Inception Labs Releases Mercury 3: Diffusion LLM Exceeds 1,000 Tokens per Second, but Benchmarks Are Questioned

Inception Labs released Mercury 3 on October 19, 2026, claiming that the diffusion-architecture LLM reaches an inference speed of 1,068 tokens per second and outperforms GPT-6 Luna and Claude Haiku 5.5.

Factual Reconstruction

According to the announcement on the Inception Labs website, Mercury 3 is the largest diffusion-architecture LLM to date, with 40 times the parameters of Mercury 1 and context length expanded from 260K tokens in Mercury 2.5 to 1M tokens. Official benchmarks show that it outperforms GPT-6 Luna and Claude Haiku 5.5 on multiple tests, at a speed of 1,068 tokens per second. Early partners have already used the model in areas including voice chat support, real-time tutoring, data engineering, and web navigation.

Third-party sources note that independent third-party test data is lacking, and that the default use of a “low reasoning effort” setting makes the benchmarks non-comparable. In measured testing, its DeepSWE coding score was 46%, lower than Claude Haiku’s 52%. On pricing, the standard rate is $0.20 per million input tokens and $0.70 per million output tokens, with a 50% discount during the initial launch period.

Mechanism Breakdown

Mercury 3 generates text through a diffusion architecture. Unlike autoregressive models, it generates multiple tokens in parallel in a single forward pass, which explains its high-speed characteristics. The official website emphasizes that the scaling laws of diffusion LLMs are similar to those of autoregressive models, and that after a 40x increase in parameter scale, intelligence improves 3x. Production cases show that the model completes 50% of Mind2Web navigation tasks in browser agent tasks, maintaining the same token count and number of steps as frontier models at lower cost.

In voice and search applications, Mercury Voice is optimized for low latency; Sense API tests show a median generation time of 0.9 seconds. In educational scenarios, the model supports near-real-time feedback, shortening wait times while maintaining quality.

Industry Impact

This release pushes diffusion architecture into the spotlight of direct competition with autoregressive models. Early adopters such as TinyFish AI, RevisionDojo, and Octen AI reported actual deployments in agents, tutoring, and search, showing that speed advantages can translate into more efficient workflows. The API is compatible with the OpenAI format for easier integration, while also offering enterprise-grade dedicated capacity and compliance controls.

The pricing strategy and 1M-token context support may affect the market landscape for real-time applications, especially in latency-sensitive areas.

Strategic Assessment

(This section is analysis, not fact.) If the diffusion architecture can consistently demonstrate scalability, it may change the competitive landscape for LLMs, but the current benchmark controversy requires more independent testing to confirm; the long-term comparison with autoregressive models depends on actual production performance, not a single speed metric.