Inception: Mercury 2

Inception: Mercury 2 breaks from the tradition of autoregressive language models by using a diffusion-based architecture that generates multiple tokens simultaneously through parallel refinement. This diffusion approach eliminates the sequential bottleneck that limits traditional LLMs, allowing the model to produce and converge tokens in a small number of steps rather than one at a time. The design philosophy centers on production-grade reasoning speed, targeting applications where latency compounds across multi-step workflows—particularly agentic loops, retrieval pipelines, and interactive coding environments where every millisecond affects user experience and operational cost.

As the world's first reasoning diffusion language model, Mercury 2 brings tunable reasoning levels to a parallel generation framework, combining the depth of chain-of-thought processes with the throughput of diffusion inference. Native tool use, schema-aligned JSON output, and OpenAI API compatibility make it straightforward to integrate into existing stacks. The model demonstrates its practical value in coding agent loops, real-time search and voice applications, and anywhere high concurrency demands consistently low response times under load.

Kilo Gatewayinception/mercury-2mercury

Open provider page Provider docs

Quick Info

Provider: Kilo Gateway
Model key: inception/mercury-2
Release date: Mar 4, 2026
Last updated: Mar 4, 2026
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.25
Output token cost: $0.75

Limits

Output tokens: 50,000 tokens
Context window: 128,000 tokens

Transparent token rates

Compare Inception: Mercury 2 pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about Inception: Mercury 2

Kilo Gateway

Coverage

Inception Labs has officially introduced Mercury 2, described by CEO Stefano Ermon as the world's fastest reasoning language model and the company's first production diffusion LLM (dLLM). Rather than decoding one token at a time, Mercury 2 generates responses through parallel refinement, converging over a small number According to Inception's launch post, Mercury 2 delivers 1,009 tokens/sec on NVIDIA Blackwell GPUs at $0.25 per million input tokens and $0.75 per million output tokens, with a 128K context window, tunable reasoning levels, native tool use, and schema-aligned JSON output. The post includes a quote from NVIDIA's Shruti

Kilo Gateway

CoverageBenchmark

ModelPicker's September 2026 review evaluates Inception's Mercury 2 at 4.08 out of 5.00, ranking it 118th overall, with pricing of $0.25 per million input tokens, $0.75 per million output tokens, and a 128K-token context window. The model achieves perfect scores in Structured Output, Strategic Analysis, Faithfulness, L The review flags critical weaknesses in safety calibration (1/5) and mediocre performance in tool calling (3/5) and constrained rewriting (3/0), indicating the model may lack robust built-in guardrails and may struggle with external API integrations. ModelPicker recommends Mercury 2 for low-cost structured data extract

Kilo Gateway

CoverageBenchmark

Artificial Analysis provides an independent benchmark assessment of Inception's Mercury 2, a proprietary model released in February 2026. It scores 15 on the Artificial Analysis Intelligence Index, placing it above average among comparable models (median: 12), and delivers measured throughput of 794.5 output tokens per The analysis notes that Mercury 2 is the reasoning version of the model and that a non-reasoning variant may also exist. Mercury 2 is characterized as above average in intelligence and well priced relative to other models at similar price points, making it a cost-efficient choice for developers seeking high-throughput