Inception Labs Mercury 2.5 Release: Faster, Cheaper Diffusion‑Based LLM with 260K Context

Mercury 2.5 raises the bar for low‑latency, diffusion‑based language models

Takeaway: Inception Labs’ Mercury 2.5 offers a 40 % intelligence increase over Mercury 2, 1,107 tokens / second on commodity NVIDIA GPUs, 260 K token context, and launch pricing of $0.04 / M input tokens – a combination that makes it the fastest, cheapest production‑grade diffusion LLM currently available.

Core improvements over Mercury 2

  • Quality: Inception claims a 40 % jump in “intelligence,” putting Mercury 2.5 on par with cost‑optimized frontier models such as GPT‑5.6 Luna (Low), Gemini 3.5 Flash‑Lite, and Claude Haiku 4.5.
  • Speed: Benchmarks show 1,107 tokens / second on widely‑available NVIDIA GPUs, a figure repeatedly echoed by community users who describe the response as “ridiculously fast.”
  • Context window: The model now supports up to 260 K tokens, enabling very long‑form interactions and extensive retrieval‑augmented generation (RAG) pipelines.
  • Pricing: Standard launch price is $0.20 / M input and $0.75 / M output. An introductory discount of 80 % reduces cost to $0.04 / M input and $0.15 / M output.
  • New capabilities: Tunable reasoning, parallel tool calls, and schema‑aligned JSON output make the model suitable for complex agentic workflows.

“Mercury 2.5 is the most capable diffusion LLM on the market. To our knowledge, it is the largest diffusion language model ever trained.” – Inception Labs blog

Real‑world production use cases

Search agents and RAG pipelines

  • Mercury 2.5 can handle dozens of model calls per search request (query rewrite, reranking, fact structuring, summarisation, answer verification) while staying within a single user interaction.
  • Users report that the model’s latency keeps end‑to‑end search experiences fluid, even when multiple reasoning steps are required.

Voice agents

  • OpenCall’s AI phone agents achieve a median model response latency of ~170 ms on production traffic.
  • After switching to Mercury 2.5, OpenCall’s P99 latency dropped from minutes to 1 second, and P50 latency fell from 0.4 s to <0.2 s, outperforming competing providers.

    “The pause a caller hears is the latency; Mercury 2.5 made that pause negligible.” – Oliver Silverstein, OpenCall CEO

Coding sub‑agents and assistants

  • Augment Code uses Mercury 2.5 for context compaction and tool search. Compaction latency fell 82 % (150 s → 27 s) and cost dropped 90 % while preserving quality.
  • The model can generate a complete web app from a few prompts in under a second, demonstrating its suitability for rapid prototyping.

Community feedback from Hacker News

Commenter Insight / Critique Relevance
networked Encountered an IP‑protection guardrail that returned a refusal message, but the model recovered and completed the turn. Highlights a built‑in safety filter that may interfere with low‑level debugging.
Sphax Expected open‑weight release due to “widely available GPUs” claim; model remains closed‑source. Clarifies licensing expectations for developers.
gertlabs Tested Mercury 2.5 preview; finds it comparable to older open‑weight models for general‑purpose chat, but notes weaker tool‑use and agentic coding performance. Provides a balanced real‑world assessment of capabilities versus frontier models.
irthomasthomas Points out that high token throughput (1.1 k t/s) can mitigate the latency penalty of multi‑model judge systems in LLM consortia. Shows strategic advantage for ensemble architectures.
mrinterweb Argues that speed alone is insufficient; accuracy matters more than raw throughput. Reminds readers to consider quality‑vs‑speed trade‑offs.
schopra909 Discusses theoretical limits of diffusion for discrete token generation, suggesting future hybrid “looped” diffusion approaches. Places Mercury 2.5 within broader research trends on diffusion in LLMs.
piterrro Uses Mercury 2.5 as a binary reranker for vector‑store results; praises the speed and deterministic output format. Demonstrates a practical low‑latency retrieval use case.
trefoiled Describes the experience on OpenRouter as “strangely fast,” confirming the benchmark claims. Anecdotal validation of token‑per‑second metric.

Overall sentiment: users are impressed by speed and cost, while noting that reasoning accuracy and tool use still lag behind top‑tier models.

Mercury Voice and Mercury Router previews

  • Mercury Voice: A diffusion‑based LLM optimized for voice assistants, delivering time‑to‑first‑token <170 ms.
  • Mercury Router: A dLLM that classifies incoming prompts and routes them to the most appropriate model (open or closed) based on quality, speed, and cost considerations.

These previews extend the Mercury ecosystem beyond pure text generation, targeting real‑time voice interactions and multi‑model orchestration.

Getting started

  • API access: Inception API, Baseten, and OpenRouter all expose Mercury 2.5.
  • Free tier: 100 M tokens are available for trial.
  • Enterprise: Dedicated capacity, autoscaling, compliance controls, and configurable data retention are offered.
  • YC benefit: Y Combinator companies can claim $500 k in deployment credits.

Outlook

Inception Labs is already training a larger diffusion model slated for release in the coming months, promising further capability gains without sacrificing speed or token efficiency. The company invites contributors interested in advancing diffusion‑based LLMs to apply.

Bottom line

Mercury 2.5 demonstrates that diffusion‑based language models can achieve frontier‑level token throughput while remaining cost‑effective and low‑latency for production workloads such as search, voice, and coding assistants. The release has been positively received for its speed and pricing, though community feedback indicates that reasoning accuracy and tool‑use still need improvement to compete with the very latest proprietary models.

Sources

Related