Live intelligence

Latest AI updates

4 updates
Industry

Nvidia bets at least $500 million on Inception's diffusion models to strengthen its open-source Nemotron lineup

Nvidia is reportedly close to acquiring AI startup Inception for at least $500 million, according to The Information. Inception builds language models using diffusion processes rather than token-by-token generation. The deal would support Nvidia's open-source Nemotron models and its robotics efforts. The acquisition is not confirmed in the source material.

NvidiaInceptiondiffusion language models
Models Editor selected

NVIDIA Releases Nemotron-Labs-Sigma, a Continuous Diffusion Language Model Family at 3B and 8B

NVIDIA released Nemotron-Labs-Sigma, a family of 14 open checkpoints at 3B and 8B parameters spanning base, instruct, few-step distilled, and self-speculative variants. The model family performs continuous diffusion in a learned 16-D embedding space, with a 4,096-token context that extends to 8,192 for Sigma-8B-instruct. NVIDIA reports Sigma-8B base averaging 71.16 on its 6-task suite versus 70.22 for NLD-8B, while the instruct 10-task average is 59.14, slightly below NLD-8B's 59.92. The self-speculative 8B checkpoint runs 3.46x faster than autoregressive decoding on one GB200 with SGLang, according to the company.

Why it matters: It gives readers a concrete open-weight model release with architecture details, benchmark results, deployment constraints, and inference-speed claims for continuous diffusion language models.

NVIDIANemotron-Labs-SigmaDiffusion language models
Research Editor selected

TCDA Adapts Denoising Schedules for Diffusion Language Models

Researchers propose Task-Calibrated Denoising Allocation (TCDA), a deployment rule that selects the cheapest certified denoising schedule for diffusion language models while providing an explicit error guarantee. TCDA uses simultaneous one-sided bootstrap bounds to certify candidate samplers against utility and guardrail criteria, falling back to the full schedule when no candidate is certified. Across seven datasets and two model families, it certifies reductions in denoiser evaluations, including factors of 3.4 on XSum and 2.0 on WMT14 for ELF-B, and 8.0 on WMT14 and 4.1 on GSM8K for LLaDA-8B. The work also notes that the ELF reference sampler draws its time grid randomly, so reproducible unconditional-generation comparisons require fixing the grid.

Why it matters: It offers a concrete, statistically certified method for reducing inference cost in diffusion language models, with reported gains across multiple benchmarks and model families.

Diffusion language modelsInference optimizationTCDA
Models

Inception Releases Mercury 3 Diffusion LLM at 1,068 Tokens per Second

Inception has launched Mercury 3, a diffusion language model that refines answers in parallel rather than token by token, claiming 1,068 tokens per second on NVIDIA GPUs and a 1M-token context window. Inception's own benchmark table shows Mercury 3 leading GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 tests, with its biggest gains on SWE-Bench Atlas (51.08) and IFBench (76.84), while Haiku 5.5 still wins DeepSWE v1 and MRCR v2. The API-only model is available through the Inception API Platform, Ope

InceptionMercury 3diffusion language models