Mercury 2.5 (retires Nov 15)

ai-tldr.dev provides a third-party technical summary of Mercury 2.5, listing it as released on September 9, 2026 with a 260K-token context window, proprietary weights (API only), and a diffusion language model architecture that generates and refines tokens in parallel. Inception reports 1,107 tokens per second on widel The model supports tool calling and structured outputs and is served through the Inception API as well as Baseten and OpenRouter, with enterprise deployments offering dedicated capacity. Pricing is listed at $0.20 per million input tokens and $0.75 per million output, with an 80% promotional rate of $0.04 input and $0.

An independent editorial piece covering Inception's Mercury 2.5 launch on September 8, 2026, distributed via Business Wire and picked up by outlets including the Las Vegas Sun. The article confirms Inception's claim of pushing past 1,100 tokens per second in live deployments and contextualizes the release within a dens The piece adds detail on two companion preview products that shipped with Mercury 2.5: Mercury Voice, targeting a voice interface layered on the same diffusion backbone, and Mercury Router, which classifies requests and routes them to the appropriate Mercury model. Inception's own materials describe Mercury 2.5 as the

BenchLM.ai's Mercury 2.5 model card records the release as September 8, 2026, describing it as Inception's diffusion reasoning model with a 260K context window and launch rates of $0.04 per million input tokens and $0.15 per million output tokens. Inception reports 79% on GPQA Diamond and 77% on IFBench, which the page The page's value is precisely its transparency: it separates verified from provisional and unmeasured rows, rather than presenting benchmark numbers as third-party validation. Capability scores and speed are absent or unranked, so this entry should be cited as evidence of the launch date, context window, pricing, and t

On September 8, 2026, Inception announced the launch of Mercury 2.5, positioning it as "the most capable dLLM and the fastest reasoning LLM in production," running over 1,100 tokens per second. The Redwood City, California-based company, founded by Stanford, UCLA, and Cornell researchers behind the first commercial dif The press release details that unlike autoregressive models that generate tokens one at a time, Inception's dLLM starts with a rough draft and refines tokens in parallel, decoupling cost and latency from reasoning depth. CEO and co-founder Stefano Ermon said: "Nobody in this industry thinks LLMs can get smarter, faster

Inception’s Mercury 2.5 uses diffusion LLM architecture to generate and refine multiple tokens in parallel rather than producing them sequentially. The model page claims 1,107 tokens per second on standard GPUs, a 260K-token context window, tunable reasoning levels, parallel tool calls, and schema-aligned JSON output for production workflows such as search agents, voice pipelines, and coding subagents. Mercury 2.5 was released on September 8, 2026, and the listing says the exact variant is going away on November 8, 2026, making its retirement imminent. It claims a 10-plus point intelligence increase over Mercury 2, although the page also reports measured throughput of 236 tokens per second and 0.78-second latency, so the headline generation-speed claim is not fully reconciled.

Benchable.ai's model detail page lists Inception Mercury 2.5, a text-in/text-out reasoning model released September 8, 2026 with a 260K context window and a diffusive architecture that generates and refines tokens in parallel rather than sequentially. The page attributes the model to Inception and records supported parameters including temperature, tool use, response format, tool choice, reasoning, stop sequences, structured outputs, and max tokens, with structured outputs, reasoning, tools, and response format listed as supported features. The same benchmark summary reports Mercury 2.5 scoring 98.0% on both Coding and Mathematics at the 99th percentile, 96.0% on Reasoning, 99.0% on Ethics, 98.0% on Email Classification, and perfect accuracy on Hallucinations and General Knowledge baselines, alongside a 100% success rate across all benchmarks. It also flags a notable weakness of 0.0% on Instruction Following (Baseline), suggesting the model needs work on precisely following directives.

Inception Labs has released Mercury 2.5, described by CEO Stefano Ermon as its most capable production model and the largest diffusion language model ever trained, with a 40% intelligence increase over Mercury 2 while keeping the same low-latency, low-cost serving profile. The company says thousands of developers and dozens of enterprises have adopted the earlier Mercury 2, whose usage grew over an order of magnitude across search, voice, and coding workloads. Mercury 2.5 generates and refines tokens in parallel rather than sequentially, hitting 1,107 tokens per second on widely available NVIDIA GPUs with a 260K-token context window. List pricing is $0.20 per million input tokens and $0.75 per million output tokens, with an 80% launch discount bringing those to $0.04 and $0.15. Capabilities include tunable reasoning, parallel tool calls, and schema-aligned JSON output, with production use in search agents, RAG pipelines, and voice agents.

A model release aggregator lists Inception: Mercury 2.5 as an API launch on 2026-09-08 with a 260k token context window, providing an independent corroboration of the launch date and context length. The entry appears alongside other recent API launches such as DeepSeek V4.1 Flash, Gemini 3.8 Flash, and Qwen3.8 Max, sit The aggregator offers no primary technical detail beyond the timestamp and context window, serving mainly as evidence that Mercury 2.5 reached general availability through OpenRouter on schedule. Inception is named as the originating provider on the Inception API, not the Kilo gateway. The thin, timestamp-only nature o