSWE v1 (51.62 against 46.31, with Luna at 37.76) and MRCR v2 (78.40 against 75.50, with Luna at 62.72). GPT-6 Luna also edged ahead on AA-Omniscience-Public (46.94 to 46.44) and AA-LCR v1.1 (78.30 to 77.67), signaling remaining limits in long-context retrieval and specialized coding tasks.
Early testing partners reported substantial operational performance gains. Web-navigation tool TinyFish completed 50% of 50 live Mind2Web tasks to beat GPT-6.1-Sol, while Octen AI achieved 0.9 seconds median generation and 2.2 seconds end-to-end response time. Partners RevisionDojo and Lorikeet similarly cited roughly one-second latencies for interactive tutoring and customer support agents. Developers can access the mercury-3 endpoint on the Inception API Platform, OpenRouter, Baseten, and Vercel using OpenAI-compatible calls with a reasoning_effort parameter. Inception is also deploying Mercury Voice and classification tool Mercury Decide. Standard list pricing is set at $0.20 per 1M input and $0.70 per 1M output tokens, discounted at launch to $0.10 and $0.35, alongside a 90% discount on cached inputs and 100M free tokens for new accounts.
This is our own summary of reporting by MarkTechPost
