Diffusion and Chinese flash models squeeze the small-model tier, and the DeepSeek ban splits buyers by origin

The plain diffusion serving tier is largely taken. Inception already sells Mercury 2.5 as a cheap, fast diffusion API at $0.20 input and $0.75 output per million tokens, discounted to $0.04 and $0.15 at launch, and it is available on Bedrock, OpenRouter and Poe, with Azure Foundry also listed. On top of that, NVIDIA hosts the open-weight DiffusionGemma for free, and open engines like dInfer and SGLang let any inference provider add the tier themselves. A new endpoint selling cheaper diffusion tokens would compete on price against a funded incumbent and free hosting, with little else to set it apart. I also could not verify that dlmrt or the 2,788 tokens per second figure exists, so do not rely on a runtime advantage until it is confirmed. The trajectory drafter is the real opening. Diffusion drafting with autoregressive checking exists only as research (DEER, DFlash, DiffuSpec), all of it token-level speculative decoding. I found no product that drafts whole multi-step tool-call sequences with a diffusion model, has a frontier model approve or repair them, and is priced and measured on cost per completed agent task. A new entrant could win by owning that agent-level orchestration and verification layer. It would work with any diffusion model, including Mercury and DiffusionGemma, and would be judged on task success and cost for coding or back-office agents. A smaller, lower-margin option is private or on-premises hosting of open-weight diffusion models for customers who cannot send data to Inception.

• Inception Labs Mercury 2 / Mercury 2.5

  A commercial diffusion LLM API and essentially the serving tier described here. Mercury 2.5 (September 2026) is priced at $0.20 per million input and $0.75 per million output tokens, cut to $0.04 and $0.15 during a launch promotion. Inception claims about 1,107 tokens per second on widely available NVIDIA GPUs. It is sold through Bedrock, OpenRouter and Poe, and its llms.txt also lists Azure Foundry. The company raised $50M in late 2025.

• Mercury 2 on OpenRouter

  Mercury 2 is already available as a routable model in the OpenRouter marketplace that agent platforms use, at $0.25 input and $0.75 output per million tokens.

• Google DiffusionGemma (formerly Gemini Diffusion) on NVIDIA NIM

  Gemini Diffusion was an experimental preview in May 2025. The research came back in June 2026 as the open-weight, Apache 2.0 model diffusiongemma-26B-A4B-it, which NVIDIA hosts free on its NIM cloud API.

• dInfer (inclusionAI / Ant Group)

  An open-source, Apache 2.0 inference framework for diffusion LLMs. The authors report over 1,100 tokens per second at batch size 1 on 8 H800 GPUs. It makes self-hosting a diffusion tier easy for any provider.

• SGLang diffusion LLM support

  A mainstream serving engine that added diffusion LLM support, with day-zero support for LLaDA 2.0 and block-wise streaming. Existing inference providers can add a diffusion tier with little effort.

• Spheron dLLM deployment guide

  A GPU cloud that publishes production setup guides for LLaDA 2 and Mercury-style diffusion models. It shows that GPU clouds are already pitching diffusion serving as a self-hosted option.

• DEER: Draft with Diffusion, Verify with Autoregressive Models

  A research paper using a diffusion model as the drafter and an autoregressive model as the checker. It reports acceptance lengths up to 32 tokens and a 5.54x speedup on HumanEval. The framing mentions agents, but this is token-level speculative decoding, not a product for drafting tool-call trajectories.

• DFlash and DiffuSpec

  More research on diffusion drafters for autoregressive verification, one using a lightweight block-diffusion drafter and one training-free. The academic idea is established, but I found no product built on it.

Checked 2026-10-17