Qwen3.5-35B-A3B-EQ-v5-GGUF

GGUF quantizations of nivvis/Qwen3.5-35B-A3B-EQ-v5 for llama.cpp, Ollama, and LM Studio.

Available quantizations

Quant	Size	BPW	Notes
F16	~65 GB	16.01	Full precision, lossless conversion
Q4_K_M	~20 GB	4.88	Best 4-bit balance, recommended for most users

Qwen3.5-35B-A3B-EQ-v5

A DPO fine-tune of Qwen3.5-35B-A3B-heretic-v2.

The tune optimized for two things:

bringing warmth, emotional intelligence, general chat improvement to Qwen 3.5 series
countering some negative tendencies of Heretic models (overwillingness to agree, be sycophantic, etc)

This is still intended as a general use model (agentic, coding, general chat). Tuning was lightly & with precision. More general benchmarks to follow.

What this model does

This model is trained to be a better conversational partner in emotionally complex situations, while maintaining base model capabilities. It:

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"~~

Key specs


Base	Qwen/Qwen3.5-35B-A3B
Parent	llmfan46/Qwen3.5-35B-A3B-heretic-v2 (decensored via MPOA+SOMA)
Fine-tune	DPO with LoRA (r=32, alpha=64)
Training data	DPO preference pairs with diverse, simulated (real-situation-based) generated dialogue

EQ-Bench 3 results

Ranked #8 on raw score* EQ-Bench 3 with only 3B active parameters — competitive with frontier models at a fraction of the compute.

#	Model	Raw Score
1	horizon-alpha	202.3
2	Kimi-K2-Instruct	202.0
3	gemini-2.5-pro-preview-06-05	200.5
4	o3	199.0
5	gpt-5	195.6
8	EQ-v5 (this model, 3B active)	193.6
10	claude-opus-4	192.6

*Table lists all models available in EQ-Bench 3 repo (so known judge, settings etc so we can be as apples to apples on raw score). Still raw score is not ideal. ELO submission pending. Better than no stats!

See the BF16 model card for full benchmarks and training details.

How to use

llama-server (OpenAI-compatible API)

llama-server \
  -m Qwen3.5-35B-A3B-EQ-v5-Q4_K_M.gguf \
  --host 0.0.0.0 --port 30000 \
  -ngl 99 --jinja

Split shards: point to the -00001-of-* file, llama.cpp auto-detects the rest.

Ollama

ollama run hf.co/nivvis/Qwen3.5-35B-A3B-EQ-v5-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

With thinking: temp=1.0, top_p=0.95, top_k=20, presence_penalty=1.5
Without thinking: temp=0.7, top_p=0.8, max_tokens=2048

Performance (Q4_K_M, single RTX 5090)

109 t/s generation
653 t/s prompt processing

Other formats

BF16 safetensors — original weights
FP8 — block-wise FP8 for vLLM/SGLang

Lineage

Qwen/Qwen3.5-35B-A3B
  → llmfan46/Qwen3.5-35B-A3B-heretic-v2 (decensored)
    → nivvis/Qwen3.5-35B-A3B-EQ-v5 (DPO for EQ)
      → this repo (GGUF quantizations)