M3TiS-33B: Nanbeige4.3-33B EXL3 3.5 bpw + DFlash2 draft head + TensorFold server (engine aacb6300)
ae9d494 verified | M3TiS-33B | |
| An independent build by CanadaApollo6, not affiliated with Nanbeige or the TensorFold authors. | |
| The model | |
| --------- | |
| Nanbeige4.3-33B, by the Nanbeige team. https://huggingface.co/Nanbeige/Nanbeige4.3-33B | |
| Licensed under the Apache License, Version 2.0 (LICENSE). | |
| Changed by CanadaApollo6: | |
| model-*.safetensors the weights, quantized to EXL3 3.5 bits per weight (exllamav3 1.6.5) on Nanbeige's own text | |
| draft/ a DFlash2 draft head, trained by CanadaApollo6 on Nanbeige4.3-33B's outputs | |
| The tokenizer and config files are Nanbeige's, with the quantization settings added to config.json. | |
| README.md is this build's card, not Nanbeige's. | |
| The server | |
| ---------- | |
| bin/tensorfold-native is built from TensorFold (https://github.com/ashhart/TensorFold), Copyright 2026 TensorFold | |
| contributors, with changes by CanadaApollo6 (SOURCE.md). Licensed under the Apache License, Version 2.0 | |
| (LICENSES/Apache-2.0.txt). TensorFold releases up to 0.5.0 were MIT-licensed, and code written before 0.6.0 keeps | |
| its MIT notice in LICENSES/MIT.txt. Third-party code and its licenses are listed in THIRD_PARTY_NOTICES.md. | |