Inception's most capable diffusion LLM yet — a text model that refines tokens in parallel to hit over 1,000 tokens per second, now with a 1M-token context window.
Last verified · 2026-10-19 · by Moe Ameen
Mercury 3 is a diffusion large language model (a dLLM) from Inception, the company founded by Stanford professor Stefano Ermon to commercialize diffusion-based text generation. It was released on October 19, 2026 and billed as the most capable diffusion-based model to date. Where a standard autoregressive model writes one token at a time from left to right, a diffusion LLM starts from a rough draft of the whole output and refines tokens in parallel — the mechanism behind its headline speed. Inception reports 1,068 tokens per second on NVIDIA GPUs and says Mercury 3 runs 4–8x faster than GPT-6 Luna and Claude Haiku 5.5 while matching or beating them on quality.
The two biggest changes over Mercury 2.5 are scale and context. Inception says Mercury 3 is more than 40x larger than Mercury 1 in total parameters, and the context window jumps to 1 million tokens, up from 260K. That large window is the feature most likely to change a creator's workflow: you can feed an entire podcast, webinar, or a quarter of transcripts into a single prompt and draft from all of it at once. It sits in the frontier speed-optimized tier — fast, cheap, and strong, rather than pitched to win the hardest reasoning tasks.
You reach it through Inception's own API platform, OpenRouter, Baseten, and Vercel, over an OpenAI-compatible endpoint — an app already wired to OpenAI can switch with minimal changes. There are 100 million free tokens to evaluate it. Standard pricing is $0.20 per million input tokens and $0.70 per million output, with a 50% launch discount ($0.10 input, $0.35 output) and cached input tokens 90% off; launch promotions expire, so budget against the standard rates.
The boundary worth planning around: Mercury 3 is a text engine, not a content workflow. It drafts words — hooks, scripts, captions, articles, JSON — extremely fast, and stops there. It does not govern brand voice, generate images or video, fan one idea into a multi-format set, or publish anywhere.
The feature that changes a creator's day with Mercury 3 is the 1M-token context. You can drop an entire 90-minute podcast transcript, a webinar, or a quarter's worth of your past posts into one prompt and have it draft a month of angles, hooks, and scripts from the whole thing at over 1,000 tokens per second — reading the mountain and returning drafts in the time a slower model skims a chapter. But a month of fast drafts grounded in a huge source is still just text, and text is not a content calendar. That is the precise hand-off into [Kompozy](/): Mercury 3 mines the long source for raw material; Kompozy turns the best of it into finished, published content the model cannot touch.
Concretely, take the strongest Mercury 3 script and in Kompozy it becomes a [Persona Short](/glossary/persona-shorts) — a HeyGen talking-head avatar reading your words with auto-captions burned in — while a drafted post becomes a brand-exact [Carousel](/glossary/hyperframes), Quote Graphics, and Photo Posts, each rewritten per platform and held to your voice by a [Persona Brief](/glossary/persona-brief) with a banned-word filter. Because Kompozy fans one idea into a set across the five output buckets and then [Autopilot](/glossary/autopilot) schedules and publishes it across the eight social platforms plus blog and email with a per-post review step, the long transcript Mercury 3 digested becomes weeks of on-brand video, images, and posts — not a document you still have to design and post by hand. For the previous release on the same rails, see [Mercury 2.5](/ai-tools/mercury-2-5); for an open diffusion model, [DiffusionGemma](/ai-tools/diffusiongemma).
Mercury 3 is a diffusion large language model from Inception, released October 19, 2026. Instead of generating one token at a time, it refines tokens in parallel, reaching 1,068 tokens per second on common GPUs. It carries a 1M-token context window and is built for fast text generation, low-latency agents, and long-document drafting.
Mercury 3 raises the context window to 1 million tokens (from 260K in Mercury 2.5), scales to more than 40x the parameters of Mercury 1, and Inception reports higher throughput plus quality on par with or better than fast frontier models like GPT-6 Luna and Claude Haiku 5.5. Both are diffusion LLMs.
It excels at drafting text at volume and from long sources: hooks, captions, video scripts, article and newsletter drafts, and structured JSON. The 1M-token context lets it draft from a whole transcript at once. It does not make images, video, or branded posts — pair it with a content engine like Kompozy for the media, branding, and publishing.
On its own Mercury 3 produces text and stops. In Kompozy you bring a Mercury-drafted script or post in and it becomes finished content: a Persona Short avatar video, a HyperFrames carousel, or Photo Posts, each refined into your brand voice via the Persona Brief and scheduled across eight social platforms plus blog and email — the steps Mercury does not do.
Standard pricing is $0.20 per million input tokens and $0.70 per million output tokens, with a 50% launch discount to $0.10 input and $0.35 output, cached input 90% off, and 100 million free tokens. Launch promotions expire, so budget against the standard rates and confirm current pricing on Inception's site.