nivvis/Qwen3.5-9B-EQ-v5.1-GGUF

Qwen3.5-9B-Heretic-v2-EQ-v5.1-GGUF

GGUF quantizations of nivvis/Qwen3.5-9B-EQ-v5.1 for llama.cpp, Ollama, and LM Studio.

Example outputs: EQ 9B v5.1 vs Vanilla Qwen3.5-9B — same prompt, same settings.

Available quantizations

Quant	Size	Notes
F16	~17.9 GB	Full precision, lossless conversion
Q4_K_M	~5.6 GB	Best 4-bit balance, recommended for most users

Qwen3.5-9B-EQ-v5.1

A DPO fine-tune of trohrbaugh/Qwen3.5-9B-heretic-v2 for emotional intelligence and empathetic response quality.

This is still intended as a general use model (agentic, coding, general chat). Tuning was light and precise — no capability regression.

What this model does

• Validates without sycophancy — empathizes with frustration without rubber-stamping bad behavior
• Sets boundaries warmly — names uncomfortable truths without lecturing
• Sounds human — conversational tone, not therapist-speak. Better tone vs vanilla Qwen 3.5, e.g. "It sounds like"

Benchmarks

All benchmarks: thinking=on, t=1.0, top_p=0.95 unless noted.

EQ-Bench 3

Rubric score = avg of 6 scored criteria × 5 (0-100), Opus 4.6 judged. Leaderboard scores from eqbench.com. Only qualitative criteria count (empathy, pragmatic EI, insight, social dexterity, emotional reasoning, message tailoring).

#	Model	Score
1	gpt-5.4	80.9
2	claude-sonnet-4-6	79.9
3	claude-opus-4-6	78.6
4	gemma-4-31B-it	72.8
5	Qwen3.5-397B	70.1
6	Qwen3.5-35b-EQ-v5.0	68.2
7	Qwen3.5-35b-EQ-v5.1	66.8
8	Qwen3.5-9b-EQ-v5.1	65.4
9	Qwen3-235B	61.4
10	gpt-4.5-preview	59.7
11	Qwen3.5-9b (vanilla)	59.1
12	o4-mini	58.1
13	DeepSeek-V3-0324	57.9
14	Qwen3.5-35b (vanilla)	55.4
15	gpt-oss-120b	51.4

HumanEval+

Model	HumanEval base	HumanEval+
EQ 9B v5.1	92.7%	86.0%
Vanilla Qwen3.5-9B	93.9%	87.2%

GSM8K

Metric	EQ 9B v5.1	Vanilla Qwen3.5-9B
Accuracy	84.9%	82.8%

IFBench

Metric	EQ 9B v5.1	Vanilla Qwen3.5-9B
Loose (leaderboard)	67.3%	65.0%
Strict	56.8%	56.8%

How to use

llama-server (OpenAI-compatible API)

llama-server \
  -m Qwen3.5-9B-Heretic-v2-EQ-v5.1-Q4_K_M.gguf \
  --host 0.0.0.0 --port 30000 \
  -ngl 99 --jinja

Ollama

ollama run hf.co/nivvis/Qwen3.5-9B-Heretic-v2-EQ-v5.1-GGUF:Q4_K_M

Thinking mode

This model supports thinking mode. To disable (for faster, direct responses):

{"chat_template_kwargs": {"enable_thinking": false}}

Sampling recommendations

Use the same settings as Qwen3.5-9B:

Mode	temp	top_p	top_k	presence_penalty
Thinking (general)	1.0	0.95	20	1.5
Thinking (coding)	0.6	0.95	20	0.0
Non-thinking (general)	0.7	0.8	20	1.5
Non-thinking (reasoning)	1.0	0.95	20	1.5

Other formats

• BF16 safetensors — original weights

Lineage

Qwen/Qwen3.5-9B
  → trohrbaugh/Qwen3.5-9B-heretic-v2 (decensored)
    → nivvis/Qwen3.5-9B-EQ-v5.1 (DPO for EQ)
      → this repo (GGUF quantizations)

License

Apache 2.0, following the base Qwen3.5 license.