Editorial illustration for Inception Releases Mercury 3 Diffusion LLM at 1,068 Tokens per Second
Mercury 3 Hits 1,068 Tokens/Second on NVIDIA GPUs
Inception Releases Mercury 3 Diffusion LLM at 1,068 Tokens per Second
Inception put out Mercury 3 this week, the fourth version of its diffusion language model, and the headline number is 1,068 tokens per second on NVIDIA GPUs. That's the kind of throughput that matters to anyone building voice bots or agent pipelines where latency, not raw reasoning depth, decides whether a product feels usable. The model also ships with a 1M-token context window, a jump from the 260K window Mercury 2.5 introduced.
Inception hasn't disclosed parameter count, but says Mercury 3 is more than 40 times larger than the original Mercury 1 from 2025. It's API-only for now, served on Inception's own NVIDIA hardware, with no open weights or quantized versions available. On the company's self-reported benchmark table, measured at medium settings, Mercury 3 beats GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 tests, with its strongest showing on SWE-Bench Atlas and its weakest on DeepSWE v1, where Haiku 5.5 still comes out ahead.
The bigger story is architectural. Mercury 3 doesn't generate text the way GPT-6 or Haiku do.
Mercury 3 has more than 40x the total parameters of Mercury 1 (mini). Average quality across standard benchmarks rose about 3x between those 2 models. Speed stayed above 1,000 tokens per second.
Why this matters
For anyone building voice bots, search tools or latency-sensitive agents, Mercury 3's 1,068 tokens/second number is the headline, but we'd push back on taking Inception's benchmark table at face value. Beating GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 self-reported tests is a nice stat, except it's Inception grading its own homework at "medium settings," whatever that means precisely. The TinyFish result is more interesting: 50 live Mind2Web navigation tasks beating GPT-6.1-Sol, but without a published Sol number there's nothing to independently verify.
Same with Octen AI's 0.9-second median generation. These are real third-party runs, which counts for something, but partial numbers without baselines are closer to marketing than evidence. If you're evaluating this for production, the 1M-token context and raw speed are worth testing yourself rather than trusting the vendor's own leaderboard.
Diffusion LLMs generating in parallel instead of token-by-token is a genuinely different architecture worth watching, especially for agent workloads where latency compounds. Just run your own benchmarks before you believe the chart.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives MiniMax M3.1 Flash and Deepseek V4.1 split China's AI pricing battle
From Archives Replika founder Kuyda launches Wabi, multiplayer AI messenger app
Common Questions Answered
What is the token throughput speed of Mercury 3 on NVIDIA GPUs?
Mercury 3 achieves 1,068 tokens per second on NVIDIA GPUs, which is a significant throughput metric for latency-sensitive applications. This speed is particularly important for voice bots and agent pipelines where low latency determines whether a product feels usable to end users.
How much larger is Mercury 3's context window compared to Mercury 2.5?
Mercury 3 features a 1M-token context window, which represents a substantial increase from Mercury 2.5's 260K-token context window. This expansion allows the model to process and maintain information across much longer sequences of text.
What improvements has Mercury 3 achieved compared to Mercury 1 (mini)?
Mercury 3 has more than 40x the total parameters of Mercury 1 (mini), and average quality across standard benchmarks rose approximately 3x between these two models. Despite these significant improvements in capability and parameters, Mercury 3 maintained speeds above 1,000 tokens per second.
Why should developers be cautious about Mercury 3's benchmark comparisons against GPT-6 Luna and Claude Haiku 5.5?
Inception reported beating GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 self-reported tests, but these results were conducted at 'medium settings' with unclear specifications and graded by Inception itself. This self-evaluation approach raises questions about the objectivity and comparability of the benchmark results.
What use cases benefit most from Mercury 3's high token throughput?
Mercury 3's 1,068 tokens per second throughput is particularly valuable for voice bots, search tools, and latency-sensitive agents where response speed directly impacts user experience. These applications prioritize fast inference over raw reasoning depth, making Mercury 3's speed a critical advantage.
Further Reading
- Inception Launches Mercury 3, Setting a New Pareto Frontier in Speed and Intelligence - Business Wire
- Mercury 3: 1,068 Tokens a Second, and No Benchmarks - OrcaRouter
- Mercury 3 Diffusion LLM: 1,068 tok/s Claimed vs Measured - ExplainX
- Mercury: Ultra-Fast Language Models Based on Diffusion - arXiv
- Introducing Mercury 2.5 - Inception Labs
