Inception Launches Mercury 3, a Diffusion LLM With 1M-Token Context at 1,000+ Tokens Per Second

On October 19, 2026, Inception announced Mercury 3, the newest release in its line of diffusion-based large language models, calling it the most capable diffusion LLM ever released. The Redwood City, California company says the model is more than 40 times the size of Mercury 1 while holding inference speed above 1,000 tokens per second, a figure it reports as 1,068 tokens per second on NVIDIA GPUs.

A second path to fast inference

Diffusion models take a different route to text than the autoregressive models that dominate the industry. Where a standard model generates text one token at a time, each token waiting on the one before it, a diffusion model starts from a rough draft of the whole answer and refines every position in parallel over a few passes. Fewer sequential steps means faster output on the same hardware.

When Inception introduced Mercury 1 in June 2025, the company notes, diffusion LLMs were still largely an academic pursuit, with early models limited to simple tasks like code autocomplete. The open question was whether the parallel-generation speed advantage would survive as the models grew more capable and took on harder tasks.

Mercury 3 is Inception’s fourth Mercury release. Mercury 2 introduced reasoning to diffusion models, and Mercury 2.5 added tunable reasoning along with a 260,000-token context window. Inception says Mercury 3 is the largest jump in intelligence of any Mercury release and expands the context window to 1 million tokens.

Key details

Inception compares Mercury 3 against frontier speed-optimized models, specifically OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5, and says Mercury 3 matches or beats their quality while running 4 to 8 times faster. Those are the company’s own benchmark numbers: MarkTechPost’s write-up of Inception’s table notes that Mercury 3 leads on 8 of the 12 listed tests at medium settings. The biggest gaps favor Mercury 3 in instruction following and agentic coding. On IFBench it scores 76.84 against 61.82 for GPT-6 Luna and 54.65 for Claude Haiku 5.5. On SWE-Bench Atlas it reaches 51.08 against 33.33 and 32.26. On GPQA Diamond it scores 89.23 against 86.36 and 84.51.

The comparisons are not all one-sided. Claude Haiku 5.5 wins DeepSWE v1 at 51.62 against Mercury 3’s 46.31, and MRCR v2 at 78.40 against 75.50. GPT-6 Luna edges ahead on AA-Omniscience-Public, 46.94 to 46.44, and on AA-LCR v1.1, 78.30 to 77.67. Long-context retrieval and longer coding tasks remain the weak spots.

Pricing is aggressive at launch. Inception lists standard rates of $0.20 per million input tokens and $0.70 per million output tokens, discounted 50 percent at launch to $0.10 and $0.35, with cached input tokens priced 90 percent below the input rate. New accounts also get 100 million free tokens, according to MarkTechPost.

Mercury 3 is available through the Inception API as well as Baseten, OpenRouter and Vercel. The API is compatible with the OpenAI API format, which Inception says lets teams switch without rewriting code, and uses the model identifier “mercury-3” with a “reasoning_effort” parameter. It is API-only: Inception has not released model weights, so developers cannot self-host it. Enterprise plans add dedicated capacity, autoscaling, compliance controls and configurable data retention. Inception also points to two companion models, Mercury Voice for voice agents and Mercury Decide for classification.

Early-access partners are already testing the model in production. Inception lists browser use agents, interactive education, real-time search agents and customer support as current uses. Michael Tong, co-founder and CTO of General Learning (RevisionDojo), is quoted saying Mercury 3 let his team give every student instant feedback without giving up reasoning, with hints and marked answers returning in about a second.

“Since we launched Mercury 1, the question we’ve heard most is whether diffusion can scale,” said Stefano Ermon, Inception’s co-founder. “We’re showing that the limits to scaling diffusion models are compute and data, not fundamental algorithmic barriers.”

Why it matters

Mercury 3 matters less as a single product than as evidence for a second inference architecture. If diffusion models keep their speed advantage while closing the quality gap with autoregressive models, latency-sensitive workloads like voice agents, browser automation and coding assistants could get meaningfully faster without a capability tradeoff. In agentic systems, where the model is called repeatedly to finish one task, small delays compound at every step, so raw generation speed has outsized value.

The launch also sharpens price competition at the fast-model tier. At $0.10 and $0.35 per million tokens during the launch period, Mercury 3 undercuts the speed-optimized offerings from the larger labs. Two caveats deserve emphasis. First, every benchmark number cited here comes from Inception’s own testing, not independent evaluation, and AI News noted it has not independently benchmarked the model. Second, the API-only release means anyone who wants the speed must route their data through Inception’s infrastructure rather than running the model themselves. Even so, Mercury 3 is a concrete test of whether the industry’s near-universal autoregressive default is the only way to build fast models.

Sources

• Inception Launches Mercury 3, Setting a New Pareto Frontier in Speed and Intelligence – BusinessWire press release, October 19, 2026
• Inception Releases Mercury 3 Diffusion LLM at 1,068 Tokens per Second – MarkTechPost, October 19, 2026
• Mercury 3 Launches: Context, Speed Claims and Pricing – AI News, October 19, 2026