Mistral Large 4 weights, Pika H3 Flex, SFX v3, Granite Switch

Five releases today, and the largest one is a follow-through: the Mistral Large 4 weights I flagged as coming at the end of the month are out.

Mistral Large 4 weights

Mistral released the weights of Mistral Large 4, which it calls Le Chonk: a 1-trillion-parameter, natively multimodal model trained on European infrastructure. The company says three more weeks of reinforcement learning since the public preview improved coding and defensive cybersecurity, and it reports gains in reasoning efficiency too.

The numbers it publishes are specific. On Cognition's private FrontierCode 2.0 benchmark, which asks whether a repository maintainer would merge the agent's change, it scores 19.7%. On Terminal-Bench 4.0 it reaches 41%, up 13 points from the preview, with 30% fewer output tokens. On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with GLM-5.3 while using 47% fewer tokens per task.

There are now three reasoning modes: low, medium and high, with high as the default. Low replaces the old "none" mode. In Surge Code's blind pairwise tests, Mistral says graders preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, and only Claude Opus 5.5 beat it.

Pika H3 Flex

Pika published H3 Flex on Hugging Face as a post-trained adapter for MiniMax H3. It generates video with synchronized audio in six sampling steps and accepts up to 12 reference inputs, so shot direction comes from references and prompts together. Local open-weight inference covers native 480p and 768p with mixed stereo audio, while Pika's hosted API goes up to 4K. The license is the MiniMax H3 community agreement, which is worth reading before commercial use.

ElevenLabs SFX v3

ElevenLabs updated its sound effects model. SFX v3 follows prompts that describe several sounds at once more reliably, produces better music and vocals, and gives cleaner audio than SFX v2 in the company's benchmark testing. It is in ElevenCreative and ElevenAPI today.

Granite Switch 4.1.1

IBM's patch release is small on purpose. It adds an updated Granite Speech model for 10 languages with native punctuation and capitalization, plus an improved Decision adapter. The base model is untouched, so you compose the new adapters and deploy. IBM says an Adapter SDK is coming so others can ship adapters the same way.

NVIDIA and interatomic potentials

NVIDIA's technical blog shows mixed-precision tricks for machine learning interatomic potentials. BF16x9 on Blackwell speeds up large matrix multiplies in the Point Edge Transformer by 2.2 to 2.5x for a 17,280-atom system. TF32x3 on Ampere speeds the full molecular dynamics step by up to 1.5x while keeping energy conservation close to FP32, where one-pass TF32 drifts.

What I'd watch next

Independent runs of Mistral Large 4 on real repositories, and whether the token-efficiency claim holds outside Mistral's chosen benchmarks. I also want to see the Granite Adapter SDK, since patching a model by swapping adapters is a cleaner story than retraining.