Glonce

Inception Labs launches Mercury 2.5 diffusion LLM, 40% smarter than Mercury 2

modelsThe Glonce editorsHacker News🔥68Glonce rating68Impact55Novelty50Relevance95Surprise40Players45Credibility78
Inception Labs launches Mercury 2.5 diffusion LLM, 40% smarter than Mercury 2

Inception Labs released Mercury 2.5, a diffusion-based large language model the company describes as its most capable production model to date. It follows Mercury 2, whose launch the company says brought thousands of developers and dozens of enterprises into production, with usage growing more than tenfold since; Inception says it used customer feedback and production failure cases from that deployment, not just benchmarks, to sharpen evals and guide training for Mercury 2.5.

Inception states Mercury 2.5 delivers a 40% increase in intelligence over Mercury 2 while keeping the same low-latency, low-cost serving profile, putting quality in the range of cost-optimized frontier models such as GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5. The company calls it the most capable diffusion LLM on the market and, to its knowledge, the largest diffusion language model ever trained, though it gives no benchmark methodology or parameter count to back either claim. On the technical side, Mercury 2.5 runs at 1,107 tokens per second on widely available NVIDIA GPUs and supports a 260K-token context window, tunable reasoning, parallel tool calls, and schema-aligned JSON output. Standard pricing is $0.20 per million input tokens and $0.75 per million output tokens; at launch, that is discounted 80% to $0.04 per million input and $0.15 per million output, with no stated end date for the discount.

Inception frames three production use cases. In search and RAG pipelines, where a single search request can trigger dozens of model calls to plan, rewrite, rerank, summarize, and check results, the company says several leading search-infrastructure companies now run Mercury in production to keep those calls fast. In voice agents, OpenCall, which builds AI phone agents for live customer calls, reports that switching to Mercury brought median model response latency close to 170 milliseconds, cut P99 response time from several minutes to about one second, and cut P50 response time from 0.4 seconds to under 0.2 seconds. In coding tools, Augment Code says moving context compaction, model routing, and MCP tool search to Mercury cut compaction latency 82%, from roughly 150 seconds to 27 seconds, reduced cost 90% while maintaining quality, and returns tool-search summaries in under a second.

Alongside Mercury 2.5, Inception previewed two related products: Mercury Voice, a diffusion model tuned for voice agents with time-to-first-token under 170 milliseconds, and Mercury Router, which reads an incoming prompt and routes it to whichever open or closed model offers the best mix of quality, speed, and cost. Mercury 2.5 is available through Inception's own API, through Baseten, and through OpenRouter, with enterprise options covering dedicated capacity, autoscaling, compliance controls, and configurable data retention. Inception is offering 100 million free API tokens to try Mercury 2.5 and $500,000 in deployment benefits to Y Combinator companies. The company also says it has already started training its next model, described as its largest yet, targeting a release in the coming months.

Key facts

  • Inception Labs says Mercury 2.5 is 40% more capable than Mercury 2 while matching cost-optimized frontier models like GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5, running at 1,107 tokens per second with a 260K-token context window.
  • Standard pricing is $0.20 per million input tokens and $0.75 per million output tokens; a launch discount cuts that 80% to $0.04 per million input and $0.15 per million output, with no stated end date.
  • OpenCall's voice-agent workload saw median model response latency near 170 milliseconds after switching to Mercury, with P99 latency dropping from several minutes to about one second and P50 from 0.4 seconds to under 0.2 seconds.
  • Augment Code moved context compaction to Mercury, cutting latency 82% (from roughly 150 seconds to 27 seconds) and cost 90% while maintaining quality.
  • Inception also previewed Mercury Voice (sub-170-millisecond time-to-first-token) and Mercury Router, and says it has already started training its next, larger model.

Why it matters

Diffusion-based language models generate output differently from the autoregressive transformers behind most frontier chatbots, which is part of why Inception Labs can claim such high throughput and low latency. Mercury 2.5 pushes that architecture further: the company calls it the most capable diffusion LLM on the market and, to its knowledge, the largest diffusion language model ever trained. The release is evidence that diffusion architectures are maturing from a research curiosity into a production option for workloads where latency and cost matter more than peak capability.

Who it affects

The stated audience is companies running latency-sensitive AI workloads: search and RAG providers that chain many model calls per query, voice-agent builders like OpenCall, and coding-assistant makers like Augment Code that split tasks across small, frequent model calls. Inception is also courting new developers directly, offering 100 million free API tokens to try Mercury 2.5 and $500,000 in deployment benefits to Y Combinator companies.

How to use it

Mercury 2.5 is available through Inception's own API, through Baseten, and through OpenRouter, with enterprise deployments supporting dedicated capacity, autoscaling, compliance controls, and configurable data retention. Standard pricing is $0.20 per million input tokens and $0.75 per million output tokens; at launch it is discounted 80% to $0.04 per million input and $0.15 per million output. The model supports tunable reasoning, parallel tool calls, and schema-aligned JSON output. Inception also previewed Mercury Voice, built for voice agents with time-to-first-token under 170 milliseconds, and Mercury Router, which reads an incoming prompt and routes it to whichever open or closed model best balances quality, speed, and cost.

How solid is it

The headline quality claims come from Inception Labs itself, not an independent benchmark. The company states a 40% increase in intelligence over Mercury 2 without naming the benchmark suite or methodology behind the figure, and gives no parameter count or other technical detail to support calling Mercury 2.5 the largest diffusion language model ever trained. The production figures are more concrete and attributed to named people: OpenCall co-founder and CEO Oliver Silverstein gives specific before-and-after latency numbers for his company's voice workload, and Augment Code gives specific latency and cost figures for context compaction; NVIDIA's Shruti Koparkar, Senior Manager of Product in its Accelerated Computing Group, is quoted endorsing the architecture's production readiness. These remain vendor-supplied case studies rather than third-party benchmarks.

Risks and caveats

The 40% quality gain and the claim of being the largest diffusion model ever trained are unverified assertions from Inception Labs, with no published methodology, benchmark suite, or parameter count behind either one. The 80% launch discount has no stated end date, so the effective price could change without notice. Mercury Voice and Mercury Router are previews, not general releases, and Inception has given no date for its next, larger model beyond "the coming months."

“After we switched to Mercury, our P99 response time dropped from several minutes to just one second, and our P50 dropped from 0.4 seconds to under 0.2.”

— Oliver Silverstein, co-founder and CEO, OpenCall