Inception "Mercury 3", 1000 tok/s diffusion LLM

Diffusion at Scale: Introducing Mercury 3 â Inception

inceptionlabs.ai

October 19, 2026

4 min read

🔥🔥🔥🔥🔥

55/100

Summary

Inception Labs has launched Mercury 3, a diffusion-based large language model that it says delivers more than 1,000 tokens per second on Nvidia GPUs while matching or exceeding frontier speed-optimized models on benchmarks. The company reports a measured speed of 1,068 tokens per second, a one-million-token context window, and a model more than 40 times larger than Mercury 1. Inception says Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks and that its benchmark intelligence rose threefold from Mercury 1 mini. Mercury 3 is available through Inception’s API platform, OpenRouter, Baseten, and Vercel, with OpenAI API compatibility. Pricing is $0.20 per million input tokens and $0.70 per million output tokens, reduced at launch to $0.10 and $0.35 respectively; cached input tokens receive a 90% discount from the standard input price. Enterprise options include dedicated capacity, autoscaling, compliance controls, and configurable data retention. Early-access users deploy the model for browser agents, tutoring, search, and customer support. Inception says Mercury 3 completed 50% of 50 live Mind2Web navigation tasks in its web-agent harness, exceeding GPT-6.1-Sol under the same token volume and agent-step count. Octen AI reported median generation time of 0.9 seconds and median end-to-end response time of 2.2 seconds in Sense API tests.

Key Takeaways

Inception Labs says Mercury 3 is a diffusion LLM that generates 1,068 tokens per second on Nvidia GPUs and supports up to one million tokens of context.
Mercury 3 costs $0.20 per million input tokens and $0.70 per million output tokens at standard pricing, with a 50% launch discount.
Inception says Mercury 3 is more than 40 times larger than Mercury 1 and delivers three times the benchmark intelligence of Mercury 1 mini.
Mercury 3 is available through Inception’s API, OpenRouter, Baseten, and Vercel, and supports OpenAI API-compatible integrations.
Inception reports that Mercury 3 completed 50% of 50 live Mind2Web tasks in its production web-agent harness, outperforming GPT-6.1-Sol under equivalent token and step limits.

What the discussion said

Commenters treated Mercury 3 less as another benchmark release than as a test of whether diffusion language models have finally become a practical serving advantage. The headline numbers drew real excitement: a million-token context, unusually low token pricing, and claimed four-to-eightfold speedups could make high-throughput coding agents and reasoning workloads dramatically cheaper. Several readers see that kind of throughput as potentially disruptive because it squeezes far more useful model output from the same GPU fleet. Early hands-on feedback also suggested a noticeable step up from Mercury 2.5. The enthusiasm came with pointed demands for evidence. Readers want the parameter count, training details, weights, tool-calling and harness support, and results across reasoning settings rather than a single medium-effort benchmark configuration. Coding is the key practical test: Mercury trails Haiku on DeepSWE, and one latency comparison reported that DeepSeek Flash still wins both responsiveness and intelligence, with Mercury suffering slow time to first token. That makes the advertised generation rate feel incomplete for interactive use. The thread also questioned why larger labs, especially Google after its earlier diffusion demonstrations, have not turned diffusion research into a major product; explanations ranged from production batching economics to the possibility that the approach still has unresolved quality or deployment weaknesses. Lack of open weights and multimodality dampened what was otherwise a cautiously optimistic reception.

Where opinion split

The sharp dispute is whether raw diffusion throughput makes Mercury 3 a superior real-world model, or merely an impressive benchmark demo. Optimists argue that thousand-token-per-second generation and low prices unlock agentic coding and reasoning at a scale autoregressive models cannot match. Skeptics counter that slower first-token latency, weaker coding-agent results, unclear tool integration, and apparently stronger competing Flash models matter more than peak decoding speed.

Read original article

Coverage & discussions

Hacker News · inceptionlabs.ai

Inception "Mercury 3", 1000 tok/s diffusion LLM ↗

Community discussions

Hacker News discussion ↗

Reported comments: 21 · 15 saved in sample

Analysis ready

Community Sentiment

Positive

Positives

A million-token context paired with launch pricing around a few tenths of a dollar per million tokens could make long-context AI workloads far less punishing financially.
Generation near 1,000 tokens per second has readers imagining coding and decision agents that can brute-force far more work per GPU than conventional autoregressive systems.
The reported jump over Mercury 2.5 and competitiveness with fast frontier models makes diffusion LLMs look like a credible production path rather than a research curiosity.
A large free-token allotment and early reports of better quality than the prior release give developers a low-friction way to test whether the claims hold.

Concerns

Mercury's lower DeepSWE result than Haiku raises the uncomfortable possibility that its post-training has not yet translated raw speed into top-tier coding-agent reliability.
One comparison found DeepSeek Flash ahead on both latency and intelligence, while Mercury's slow first-token time undercuts the appeal of its huge steady-state throughput.
Benchmarking only a medium reasoning setting and omitting a quality-versus-effort curve leaves readers unable to judge where the model breaks or how fast it finishes whole tasks.
Closed weights, missing multimodality, and unanswered questions about model size, tools, and harness support limit its appeal to builders who want to inspect and integrate the model deeply.