Inception launches Mercury 3 diffusion LLM: 1M context, 1,068 tokens/sec, $0.20/$0.70 per 1M tokens
Inception released Mercury 3, a diffusion-based LLM with a 1M-token context window that it says runs at 1,068 tokens per second on NVIDIA GPUs. Pricing is $0.20 per 1M input and $0.70 per 1M output tokens, with a 50% launch discount. Inception claims quality on par with or better than GPT-6 Luna and Claude Haiku 5.5 at 4–8x the speed.
Mercury 3 — Quick Specs
Inception released Mercury 3 on October 19, 2026, a diffusion-based large language model with a 1M-token context window. According to the company, it generates 1,068 tokens per second on NVIDIA GPUs. Inception describes it as the most capable diffusion LLM released to date.
Specifications
- Context window: 1M tokens, up from 260K for Mercury 2.5
- Speed: 1,068 tokens per second on NVIDIA GPUs (company figure)
- List pricing: $0.20 per 1M input tokens, $0.70 per 1M output tokens
- Launch pricing: 50% off, or $0.10 input and $0.35 output per 1M tokens (Inception did not say how long the discount lasts)
- Cached tokens: 90% off the input price
- Parameter count: not disclosed. Inception says the model is more than 40x larger than Mercury 1 in total parameters.
- Training cutoff: not disclosed
Performance claims
Inception claims Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running 4–8x faster. The announcement does not give specific benchmark scores, and the comparisons have not been independently verified.
The company also says intelligence across standard benchmarks has risen 3x from Mercury 1 (mini) to Mercury 3. It argues this shows autoregressive-style scaling laws also hold for text diffusion.
The announcement includes the following customer-reported results:
- TinyFish AI: In its production web-agent harness, Mercury 3 completed 50% of 50 live Mind2Web navigation tasks. TinyFish says that is significantly more than GPT-6.1-Sol, using the same token volume and agent steps.
- Octen AI: In Sense API tests for smart-glasses search, Octen reports a median generation time of 0.9 seconds and a median end-to-end response time of 2.2 seconds. Octen says Mercury 3 gave the best speed-quality balance among the fast models it tested.
- RevisionDojo and Lorikeet: Both report quality at or above their previous models, and Lorikeet reports lower latency. Neither supplied figures.
All of these results come from early-access partners quoted in Inception's own post.
Availability
Mercury 3 is available through the Inception API Platform, OpenRouter, Baseten, and Vercel. It is OpenAI API compatible. Inception offers 100 million free tokens for API trials and a chat interface. Enterprise options include dedicated capacity, autoscaling, compliance controls, and configurable data retention.
Inception also describes two related products. Mercury Voice, announced September 29, targets time to first answer token in voice applications. Mercury Decide is a classification model for high-volume decision tasks.
What this means
Diffusion LLMs have mostly been viewed as a fast but limited alternative to autoregressive models. Mercury 3 is the first evidence from a commercial vendor that the approach scales to a 1M-token context and competes with small frontier models. That evidence is still company-reported. Without published benchmark numbers or third-party evaluations, the claim of parity with GPT-6 Luna and Claude Haiku 5.5 remains unproven.
The pricing and speed matter most for latency-bound workloads such as voice agents, browser agents, and real-time search. In those settings, generation time compounds across steps. At $0.20/$0.70 list price, Mercury 3 is positioned as a cost-efficient option for them.
The next step is independent testing on long-context retrieval, tool use, and reasoning-heavy tasks. Those results will show whether the 1,068 tokens-per-second figure holds on workloads other than Inception's own.
Related Articles
DeepSeek V4.1 Flash 1017 launches with 552B CED backbone, 1M context, $0.15/$0.60 per 1M tokens
DeepSeek has released V4.1 Flash 1017, an October 17, 2026 snapshot built on its new Causal Encoder-Decoder architecture. It uses a 552B-parameter sparse MoE backbone with a 1M-token context window, native image understanding, and pricing of $0.15 input / $0.60 output per 1M tokens on OpenRouter.
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.
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...