GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-Distill-GGUF

GLM 4.7 Flash x Claude 4.5 Opus (High Reasoning)

This model was trained on a small reasoning dataset of Claude Opus 4.5, with reasoning effort set to High.

• 🧬 Datasets:

  • TeichAI/claude-4.5-opus-high-reasoning-250x

• 🏗 Base Model:

  • unsloth/GLM-4.7-Flash

• ⚡ Use cases:

  • Coding
  • Science
  • Deep Research

• ∑ Stats (Dataset)

  • Costs: $ 52.30 (USD)
  • Total tokens (input + output): 2.13 M

---

How to run

For specific instructions/commands to serve this model locally using vLLM, SGLang, or transformers please see the instructions from the original model's card

For detailed instructions getting started with Llama.cpp please refer to the unsloth guide

Sampling Parameters

z-ai recommends the following sampling parameters for this model:

• For general use-case: --temp 1.0 --top-p 0.95
• For tool-calling: --temp 0.7 --top-p 1.0
• If using llama.cpp, set --min-p 0.01 as llama.cpp's default is 0.05
• Sometimes you'll need to experiment what numbers work best for your use-case.

If you experience any issues with these parameters, some users have reported better results when lowering temperature to 0.5-0.6

---

Benchmarks

Model Comparison vs Base

• Base model: zai-org/GLM-4.7-Flash

Aggregate Comparison

Detailed Results

MMLU Subject Breakdown

MMLU Detailed Results

Benchmark Config

• Quantization: 4bit
• Temperature: 0.0
• Top P: 1.0
• Top K: 0
• Repetition Penalty: 1.0

All results were obtained through the official lm evaluation harness

---

This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.