Inception Launches Mercury 2.5 Preview, a Diffusion LLM Claiming 1,107 Tokens/Sec
Inception released Mercury 2.5 Preview, a diffusion-based language model that generates tokens in parallel rather than sequentially, claiming throughput of 1,107 tokens per second on standard GPUs. The model is available on OpenRouter with a 260K context window and an 80% launch discount through September 8, 2026.
Mercury 2.5 Preview — Quick Specs
Inception Releases Mercury 2.5 Preview
Inception has released Mercury 2.5 Preview, the latest version of its diffusion large language model (dLLM), available now through OpenRouter. Unlike standard autoregressive models that generate tokens one at a time, Mercury 2.5 produces and refines multiple tokens in parallel, a diffusion-based approach the company says makes it the fastest reasoning LLM currently available.
Key specs
- Context window: 260,000 tokens
- Throughput: 1,107 tokens/sec on standard GPUs, according to Inception
- P50 latency: 1.38 seconds, with observed throughput of 223 tokens/sec on OpenRouter's tracked provider endpoint
- Pricing: $0.20 per 1M input tokens / $0.75 per 1M output tokens (list price); currently $0.04 / $0.15 per 1M tokens under an 80% launch discount running through September 8, 2026 at 07:00 UTC
- Cache read pricing: $0.02 per 1M tokens (discounted from $0.75)
- Release date: August 31, 2026
Claimed performance
Inception claims Mercury 2.5 delivers a "10+ point jump in intelligence" over its predecessor, Mercury 2, though the company did not specify which benchmark suite produced that figure. Inception positions the model's quality as comparable to cost-optimized frontier offerings, specifically naming GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5 as reference points. None of these comparisons have been independently verified.
Features
Mercury 2.5 supports tunable reasoning levels, letting developers dial compute up or down depending on task difficulty. It also supports parallel tool calls and schema-aligned JSON output, features aimed at agentic and structured-output workloads. Inception markets the model for latency-sensitive production use cases: search agents, voice pipelines, and coding subagents, where the cost of sequential token generation compounds across long interaction chains.
Availability data
OpenRouter's tracked deployment shows 100% uptime over the trailing three days but 97.13% availability over the same window, with 24-hour availability at 97.19%. OpenRouter notes that failed requests can be routed to alternate healthy providers if a request's routing filters allow it.
What this means
Mercury 2.5 is notable less for a specific benchmark score and more for its architecture: diffusion-based generation is still a minority approach among production LLMs, and a claimed 1,107 tokens/sec would be a meaningful jump over typical autoregressive throughput on comparable hardware. If Inception's speed claims hold up under independent testing, this positions Mercury 2.5 as a serious option for latency-bound applications like voice agents and multi-step coding subagents where waiting on token-by-token generation is the primary bottleneck.
The comparisons to GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5 suggest Inception is targeting the budget-tier segment of the market rather than competing with top-end frontier models on raw capability. The 80% discount pricing — effectively $0.04/$0.15 per 1M tokens against a $0.20/$0.75 list price — is a temporary promotional window ending September 8, 2026, after which buyers should expect list pricing to apply. As with all vendor-reported throughput and quality claims, third-party benchmarking will be needed to confirm the stated gains over Mercury 2.
Related Articles
Moondream 3.1: 9B MoE vision model with 2B active parameters, 32K context, $0.30/$2.50 per 1M tokens
Moondream has released Moondream 3.1, a mixture-of-experts vision-language model with 9B total and 2B active parameters and a 32K-token context window. It is listed on OpenRouter at $0.30 per 1M input tokens and $2.50 per 1M output tokens, with an optional reasoning mode.
ScaleDown launches Summarize, a 1M-context abstractive summarization model at $0.04 per 1M input tokens on OpenRouter
ScaleDown has released Summarize, an abstractive summarization model with a 1M-token context window, listed on OpenRouter on October 13, 2026. OpenRouter shows $0.04 per 1M input tokens (20% off the $0.05 list price) and $0 per 1M output tokens. Parameter count and benchmark scores have not been disclosed.
Moondream 3 Preview lists on OpenRouter: 9B MoE vision model, 2B active, 32K context, $0.30/$2.50
Moondream 3 Preview is a vision-language model with 9B total and 2B active parameters (mixture-of-experts), a 32K-token context window and an optional reasoning mode. OpenRouter lists it at $0.30 per 1M input tokens and $2.50 per 1M output tokens, served by Prime Intellect.
Unbiased releases Pareto 26.9: 262K-context composite model at $2.50/$7.50 per 1M tokens
Unbiased has released Pareto 26.9, a multimodal composite model with a 262K-token context window, priced at $2.50 per 1M input tokens and $7.50 per 1M output tokens. It is listed on OpenRouter with a Oct 13, 2026 release date. A cheaper 1.0M-context preview, Pareto 26.10, is already listed alongside it.
Comments
Loading...