![\"Hand-drawn](\"https:\/\/rngoewtqzlssydnkvcdn.supabase.co\/storage\/v1\/object\/public\/article-images\/ef7decb0194246f8ac740e30aa022574.webp\"\/)
<\/p>\n\n\n\n<\/p>\n\n\n\n
Most AI speed problems don’t call for a bigger model. They call for a faster one. Mercury 2.5, Inception’s newest diffusion language model, is built on that idea: instead of generating text one token at a time, it produces and refines many tokens at once. OpenRouter clocks it at roughly 1,107 tokens per second on standard GPUs (buildfastwithai.com<\/a>). For workloads built on many quick model calls rather than one hard question, that number is the whole story.<\/p>\n\n\n\n This review covers what Mercury 2.5 does well, where it falls short, and how to slot it into HotBot alongside other tiers. If you want to run the prompts below as you go, you can try Mercury 2.5 on HotBot<\/a> directly.<\/p>\n\n\n\n Mercury 2.5 is a diffusion large language model, or dLLM. Rather than writing tokens in sequence, it generates and refines multiple tokens in parallel, which works more like iterative editing than left-to-right composition (datacamp.com<\/a>). That design is why the model chases throughput instead of frontier-level reasoning.<\/p>\n\n\n\n Inception released Mercury 2.5 as its most capable dLLM, reporting a 10-point intelligence gain over Mercury 2 and placing it alongside cost-optimized frontier models like GPT-5.6 Luna, Gemini 3.5 Flash-Lite, and Claude Haiku 4.5 (businesswire.com<\/a>). The update pushed the context window to 260K tokens, up from 128K, and added tunable reasoning, native tool use, JSON mode, and parallel tool calls.<\/p>\n\n\n\n One caveat matters. Mercury 2.5 is a Preview release, so its APIs, pricing, and performance can still shift (buildfastwithai.com<\/a>). Treat it as a fast, cost-efficient workhorse, not a settled frontier system.<\/p>\n\n\n\n The model earns its keep when latency compounds or when an application fires off many calls. Inception and independent reviewers point to query rewriting, extraction, fast coding iteration, search ranking, support routing, and subagent tasks as its natural territory (buildfastwithai.com<\/a>).<\/p>\n\n\n\n The production numbers hold up. Augment Code moved context compaction to Mercury and cut latency by 82%, from about 150 seconds to 27 seconds, while dropping cost 90% and holding quality steady; tool-search summaries came back in under a second (inceptionlabs.ai<\/a>). That’s the profile to watch for: repeated, structured calls where speed and price outweigh a single hard reasoning pass.<\/p>\n\n\n\n Mercury 2.5 sits between tiny efficient models and premium frontier systems (buildfastwithai.com<\/a>). On LM Market Cap it posts an LMSYS Arena Elo of 1347 and falls under the Coding category, though the site notes it has no task benchmark data yet (lmmarketcap.com<\/a>). Independent benchmark coverage remains thin, so read any quality claims as directional.<\/p>\n\n\n\n For a single exceptionally hard reasoning task, a large frontier system is the better call (buildfastwithai.com<\/a>). Model routing usually wins: let Mercury take the fast, high-volume path and pass the genuinely difficult requests to a heavier model. In HotBot, you can switch models per conversation from the full model catalog<\/a>.<\/p>\n\n\n\n The confirmed specs, useful when you compare Mercury 2.5 against other options:<\/p>\n\n\n\nWhat Makes Mercury 2.5 Different<\/h2>\n\n\n\n
What Mercury 2.5 Is Best At<\/h2>\n\n\n\n
Where it lands on quality<\/h3>\n\n\n\n
When to route elsewhere<\/h3>\n\n\n\n
Mercury 2.5 Specifications at a Glance<\/h2>\n\n\n\n