# Mercury 3, CERN and Mistral, Mamba-2 in cuTile, TorchNeuron

> Inception ships Mercury 3 at over 1,000 tokens per second, CERN and Mistral plan a collider physics model, plus cuTile and TorchNeuron tooling.

Published 19 Oct 2026. Coverage: 18 Oct 2026 to 19 Oct 2026. By Ali Asif.

## Items

1. **Mercury 3: Diffusion at scale** (Inception Labs, Frontier): A diffusion-based LLM with a 1M token context that Inception says runs at 1,068 tokens per second and costs $0.20 per million input tokens. [Primary source](https://www.inceptionlabs.ai/blog/introducing-mercury-3)
2. **CERN and Mistral join forces to accelerate AI for science** (CERN, Science): A statement of intent to build a Large Physics Model, a foundation model for collider physics trained on simulated LHC collisions. [Primary source](https://home.cern/cern-and-mistral-join-forces-to-accelerate-ai-for-science)
3. **Accelerating Mamba-2 with forward pass kernels in NVIDIA CUDA Tile** (NVIDIA, Agents and coding): Three cuTile kernels for the Mamba-2 forward pass, reported 1.14x to 1.82x faster than the open-source Triton reference on a B300. [Primary source](https://developer.nvidia.com/blog/accelerating-mamba-2-with-high-performance-forward-pass-kernels-in-nvidia-cuda-tile)
4. **TorchNeuron: running HF models on AWS Trainium** (Hugging Face, Agents and coding): A PyTorch backend that exposes Trainium as a native device, covering 89% of the top 100 text-generation architectures versus 46% before. [Primary source](https://huggingface.co/blog/hf-torchneuron)

One model launch, one science partnership and two tooling posts today. I read each primary page; I ran none of the code, and the figures are the publishers' own.

## Mercury 3 from Inception Labs

Inception describes Mercury 3 as the most capable diffusion-based language model released to date. Instead of writing one token at a time, a diffusion model refines many tokens in parallel, which is where the speed comes from. The post reports 1,068 tokens per second on NVIDIA GPUs and says the model beats GPT-6 Luna and Claude Haiku 5.5 on a range of benchmarks while running 4x to 8x faster. It is roughly 40x larger in total parameters than Mercury 1, and the context window grows from 260K tokens on Mercury 2.5 to 1M.

Partner numbers are in the post too. TinyFish reports Mercury 3 completing 50% of 50 live Mind2Web navigation tasks, and Octen reports a median end-to-end search response of 2.2 seconds.

Pricing is $0.20 per million input tokens and $0.70 per million output tokens, with a 50% launch discount and 90% off cached input. It is OpenAI API compatible and available on Inception's platform, OpenRouter, Baseten and Vercel, with 100 million free tokens to start. For anyone building agents, that combination of price, speed and a drop-in API is easy to try, so I will.

## CERN and Mistral plan a Large Physics Model

On 15 October CERN and Mistral signed a statement of intent for a long-term partnership on foundation models for scientific discovery, starting with high-energy physics. The stated goal is a single pretrained Large Physics Model, described as the first foundation model for collider physics, trained on billions of simulated proton-proton collisions at the LHC. Intended tasks include reconstructing collisions, identifying particles, flagging unexpected signals and evaluating future detector designs.

The case for it is that reconstruction today runs through a chain of specialised algorithms, and one model could learn across that chain. CERN holds about one exabyte of LHC data, and the High-Luminosity LHC, expected around 2030, should raise collisions tenfold. This is an intent, not a release, so there is nothing to benchmark yet.

## Mamba-2 kernels in cuTile

NVIDIA's developer blog describes Mamba-2 forward pass kernels written in CUDA Tile, a Python DSL, with Jane Street ML engineers. Three kernels cover the cumulative decay, the state update across chunks and the fused output step. On a single B300 with 16,384 tokens, they run 1.14x to 1.82x faster than the open-source Triton reference, with the gap growing as sequence length and state size rise. These are kernel-level results, not end-to-end training numbers, and the code is available through TileGym.

## TorchNeuron on Trainium

Hugging Face says TorchNeuron enters public beta on 20 October and installs with `pip install torch-neuronx`. It makes Trainium a native `neuron` device, so Transformers, Diffusers, Accelerate, PEFT and TRL run with the same classes used elsewhere. Text generation support covers 89% of the top 100 architectures, up from 46% with Optimum Neuron v0.4.5. Hugging Face also reports a vision transformer kernel up to 2.4x faster.

## What I'd watch next

Independent latency and quality tests of Mercury 3 on tool-calling workloads, and whether CERN and Mistral publish a first dataset or evaluation plan. I also want to see end-to-end Mamba-2 training numbers, and how the Trainium beta holds up once it is public.

## My take

Mercury 3 is the item I would test first, because a price of $0.20 per million input tokens at that speed is a real option for agent loops where latency is the bottleneck. I read all four pages and ran none of the code, so every number here is the publisher's own.

## Sources

- https://www.inceptionlabs.ai/blog/introducing-mercury-3
- https://home.cern/cern-and-mistral-join-forces-to-accelerate-ai-for-science
- https://developer.nvidia.com/blog/accelerating-mamba-2-with-high-performance-forward-pass-kernels-in-nvidia-cuda-tile
- https://huggingface.co/blog/hf-torchneuron

Canonical: https://heyaliasif.com/ai/2026-10-19-mercury-3-cern-mistral-cutile-torchneuron