#1
#1 out of 1
technology5h ago
Multiverse Computing releases a compressed version of OpenAI's gpt-oss-120B
- Multiverse Computing released HyperNova 60B 2602, a compressed OpenAI-compatible model for free on Hugging Face.
- The model reportedly cuts memory needs from 61GB to 32GB while keeping near-parity tool-calling performance.
- CompactifAI, Multiverse’s compression tech, reorganizes transformer weights using quantum-inspired tensor networks after training.
- The process is post-training and does not require retraining or access to original training data.
- CompactifAI targets memory-bound scenarios, enabling deployment on cloud, on-prem, and edge environments.
- The technique is architecture-agnostic within transformers and does not alter model APIs.
- The compression claims up to a 93% memory reduction with manageable accuracy loss.
- Executives frame the work as an answer to sovereign AI and energy-use concerns.
- The model supports various hardware targets and can run on edge devices with suitable accelerators.
- Experts view the release as a potential path to more affordable AI deployment amid rising infrastructure costs.
Vote 0