中文Sign up

vanch007/Index-Translate-35B-A3B-preview-oQ6e-fp16-mtp

Source: 🤗text-generationapache-2.036B params3B activated73 GBsafetensors✓ 8 checksumsupdated 18d ago
or submit a torrent

Index-Translate-35B-A3B-preview-oQ6e-fp16-mtp

MLX safetensors conversion of IndexTeam/Index-Translate-35B-A3B-preview, using oMLX's native enhanced oQ quantizer. Source revision: 207935a9a7cb331e14dcecd6938af4878ecee73b.

The per-module precision map, tensor shapes and storage dtypes match symrex/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V13-dequantized-oQ6e-fp16-mtp. All model weights and tokenizer files come from Index-Translate; no reference-model weights or calibration statistics were copied.

Precision and calibration

  • Default affine quantization: 6 bits, group size 64.
  • Exact reference overrides: 187 modules at 8-bit/group-64, 30 at 6-bit/group-128, and 123 at 8-bit/group-128.
  • 483 quantized modules; unquantized text weights and quantization scales/biases use FP16, while vision weights use FP32 as in the reference.
  • Native source MTP retained: 42 tensors after MLX conversion.
  • Source-specific oQe activation calibration: requested 128 sequences × 512 tokens, processed 128 sequences with adaptive expert coverage. See oq_imatrix_report.json for actual coverage.
  • Safetensors weights: 31.18 GB.

The native quantizer selects a memory-safe source-derived calibration proxy when the full checkpoint exceeds its live memory budget. The source-specific imatrix is never borrowed from the alignment model.

oMLX usage

Download this repository into an oMLX model directory and refresh the model list. For translation, use greedy decoding (temperature=0) and disable thinking (chat_template_kwargs={"enable_thinking": false}). Use the source model's training-aligned prompt:

请将以下文本翻译为英语，直接输出翻译结果，不要进行任何解释。

你好，世界！

The checkpoint configuration declares a 262,144-token context; practical context length depends on available memory. Native MTP weights are present and the oMLX MTP path passed a short greedy-parity test. Local defaults keep MTP disabled, matching the reference's current settings. No speed improvement or broad translation benchmark result is claimed for this conversion.

Validation

Source download integrity, complete tensor index, every reference tensor shape/dtype, and the exact quantization map passed verification. oMLX translation tests passed Chinese-to-English, English-to-Chinese, JSON/placeholder preservation, and native MTP greedy-output parity. Runtime library versions: {"mlx": "0.32.2", "mlx-lm": "0.31.4.dev132+g94cdcae13", "mlx-vlm": "0.7.1"}. These are smoke tests, not a multilingual quality benchmark. The model's published translation training and benchmark results are described in the original model card.

License

Apache-2.0, inherited from the source checkpoint. See LICENSE.