模型與推論
Mercury 2.5 Continues Diffusion-Based Text Generation, but Third-Party Gateway Tests Show Throughput Below the Official Peak
Inception has released Mercury 2.5 with a 260,000-token context window, adjustable reasoning, and parallel tool calling, continuing to replace purely autoregressive decoding with iterative refinement of multiple tokens. The company claims 1,107 tokens per second, but Vercel AI Gateway reports a real-world median of about 616 tokens per second, so deployers should still retest it with their own prompts and concurrency conditions.
On September 8, Inception launched [Mercury 2.5](https://www.inceptionlabs.ai/blog/introducing-mercury-2-5), continuing the Mercury series’ diffusion language model approach. Traditional autoregressive LLMs generate tokens strictly from left to right; Mercury instead uses multiple denoising rounds to refine candidate text in parallel, aiming to parallelize part of the decoding process. This design is particularly suited to search agents, voice services, and subagent workflows, where a single user request may involve multiple short inference steps—including query rewriting, routing, summarization, and structured output—and the latency of each step accumulates across the full chain.
The new version offers a 260,000-token context window, adjustable reasoning effort, parallel tool calling, and JSON output conforming to a specified schema. Inception claims that the model can reach 1,107 tokens per second on NVIDIA GPUs and delivers a 40% increase in overall “intelligence” compared with Mercury 2. The company also says that after Augment Code adopted it for context compression, latency fell from about 150 seconds to 27 seconds. However, the parameter count, training data, hardware model, batch configuration, and full evaluation methodology have not been disclosed. These remain vendor- and customer-reported results and should not be treated as directly reproducible benchmarks.
External services provide a more practical validation point. [Vercel AI Gateway](https://vercel.com/ai-gateway/models/mercury-2.5) already exposes routes for OpenAI Chat Completions, Responses, Anthropic Messages, and the AI SDK. As of September 10, the page showed median throughput of about 616 tokens per second and time to first token of roughly 1.4 seconds—well below the official peak. The gap does not necessarily indicate an error in the model or either test; it may result from shared-service conditions, reasoning effort, context length, batching, and network overhead.
Engineering teams should measure time to first token, steady-state throughput, total completion time, and tool-calling accuracy separately instead of comparing only tokens per second. The model is available exclusively through an API, with no open weights. Current pricing also includes an 80% launch discount, so architectural advantages, service-layer capacity, and promotional pricing must be evaluated independently.