Arien0/Negentropy-claude-opus-4.7-9B-AWQ-BF16-INT8-AR-MTP

Model Card for Model ID

This is a mixed BF16-INT8 AWQ layer quantization, with working MTP (speculative decoding) via llmcompressor.

Model Details

📖 Model Description:

The "AR" in the name refers to my personal calibration mixed dataset used for this quant.

Fixed chat_template with "froggeric/Qwen-Fixed-Chat-Templates"

Working MTP with VLLM flag:

--speculative-config '{"method":"mtp","num_speculative_tokens":2}'

Tested with VLLM 0.19.1 and transformers 5.6.2

Recommended flags:

--enable-auto-tool-choice
--reasoning-parser qwen3
--tool-call-parser qwen3_xml

🏆 Ranking: Qwen3.5-9B Custom Series (AWQ-INT8 Benchmarks)

Rank	Model / Version	HumanEval (Code) ↑	Winogrande (Logic) ↑	HellaSwag (Context) ↑	WikiText (PPL) ↓	Verdict
1st	Negentropy-4.7-AR	0.7256	0.7474	0.7780	9.7812	Best for Agents & Coding
2nd	Qwopus-v3.5-CK	0.6829	0.7395	0.7807	9.6310	Most Balanced Performance
3rd	Qwopus-v3-NM	0.6768	0.7395	0.7820	9.6056	Solid All-Rounder
4th	DeepSeek-V4-Flash-AR2	0.6646	0.7451	0.7830	9.5984	Best for Natural Chat

🔍 Key Takeaways:

• Negentropy-4.7 (AR): Absolute winner for Agentic workflows. Its HumanEval score (0.7256) is massive for a 9B model, outperforming the others by a +4-6% margin. It is the most capable for logical tasks and complex instruction following.
• Qwopus-v3.5 (CK): Represents the best coding evolution within the Qwopus series, serving as the best alternative when a balance between reasoning and linguistic stability is required.
• DeepSeek-V4-Flash (AR2): While the least powerful in programming, it maintains the lowest Perplexity and the best HellaSwag, making it the most "human-like" and coherent for general-purpose conversation without tools.
• Efficiency: The Negentropy-4.7 (AR) was also the fastest to complete HumanEval (24s), suggesting high attentional efficiency. Combined with the injected MTP tensors, it offers the best intelligence-to-latency ratio.
• Quantization Impact: All AWQ-INT8 versions maintain near-lossless fidelity compared to their BF16 bases, but with an inference speed boost of approximately 30-40%.

🙏 Acknowledgements:

• Jackrong: for finetune the base model
• HuggingFaceH4: for the dataset
• cyankiwi: for the dataset
• cpatonn: for the quant recipe
• froggeric: for the fixed chat_template