Instructions to use ChiefBroseph/Hy-Memory-30B-A3B-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use ChiefBroseph/Hy-Memory-30B-A3B-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M

Use Docker

docker model run hf.co/ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use ChiefBroseph/Hy-Memory-30B-A3B-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ChiefBroseph/Hy-Memory-30B-A3B-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ChiefBroseph/Hy-Memory-30B-A3B-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M

Ollama
How to use ChiefBroseph/Hy-Memory-30B-A3B-GGUF with Ollama:
```
ollama run hf.co/ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M
```
Unsloth Desktop

How to use ChiefBroseph/Hy-Memory-30B-A3B-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use ChiefBroseph/Hy-Memory-30B-A3B-GGUF with Docker Model Runner:
```
docker model run hf.co/ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M
```

Lemonade

How to use ChiefBroseph/Hy-Memory-30B-A3B-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.Hy-Memory-30B-A3B-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use ChiefBroseph/Hy-Memory-30B-A3B-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use ChiefBroseph/Hy-Memory-30B-A3B-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "ChiefBroseph/Hy-Memory-30B-A3B-GGUF:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Hy-Memory-30B-A3B — 24GB VRAM Optimized Asymmetric Hybrid GGUF (imatrix)

An Asymmetric Hybrid Q4_K_M (Q8_0 / Q6_K / Q4_K_M) GGUF quantization of tencent/Hy-Memory-30B-A3B, calibrated with an importance matrix (imatrix) and tuned specifically for single 24 GB GPUs (RTX 3090 / RTX 4090) as well as zero-VRAM CPU sidecar deployments.

Why Asymmetric Hybrid Quantization for `Hy-Memory-30B-A3B`?

Hy-Memory-30B-A3B (HYV3ForCausalLM / hy_v3 architecture) is a sparse Mixture-of-Experts memory writer model with 30.66B total parameters and ~3.0B active parameters per token (39 MoE layers, 128 routed experts, top-8 routing + 1 shared expert, and 8:1 GQA).

Because 94.5% of the model's parameters (28.97B) live inside the 128 sparse routed experts while only 5.5% (1.69B) make up the always-active backbone, uniform quantization leaves massive quality on the table. Paying just ~0.27 GiB extra keeps the entire always-active backbone at near-lossless Q6_K/Q8_0 precision while fitting comfortably in 24 GB VRAM with room for a 64k–128k q8_0 KV cache:

Component	Tensors	Parameter Share	Quantization	Why
Output / LM Head	`output.weight`	`0.26B` (0.8%)	`Q8_0`	Preserves exact JSON token probabilities, brackets, and schema keys
GQA 8:1 Attention	`attn_qkv.weight`, `attn_output.weight`	`0.50B` (1.6%)	`Q6_K`	8 query heads share 1 KV head across 256k RoPE context; high sensitivity to quantization noise
Shared Experts & Dense Layer 0	`ffn__shexp.weight`, `blk.0.ffn_.weight`	`0.68B` (2.2%)	`Q6_K`	Active on 100% of tokens
Token Embeddings	`token_embd.weight`	`0.26B` (0.8%)	`Q6_K`	Clean input representation across long contexts
MoE Router Gates & Norms	`ffn_gate_inp.weight`, `*_norm.weight`	`0.01B` (<0.1%)	`F32`	Exact top-8 expert routing decisions without router drift
128 Routed Experts (Layers 1–39)	`ffn_gate_up_exps.weight`, `ffn_down_exps.weight`	`28.97B` (94.5%)	`Q4_K_M` (imatrix)	Each expert activates on only 6.25% of tokens; calibrated with a 575-chunk `hy_v3` imatrix

File Summary

File	Quant Tier Breakdown	Size	Avg BPW	24 GB GPU Fit (RTX 3090 / 4090)
`Hy-Memory-30B-A3B-Hybrid-Q4_K_M.gguf`	`Q8_0` head + `Q6_K` attn/shared/embd + `Q4_K_M` (imatrix) experts	17.36 GiB (`17,774 MiB`)	4.96 BPW	Full offload (`-ngl 99`) + up to 128k `q8_0` KV cache (~19.2 GB total VRAM)

Includes Upstream `hy_v3` GGUF Converter Fix (`expert_feed_forward_length = 768`)

In upstream llama.cpp, convert_hf_to_gguf.py (HunyuanMoEModel.set_gguf_parameters in conversion/hunyuan.py) reads moe_intermediate_size (1536) instead of expert_hidden_dim (768) for HYV3ForCausalLM, causing llama.cpp to reject converted GGUFs on load with a tensor shape mismatch (blk.1.ffn_down_exps.weight expected [768, 2048, 128], got [1536, 2048, 128]).

