Inception Mercury 3: Diffusion LLM Hits 1,068 Tokens/Sec

Inception Mercury 3 is the diffusion language model that wants to make "fast and dumb" a thing of the past. Released October 19, it writes whole answers in parallel instead of token by token, and Inception clocks it at 1,068 tokens per second on Nvidia GPUs. The company says it beats GPT-6 Luna and Claude Haiku 5.5 on most of its own benchmarks at four to eight times their speed.

What actually happened

Inception's launch post frames Mercury 3 as proof that "text diffusion scales." Most chatbots are autoregressive: they predict one token, then the next. A diffusion model drafts the full response and refines every position over a few passes, which is why it can be so fast. The open question since Mercury 1 has been whether that approach can get smart enough for real work.

Inception's answer is a set of company-run benchmarks. Per SiliconANGLE's write-up, Mercury 3 beat GPT-6 Luna and Haiku 5.5 on eight of 12 tests. Its biggest lead came on SWE-Bench Atlas, a coding test, at 51.1 versus 33.3 for GPT-6 Luna and 32.3 for Haiku 5.5. It trailed Haiku on DeepSWE, 46.3 to 51.6, and on one long-context test. No outside group had published independent results at launch.

The speed comparison is the headline. SiliconANGLE reports Haiku 5.5 at about 240 tokens per second and GPT-6 Luna at about 130 in the same setup, against Mercury 3's 1,068.

Who is already using it

Inception lists early partners in latency-sensitive jobs. TinyFish AI says Mercury 3 completed 50% of 50 live Mind2Web web-navigation tasks in its harness. Octen AI, which builds search for smart glasses, reports a median generation time of 0.9 seconds. Tutoring app RevisionDojo says hints and marked answers now come back in about a second, and support platform Lorikeet says it beat other models on latency in early tests.

Those are customer quotes in a launch post, so read them as marketing with numbers attached. They do show where Inception Mercury 3 is aimed: voice agents, browser agents, real-time search and support bots, where every second of waiting feels broken.

Mercury 3 also anchors a small product family. Mercury Voice, launched in late September, is tuned for time to first answer token in voice apps, and Mercury Decide is a classification model Inception says beats Jev on JevBench and LegalBench. Co-founder and CEO Stefano Ermon told SiliconANGLE the limits of diffusion come down to "compute and data, not fundamental algorithmic barriers," and that training on a successor is already under way.

The price math

At the 50% launch discount, Mercury 3 costs $0.10 in and $0.35 out per million tokens. SiliconANGLE notes GPT-6 Luna and Haiku 5.5 list at about $0.10 and $0.50 for typical prompts. So Mercury 3 is roughly price-matched on input and a bit cheaper on output during the promo, then pricier at list. The pitch is not "cheapest," it is "same tier, several times faster."

Inception also offers 100 million free API tokens to try it. The company raised a $50 million round led by Menlo Ventures last November, with Microsoft's M12 participating, and Reuters reported in May that Microsoft held acquisition talks. No deal has been announced.

What to do as a reader

If you build agents, run a quick A/B: same prompts, same tool calls, Mercury 3 versus your current small model, and measure total task time, not tokens per second. Diffusion speed helps most when an agent makes many short calls in a row. Our Claude Haiku 5.5 review and the MiniMax M3.1 Flash launch are the obvious baselines, along with DeepSeek V4.1 Flash.

Wait for third-party evals before betting a product on the benchmark table. If Inception Mercury 3 holds up outside its own lab, it is the first diffusion model that is both fast and in the same league as frontier speed models. If not, it is still a great option for latency-bound work where "good enough in one second" beats "great in five."

Sources: Inception launch post and SiliconANGLE coverage.