Instructions to use IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-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 IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-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 IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF
# Run inference directly in the terminal:
llama cli -hf IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF
# Run inference directly in the terminal:
llama cli -hf IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF

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 IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF
# Run inference directly in the terminal:
./llama-cli -hf IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF

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 IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF
# Run inference directly in the terminal:
./build/bin/llama-cli -hf IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF

Use Docker

docker model run hf.co/IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF

LM Studio
Jan

vLLM

How to use IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-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": "IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF

Ollama
How to use IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF with Ollama:
```
ollama run hf.co/IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF
```
Unsloth Desktop

How to use IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF

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": "IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF with Docker Model Runner:
```
docker model run hf.co/IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF
```

Lemonade

How to use IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF

Run and chat with the model

lemonade run user.Iris-mini-MTP-APEX-I-MiniPlus-GGUF-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-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 IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF

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 IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF

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 "IsValorum/Iris-mini-MTP-APEX-I-MiniPlus-GGUF" \
  --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"

Iris-mini APEX-I-MiniPlus (Native MTP) GGUF

Handcrafted Adaptive Quantization · From 4GB Laptops to 24GB Full 256K Context Powerhouses

Welcome to the APEX-I-MiniPlus release of AllSpark-Research/Iris-mini (Qwen3.5-MoE 35B architecture).

Nearly every GGUF on Hugging Face is built by automated batch bots running uniform quant scripts across hundreds of models without looking at tensor architectures. They treat sensitive language heads the same as internal routing weights, squashing critical layers and stripping native speculative decoding.

This release was built differently. APEX-I-MiniPlus is a 100% handcrafted, custom-engineered quantization. Every single tensor was audited, mapped to custom bit-widths, and calibrated against an empirical importance matrix (imatrix) to squeeze maximum intelligence into 13.82 GiB (14.84 GB).

Whether you run on a humble 4GB laptop GPU or a 24GB monster workstation, this build was tuned to extract the absolute best performance possible out of both worlds.

💎 Why Handcrafted APEX Beats Generic Bot Conversions

Automated conversion scripts force a single flat quantization type across all 40 layers. In contrast, our custom APEX-I-MiniPlus architecture applies surgical precision where it matters most:

6-Bit Uncompromised Language Head (output.weight = Q6_K): Bot conversions flatten output heads to 3-bit, causing vocabulary drift across the 248k token space. We hand-assigned Q6_K to ensure crystal-clear token probabilities and zero degradation in complex formatting.
Surgical Core Expert Compression (Layers 10–29 = IQ3_XXS): Instead of degrading the whole model, we concentrated compression on the internal routed experts (3.06 bpw with imatrix), harvesting huge VRAM savings while keeping core reasoning intact.
Armored Semantic Boundaries (Layers 0–9 & 30–39 = Q3_K): The input ingestion layers and final synthesis layers are protected in higher precision, preserving prompt adherence and nuanced instruction-following.
Fully Preserved Native Multi-Token Prediction (blk.40 = Q3_K / Q4_K / Q5_K): Generic releases break or strip the MTP head with --no-mtp. We carefully preserved and calibrated the native prediction block so you can unleash zero-cost speculative decoding.

⚡ The Powerhouse Experience: 24GB GPUs Run Full 256K Context in VRAM!

If you have a 24GB GPU (RTX 3090, RTX 4090, or professional hardware), this is the quantization you have been waiting for.

Standard community 3-bit quants (IQ3_S, Q3_K_S) hover around 15.6 to 15.8 GiB in weights alone. Once you add compute buffers and extended context, memory easily surges past 22.8 GiB, triggering instant CUDA Out Of Memory crashes on Windows displays.

APEX-I-MiniPlus changes the game: At 13.82 GiB, you can run the entire native 262,144 (256K) context window with MTP enabled at just 20.25 GiB total VRAM:

Context Length	Model Weights	KV Cache (q8_0, 4 slots)	Compute Buffers	Draft Cache (MTP)	Total GPU VRAM	Hardware Reality
32,512 (32k)	`13.82 GiB`	`0.57 GiB`	`1.79 GiB`	`0.31 GiB`	`16.50 GiB`	Full offload on 24GB; 38/40 layers on 16GB
64,512 (64k)	`13.82 GiB`	`0.90 GiB`	`1.93 GiB`	`0.37 GiB`	`17.02 GiB`	Effortless full offload on 24GB GPUs
128,640 (128k)	`13.82 GiB`	`1.55 GiB`	`2.20 GiB`	`0.49 GiB`	`18.07 GiB`	Effortless full offload on 24GB GPUs
192,640 (192k)	`13.82 GiB`	`2.20 GiB`	`2.48 GiB`	`0.61 GiB`	`19.11 GiB`	Effortless full offload on 24GB GPUs
262,144 (Full 256K)	`13.82 GiB`	`2.90 GiB`	`2.78 GiB`	`0.75 GiB`	`20.25 GiB`	🔥 ENTIRE 256K NATIVE CONTEXT IN VRAM!

You get the full 256k window, active MTP acceleration, and still keep nearly 4 GiB of breathing room for display drivers and OS tasks!

🚀 The Everyday Miracle: 35B MoE Running on Budget Laptops!

Tested Live by the Creator in Unsloth Studio

You do not need a high-end desktop rig to experience this model. The creator actively uses and tested this build on a budget laptop setup:

GPU VRAM Offload: Uses only 3.8 GB VRAM (fits comfortably on 4GB / 6GB laptop GPUs like the RTX 4050, 3050, or older 1660 Ti/2060).
System Memory: Standard 32 GB DDR4 @ 3200 MHz holds the remaining layers in RAM.
Platform: Unsloth Studio.

