NVIDIA and CohereLabs: Revolutionary New Open Models

⚡

Key Takeaways

1NVIDIA unveils Nemotron Super, a 120 billion parameter model, offering a major advancement in multilingualism.

2CohereLabs launches Cohere Transcribe, an audio-to-text transcription model, supporting 14 languages under Apache 2.0 license.

3SarvamAI and Mistral innovate with models tailored for Indic languages and hybrid capabilities.

💡Why it matters — These innovations enhance the accessibility and diversity of open models, which are essential for specialized and economic applications.

A New Era for Open Models

The latest edition of open artifacts highlights a series of innovative models that broaden the horizons of technological applications. Unlike previous editions dominated by giants like Qwen and DeepSeek, this selection stands out for its diversity and originality, covering areas such as optical character recognition, audio transcription, and even the demonstration of mathematical theorems.

This month's artifacts come from a much larger pool of open model creators. This hints at a promising future for open models, where the need for domain-specific and cost-effective models is seen as essential to complement the most powerful closed agents.

NVIDIA and CohereLabs in the Spotlight

NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4

NVIDIA has launched its highly anticipated model, the Nemotron Super, which features 120 billion parameters, of which 12 billion are active. This model offers an impressive context window of one million and supports several popular languages. It utilizes LatentMoE and NVFP4 technologies during pre-training, marking a first for open models.

Cohere Transcribe

CohereLabs has introduced Cohere Transcribe, a speech-to-text conversion model based on the conformer architecture, similar to NVIDIA's. This model supports 14 languages, including some AIPAC languages and Arabic, and is released under the Apache 2.0 license. Cohere claims that this model outperforms both open and closed models of similar size.

Innovations from SarvamAI and Mistral

Sarvam-105b

The Indian startup SarvamAI has expanded its datasets, reaching between 12 and 16 trillion tokens, and its models, reaching up to 105 billion parameters. These models are particularly sought after for Indic languages, surpassing the SOTA open models.

Mistral-Small-4-119B-2603

Mistral offers a hybrid model with 119 billion parameters, combining previous generations to provide advanced coding capabilities.

Other Notable Models

Among other notable models is Zeta-2 from Zed Industries, an open-source code editor based on Seed-Coder-8B, trained on open-source code by users who explicitly consented to data collection.

General Models

NVIDIA has also introduced the gpt-oss-puzzle-88B, an optimized version of GPT OSS 120B, which replaces some global attention layers with windowed attention. This model is designed to improve inference efficiency for heavy reasoning workloads.

AllenAI unveiled Olmo-Hybrid-7B, a hybrid model combining attention and GDN (gated DeltaNet). NVIDIA also presented a compressed version of its Nemotron model with the NVIDIA-Nemotron-3-Nano-4B-BF16.

Multimodal Models

YuanLabAI launched Yuan3.0-Ultra, a multimodal model with 1 trillion parameters, pre-trained on 2.2 trillion tokens. Meituan-LongCat introduced LongCat-Next, a multimodal model capable of processing text, vision, and audio.

IBM-Granite presented granite-4.0-1b-speech, a small speech-to-text conversion model supporting six languages. Microsoft launched Phi-4-reasoning-vision-15B, a model using the SigLIP-2 vision encoder.

Special Models

Miromind-AI introduced MiroThinker-1.7, a fine-tuned version of Qwen 235B for agentic workflows. Prior-Labs updated its tabular prediction model with tabpfn_2_6, slightly larger than its predecessor.

Facebook updated SAM 3 with sam3.1, under a restrictive license. Hcompany launched Holotron-12B, a policy model for CUA agents. Meituan-LongCat also introduced LongCat-Flash-Prover, a fine-tuned Lean4 of the large LongCat model.

RAG Models

Baidu launched Qianfan-OCR, an OCR model under the Apache 2.0 license. Datalab-TO updated the Chandra OCR model with chandra-ocr-2, under a restrictive license.

LightonAI presented Reason-ModernColBERT, a SOTA retrieval model under a non-commercial license. ChromaDB introduced context-1, a fine-tuned version of GPT-OSS for agentic search. Finally, Rednote-Hilab updated the dots.ocr model with dots.mocr, supporting SVG outputs but with additional restrictions on its use.