Mercury 2.5
Mercury 2.5 by Inception is a diffusion LLM refining tokens in parallel at 1,100+ tokens/sec, with tunable reasoning, 260K context, and parallel tool calls.
Model Overview
Capabilities, design details, and architectural traits
Mercury 2.5 - Inception's largest diffusion language model
Mercury 2.5 is a diffusion large language model (dLLM) from Inception. Instead of generating text one token at a time like autoregressive LLMs, it starts from a rough draft and refines many tokens in parallel through denoising steps. Inception describes it as the most capable production dLLM on the market and, to its knowledge, the largest diffusion language model ever trained.
What sets Mercury 2.5 apart
| Trait | Detail |
|---|---|
| Diffusion architecture | Generates tokens in parallel via a coarse-to-fine denoising process, rather than decoding one token at a time |
| Throughput | 1,107 tokens per second |
| Reasoning without the wait | Tunable reasoning lets a full reasoning pass finish inside the latency budget of a natural conversation |
| Context window | 260K tokens, extended from 128K in Mercury 2 |
| Agentic features | Parallel tool calls, native tool use, and schema-aligned JSON mode |
| Intelligence tier | A 40% quality jump over Mercury 2, comparable to cost-optimized frontier models like GPT 5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5 |
| Production focus | Latency-sensitive workloads: search agents and RAG pipelines, voice agents, and coding subagents |
Built for workloads that call the model many times
Mercury 2.5 targets pipelines where a single user action triggers dozens of model calls. In search, it handles query rewriting, reranking, and summarization fast enough to stay inside one interaction. In voice, it keeps reasoning latency low enough for live phone agents. In coding, it powers supporting calls like context compaction, model routing, and tool search, where slow supporting models would otherwise compound into long waits.
The Mercury family is expanding
Alongside Mercury 2.5, Inception previewed Mercury Voice, a dLLM optimized for voice agents with time-to-first-token under 170 milliseconds, and Mercury Router, a dLLM that reads incoming prompts and routes them to the best mix of quality, speed, and cost across open and closed models.
Benchmark Performance
Independent evaluations · Artificial Analysis
Accuracy & Capability Details
Compare Models Side-by-Side
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | |||
|---|---|---|---|
| General Info | |||
| Provider | Inception | Anthropic | Anthropic |
| Release Date | September 8, 2026 | September 22, 2026 | September 28, 2026 |
| Knowledge Cutoff | - | - | Jun 2026 |
| Context & Limits | |||
| Context Window | 260K | 1M Best Context Window | 1M Best Context Window |
| Pricing (per 1M tokens) | |||
| Input Pricing | $0.25 Best Input Pricing | $4 | $2 |
| Output Pricing | $0.75 Best Output Pricing | $20 | $10 |
| Modalities | |||
| Inputs | text | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 12.3 | 57.6 Best Intelligence Index | 56.0 |
| Coding Index | - | - | - |
| Agentic Index | - | - | - |
Humanity's Last Exam
Extremely difficult logical reasoning and knowledge.
Long Context Reasoning
Logical reasoning over long context windows.
SciCode Benchmark
Scientific coding and mathematical modeling.
Independent evaluation data provided by Artificial Analysis. To view the latest benchmarks and full details, visit their official site.
Top AI Models
Leading alternatives by intelligence score
