robertbevanofhermes-qwen3.5-35b-a3b

Note: These models are optimized for use within an agentic harness (e.g. Hermes Agent) and may behave unexpectedly in raw inference without a system prompt. Capability benchmarks are strong but conversational behavior outside of a structured harness is not reliable. I am currently working on v2 to address this and reduce harness dependency.

<div align="center"> <h3>Support This Work</h3> <p> I'm a PhD student who also happens to spend way too much time fine-tuning, merging, and quantizing open-weight models on rented H100s and a local DGX Spark. It's a hobby that got out of hand. <br><br> If my uploads have been useful to you, consider buying a PhD student a coffee. It goes a long way toward keeping these experiments running. </p> <p><a href="https://ko-fi.com/djlougen">&#9749; ko-fi.com/djlougen</a></p> </div>

---

Hermes Qwen3.5 35B-A3B GGUF

GGUF quantizations of a Qwen3.5-35B-A3B model fine-tuned on NousResearch/hermes-function-calling-v1 for structured function calling and tool use.

Base Model

• Architecture: Qwen3.5 MoE (Mixture of Experts) — 35B total parameters, ~3B active per token
• Base: Qwen/Qwen3.5-35B-A3B
• Context Length: 262,144 tokens
• Experts: 256 total, 8 active per token

Fine-Tuning Details

• Method: LoRA via Unsloth + TRL SFTTrainer
• LoRA Rank (r): 32
• LoRA Alpha: 32
• LoRA Dropout: 0
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Training Precision: bf16
• Optimizer: AdamW 8-bit
• Learning Rate: 2e-4 with cosine scheduler
• Warmup Steps: 10
• Epochs: 3
• Batch Size: 2 per device, 8 gradient accumulation steps (effective batch size 16)
• Max Sequence Length: 4,096 tokens
• Weight Decay: 0.01
• PEFT Version: 0.18.1

Training Dataset

NousResearch/hermes-function-calling-v1 — a function-calling dataset following the Hermes Function-calling Standard. Includes:

• Cleaned Glaive Function Calling samples
• Advanced JSON structured output (agentic, multi-turn)
• Single-turn JSON structured output samples

Conversations were formatted using ChatML (<|im_start|> / <|im_end|>) with role mapping: system, human -> user, gpt -> assistant, tool.

Quantization

All quantizations were produced using llama.cpp with an importance matrix (imatrix) computed from WikiText-2 calibration data for improved quality at lower bit depths.

Available Quants

All quantizations verified: 733 tensors, GGUF v3.

Choosing a Quant

• Q8_0 (36.9 GB): Closest to full precision. Use if you have the VRAM/RAM.
• Q6_K / Q5_K_M (28.5 / 24.7 GB): Good balance of quality and size for most use cases.
• Q4_K_M (21.2 GB): Popular sweet spot — significant size reduction with minimal quality loss.
• IQ4_NL / IQ4_XS (19.8 / 18.7 GB): Importance-matrix 4-bit — can outperform standard Q4 quants at similar size.
• IQ3_M / IQ3_S (15.4 / 15.3 GB): Importance-matrix 3-bit — good quality for the size with imatrix calibration.
• IQ2_M and below (11.7 GB and smaller): Extreme compression with imatrix. Quality degrades progressively.
• IQ1_M / IQ1_S (8.2 / 7.5 GB): Maximum compression. Expect significant quality loss.
• IQ3_M and below: For constrained environments. Quality degrades progressively.
• IQ2 / IQ1: Extreme compression. Expect notable quality degradation.

Usage

llama.cpp

llama-cli -m hermes-qwen3.5-35b-a3b-Q4_K_M.gguf -p "You are a helpful assistant." -cnv

LM Studio / Ollama / KoboldCpp

Download any GGUF file and load it directly.

Credits

• Base Model: Qwen Team
• Training Dataset: NousResearch
• Fine-Tuning Framework: Unsloth
• Quantization Tooling: llama.cpp