NeoHorse-1-9B FP8 Calibrated

NeoHorse is a recent Qwen finetune that we started using and liked. You can find it here: TokenRhythm/NeoHorse-1-9B

This version quantizes both the weights and the cache to FP8 to run on smaller hardware. We kept the FP8 weights at block over dynamic quant for higher quality.

Quantization

• FP8 block-scaled weights with 128 x 128 weight blocks
• Dynamic grouped FP8 input activations
• Static tensor-wise FP8 KV-cache scales calibrated from 512 UltraChat samples
• Calibration sequence length: 2,048 tokens
• Quantized with LLM Compressor 0.13.0
• Serialized with the compressed-tensors format

The embedding table, language-model head, and sensitive hybrid linear-attention state remain in the source precision.

Validation

The packaged validation report records:

• 643 readable output tensors
• 200 FP8 weight tensors and 200 matching weight-scale tensors
• 8 calibrated K-cache scales and 8 calibrated V-cache scales
• 74 protected tensors that exactly match the source checkpoint
• complete SHA-256 checksums for the release files

See VALIDATION_REPORT.json, recipe.yaml, and SHA256SUMS for the recorded artifact checks.

Compatibility

The checkpoint uses the Qwen3.5 text architecture and the compressed-tensors FP8 format. Use a recent runtime that supports both the architecture and FP8 block quantization. Our tests were on vLLM on the DGX Spark.

License

The source model is licensed under Apache 2.0. See the source repository for its model documentation and usage terms.