This build was converted with --fuse-gate-up-exps --fuse-qkv and the corrected hy_v3.expert_feed_forward_length = 768 metadata, so it loads cleanly out of the box on llama.cpp (b9993+, verified on b11695).

Importance Matrix (`imatrix`) Calibration

Calibrated across 575 chunks (~1 MB) formatted with native hy_v3 special tokens (<｜hy_begin▁of▁sentence｜>, <｜hy_User｜>, <｜hy_Assistant｜><think>\n\n</think>) covering:

Multi-turn conversational logs and episodic summarization prompts
Structured JSON memory state extraction & schema-constrained updates (add / update / delete / noop)
Technical Markdown documentation and Python/JavaScript source code (preventing MoE router collapse on single-domain text)

Recommended Launch Configurations (`llama-server`)

Important (no_think + temp 0.0): Per Tencent's official Hy-Memory-30B-A3B evaluation setup, run with thinking disabled (no_think) and greedy decoding (--temp 0.0) for deterministic, clean JSON memory operations without <think> trace overhead.

1. Full 24 GB GPU Mode (RTX 3090 / 4090 — 128k Context)

llama-server \
  -m Hy-Memory-30B-A3B-Hybrid-Q4_K_M.gguf \
  -ngl 99 \
  -c 131072 \
  --flash-attn on \
  --cache-type-k q8_0 \
  --cache-type-v q8_0 \
  --jinja \
  --reasoning off \
  --chat-template-kwargs '{"reasoning_effort":"no_think"}' \
  --temp 0.0

2. Zero-VRAM CPU Sidecar Mode (Run Alongside Another Local LLM / ComfyUI)

Because only 3B parameters are active per token, Hy-Memory-30B-A3B runs remarkably fast on system RAM (~15–20 tok/s on modern DDR5 desktops) while using 0 MB of GPU VRAM — ideal as an always-on background memory consolidator alongside a primary GPU model:

CUDA_VISIBLE_DEVICES="" llama-server \
  -m Hy-Memory-30B-A3B-Hybrid-Q4_K_M.gguf \
  -ngl 0 \
  --device none \
  -c 131072 \
  --flash-attn on \
  --cache-type-k q8_0 \
  --cache-type-v q8_0 \
  -t 16 \
  --jinja \
  --reasoning off \
  --chat-template-kwargs '{"reasoning_effort":"no_think"}' \
  --temp 0.0

3. Unsloth Studio "Extra Arguments" Box

Unsloth Studio passes --jinja automatically. Paste the following single line into Extra Arguments:

--reasoning off --chat-template-kwargs {"reasoning_effort":"no_think"} --temp 0.0 --cache-type-k q8_0 --cache-type-v q8_0

Exact `llama-quantize` Recipe Used

llama-quantize \
  --imatrix imatrix.dat \
  --tensor-type "output.weight=q8_0" \
  --tensor-type "token_embd.weight=q6_k" \
  --tensor-type "attn_qkv=q6_k" \
  --tensor-type "attn_output=q6_k" \
  --tensor-type "shexp=q6_k" \
  --tensor-type "blk.0.ffn=q6_k" \
  hy-memory-bf16.gguf \
  Hy-Memory-30B-A3B-Hybrid-Q4_K_M.gguf \
  Q4_K_M

Downloads last month: 144

GGUF

Model size

30B params

Architecture

hy_v3

Hardware compatibility

4-bit

Model tree for ChiefBroseph/Hy-Memory-30B-A3B-GGUF

Base model

tencent/Hy-Memory-30B-A3B

Quantized

(21)

this model

Hy-Memory-30B-A3B — 24GB VRAM Optimized Asymmetric Hybrid GGUF (imatrix)

Why Asymmetric Hybrid Quantization for Hy-Memory-30B-A3B?

File Summary

Includes Upstream hy_v3 GGUF Converter Fix (expert_feed_forward_length = 768)

Importance Matrix (imatrix) Calibration

Recommended Launch Configurations (llama-server)

1. Full 24 GB GPU Mode (RTX 3090 / 4090 — 128k Context)

2. Zero-VRAM CPU Sidecar Mode (Run Alongside Another Local LLM / ComfyUI)

3. Unsloth Studio "Extra Arguments" Box

Exact llama-quantize Recipe Used

Model tree for ChiefBroseph/Hy-Memory-30B-A3B-GGUF

Why Asymmetric Hybrid Quantization for `Hy-Memory-30B-A3B`?

Includes Upstream `hy_v3` GGUF Converter Fix (`expert_feed_forward_length = 768`)

Importance Matrix (`imatrix`) Calibration

Recommended Launch Configurations (`llama-server`)

Exact `llama-quantize` Recipe Used