Inception Labs this week shook up the AI race with Mercury 2, a new “diffusion” language model the company bills as the world’s fastest reasoning LLM. In benchmark and customer tests, Mercury 2’s standout claim is raw throughput: roughly 1,000 tokens per second versus about 89 tokens/sec for Anthropic’s Claude Haiku 4.5 Reasoning and 71 tokens/sec for OpenAI’s GPT-5 Mini. That puts it squarely in the same high-speed bracket Google later associated with its own DiffusionGemma — welcome to what some are calling the diffusion era of large language models.
What diffusion models do differently
- Traditional chatbots generate text one token at a time, checking each step as they go. Diffusion models instead initialize a block of text with noisy placeholder tokens and refine that block in several parallel passes until a final answer emerges — a technique borrowed from image generators like Stable Diffusion.
- The result is much higher parallel throughput and a snappier “flow” for long sessions: instant autocompletes, faster iterations on code or plans, and subagents that can run many quick utility calls without dragging down the whole system.
Benchmarks and head-to-heads
- On AIME 2026 (based on real American Invitational Mathematics Examination problems, scored as percent solved), Mercury 2 scored 90%. Google’s DiffusionGemma scored 69.1% on the same test, while standard (non-diffusion) Gemma 4 scored 88.3%.
- On GPQA, a PhD-level science benchmark, the gap narrows: Mercury 2 at 77% vs. DiffusionGemma’s 73.2%. Google’s own guidance still recommends standard Gemma 4 for applications that need the absolute highest quality, noting DiffusionGemma trails it across the board.
Real-world performance and cost
- Mercury 2’s speed claims aren’t just lab numbers. Augment Code, an AI coding-agent company, swapped Mercury 2 in for Anthropic’s Claude Opus 4.7 on a context-compaction subagent and reported an 82% latency drop and a 90% cost reduction, while maintaining comparable output quality (per a joint case study).
Origins and funding
- Inception’s approach builds on diffusion research from founder Stefano Ermon, a Stanford professor who co-authored early score-based diffusion work used in image generation. The startup raised a $50 million round with backing from Nvidia’s venture arm and individual investors Andrew Ng and Andrej Karpathy. Mercury 2 is currently available via API/cloud — the model weights aren’t public.
Practical caveats and the new architecture
- Diffusion LLMs excel where latency and high-volume throughput matter (real-time editing, many small utility calls, voice interfaces, etc.), but they’re not necessarily the best fit for the absolute hardest frontier reasoning tasks, where larger autoregressive models may still hold an edge.
- Architecturally, the big shift is toward orchestras of specialized subagents (reasoners, summarizers, routers, checkers). Sequential token-by-token models make many utility calls slow and expensive; parallel diffusion models make those calls cheap enough to use liberally.
- The ecosystem is still catching up: local runtimes, agent frameworks, and other infrastructure need to mature to make diffusion models seamless everywhere.
Where this matters for crypto and web3
- Faster, cheaper LLMs lower the friction for latency-sensitive on-chain and off-chain services:
- real-time developer tools for smart contract coding and “vibe coding” that keep pace with edits;
- multi-agent support systems and bots for DAOs that require many quick sub-calls;
- low-latency voice or chat interfaces for wallets, dApps, or on-call node operators;
- cheaper inference costs for oracle preprocessing, monitoring, and alerting pipelines.
- At scale, higher throughput on commodity GPUs can translate into meaningful cost and energy savings for projects that run lots of AI calls.
Bottom line
Mercury 2 pushes diffusion LLMs into the “fast and good” quadrant, delivering dramatic latency and cost improvements for throughput-heavy tasks while keeping competitive quality. It won’t replace every model class, but for crypto builders and other developers focused on speed, responsiveness, and multi-agent systems, diffusion models like Mercury 2 open new practical possibilities — provided the surrounding tooling and runtimes catch up.
Read more AI-generated news on: undefined/news