Real Measured Benchmark Numbers:

Streaming Generation Speed: 23 to 26+ tokens/second sustained output!
Document Ingestion (Prefill): 300 to 410 tokens/second (digesting 1,300+ prompt tokens in just 3 to 4 seconds).

Creator Pro Tip for Hybrid DDR4 Users (Turn MTP OFF):
If you are offloading partially to RAM with 3.8 GB in VRAM, turn Speculative Decoding (MTP) OFF.
The technical reason: Multi-Token Prediction performs draft validation passes. While GPU VRAM has massive bandwidth (GDDR6 at 500+ GB/s), system DDR4 bandwidth is limited to approx. 25–40 GB/s. Extra verification passes across the DDR4 bus introduce memory latency that slows throughput to approx. 17–19 t/s. Simply by leaving MTP off on DDR4 setups, the model generates pure autoregressive tokens at an astonishing 23 to 26+ tokens/second!

⚡ Native Multi-Token Prediction (MTP) Benchmarks

For users with full GPU offload (24GB GPUs or Apple Silicon), the preserved blk.40.nextn head acts as an internal co-pilot:

Empirical Draft Acceptance Rate: 58.8% to 65.5% of candidate tokens accepted!
Effective Generation Yield: 1.60 to 1.66 tokens per forward step (peaks up to 2.0 tokens/step on continuous prose).
Net Acceleration: Approx. 1.6x faster token streaming over conventional decoding.

🏆 The APEX Lineup: Mini vs. MiniPlus vs. Compact

The APEX quantization family offers custom profiles designed for specific hardware targets:

Metric / Feature	APEX-I-Mini	APEX-I-MiniPlus (This Model)	APEX-I-Compact	Generic Flat Bot (`IQ3_S`)
Quantization Profile	Custom APEX	Custom Handcrafted APEX	Custom APEX	Automated Bot Script
Model Weights Size	approx. 12.2 GiB	13.82 GiB (`14.84 GB`)	approx. 16.1 GiB	approx. 15.6 GiB
Full 256k Context (24GB GPU)	✅ Fits	✅ 20.25 GiB (Stable in Unsloth Studio)	⚠️ Tight (>22.5 GiB)	❌ OOM Crash (>22.8 GiB + OS)
Budget Laptop (4GB VRAM + RAM)	✅ Runs	✅ 23–26+ t/s in Studio	⚠️ Slower in RAM	⚠️ Sluggish & uncalibrated
Output Head Precision	Q6_K	Q6_K (6-Bit Near-Lossless)	Q6_K	3-Bit (Severe lexical drift)
Core Routed Experts (10–29)	IQ2_S (2.5 bpw)	IQ3_XXS (3.06 bpw + imatrix)	Q3_K (3.44 bpw)	3.44 bpw (Uncalibrated flat)
Edge Experts (0–9, 30–39)	Q3_K	Q3_K (Preserved boundaries)	Q4_K (4.5 bpw)	3.44 bpw (Flat)
Native MTP Draft Layer	Preserved	Fully preserved in Q3/Q4/Q5	Preserved	Broken / Stripped (<20%)
Reasoning & Code Fidelity	Good	Superior (>3 bpw threshold)	Maximum	Degraded by flat logits

🛠️ Handcrafted Layer-by-Layer Architecture

Component	Target Layers	Quant Type	Rationale
Output Head (`output.weight`)	Final projection	`Q6_K`	Preserves probability distributions across 248k vocabulary tokens
Token Embeddings	Input projection	`Q3_K`	High semantic input fidelity
Core Routed Experts	Layers 10 to 29	`IQ3_XXS`	Maximum parameter compression (3.06 bpw) with importance matrix guidance
Core Shared Experts	Layers 10 to 29	`Q4_K`	High-precision shared expert routing
Edge Routed Experts	Layers 0 to 9 & 30 to 39	`Q3_K`	Protects prompt ingestion and response synthesis boundaries
Edge Shared Experts	Layers 0 to 9 & 30 to 39	`Q5_K`	Pristine shared pathway fidelity
MTP Prediction Head	Block 40 draft layer	`Q3_K / Q4_K / Q5_K`	High-precision draft prediction for speculative speedup
Norms & State Buffers	Norms, biases, SSM buffers	`F32`	32-bit floating point for absolute numerical stability

📖 Recommended Setup & Best Practices

In Unsloth Studio (Creator-Verified Setup):

All metrics in this card were recorded in Unsloth Studio using the standard Automatic (Auto) mode, tuning only three essential parameters:

Context Checkpoints = 1 (CRITICAL): Set Context Checkpoints to 1. This prevents Unsloth Studio from multiplying KV cache reservations across multiple checkpoints, keeping VRAM flat and rock-solid!
KV Cache Dtype = q8_0: Set KV Cache Dtype to q8_0 (or q4_0 if VRAM is extremely tight) to preserve full attention quality at minimal memory cost.
Context Length Slider: Dial your context to taste (from 32k on laptops up to the full 262,144 on 24GB GPUs!).
Speculative Decoding:
- On Budget Laptops (4GB VRAM / DDR4 RAM): Set to OFF for the verified 23 to 26+ tokens/second!
- On 24GB GPUs / Full Offload: Set to MTP with 1 Draft Token for approx. 1.6x accelerated output.

In llama.cpp / LM Studio / Ollama:

Run with your standard GPU offload configuration. For high-end GPUs, enable native MTP draft decoding (draft count 1); for partial CPU/RAM offload, run standard autoregressive decoding for top throughput.

Original Model & Credits

Original Base Model: AllSpark-Research/Iris-mini
Architecture: qwen35moe (35B Mixture of Experts with Native Multi-Token Prediction)
License: Apache-2.0
Quantization: Custom Handcrafted APEX-I-MiniPlus by IsValorum