H Company has released NeoMME, a family of open-source multimodal encoders designed for visual document retrieval. The models, available in 260M and 800M parameter sizes, utilize a shared bidirectional Transformer architecture to improve efficiency on the ViDoRe v3 benchmark by eliminating the need for causal decoders.
The release of these open-source encoders provides a more efficient path for streaming and media enterprises to index high-resolution visual assets without the heavy compute costs of traditional generative vision-language models. By bypassing OCR and preserving document layouts, the technology streamlines the retrieval of complex metadata from charts and tables within large media libraries. This shift toward smaller, specialized bidirectional Transformers challenges the industry's reliance on massive causal decoders for non-generative tasks. As multimodal search becomes a standard feature for content management systems, watch for the adoption rate of these Apache 2.0 licensed models within Hugging Face Transformers to gauge their impact on production-scale retrieval-augmented generation pipelines.
The release of NeoMME arrives amid intensifying competition in open-source multimodal retrieval models. Hugging Face, which hosts NeoMME and the ViDoRe benchmark, has become the primary distribution channel for vision-language retrieval models, with ColPali and its derivatives establishing a new baseline for late-interaction visual document retrieval since mid-2025. The ViDoRe benchmark itself has evolved through multiple versions to test increasingly complex document layouts, and NeoMME's bidirectional architecture represents a deliberate departure from the causal-decoder approach that ColPali and similar models inherited from generative vision-language models. The business case for smaller, specialized encoders is gaining traction across the AI infrastructure stack. Nokia announced in June 2026 that its Autonomous Network Fabric would run on AWS, combining agentic AI with intent-based networking for cross-domain automation, illustrating how enterprises are shifting from monolithic foundation models toward modular, task-specific AI components that can be orchestrated across cloud environments. Similarly, Ericsson launched its AI in RAN commercial software subscription on June 11, claiming up to 20% higher downlink throughput across more than 15 live deployments, demonstrating that specialized AI models deployed at scale can outperform general-purpose systems on narrow tasks. This pattern of domain-specific efficiency over brute-force scale mirrors the architectural bet NeoMME makes for document retrieval. On the technical front, the bidirectional Transformer approach that NeoMME employs draws on the ModernBERT lineage, which demonstrated that encoder-only architectures can match or exceed decoder-based models on retrieval and classification tasks at a fraction of the compute cost. Ericsson's agentic AI blueprint places a Telco DataOps Platform at the center of its OSS/BSS architecture, using streaming data pipelines to clean and contextualize information before agents make decisions, a pattern that parallels how NeoMME's encoders are designed to feed structured embeddings into enterprise AI agent architectures for retrieval-augmented generation systems. The Apache 2.0 licensing under which H Company released NeoMME removes commercial restrictions, positioning the models for integration into proprietary content management and media asset platforms where permissive licensing is a prerequisite for production deployment.
H Company has released NeoMME, a family of open-source multimodal encoders available under the Apache 2.0 license. By utilizing a bidirectional Transformer architecture, these models eliminate the need for causal decoders, significantly reducing compute costs while maintaining high performance for indexing high-resolution visual assets in media and streaming libraries.
NeoMME uses a shared bidirectional Transformer architecture that removes the need for causal decoders. This design reduces compute overhead, supports 16,384 tokens for processing two 4K UHD images simultaneously, and enables faster encoding throughput compared to traditional generative vision-language models.
The NeoMME-Retriever-260M model achieves an nDCG@10 score of 0.523, matching the performance of ColQwen2.5 while utilizing 14 times fewer parameters.
H Company released the NeoMME multimodal encoders under the Apache 2.0 license, which removes commercial restrictions and allows for integration into proprietary content management and media asset platforms.
By bypassing OCR preprocessing and using compression techniques like hierarchical token pooling and asymmetric quantization, NeoMME reduces storage requirements by up to 255 times and achieves encoding speeds of 51 pages per second on an NVIDIA L40S.
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source