Instructions to use IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF with MambaSSM:

from mamba_ssm import MambaLMHeadModel

model = MambaLMHeadModel.from_pretrained("IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF")

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-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/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0
# Run inference directly in the terminal:
llama cli -hf IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0
# Run inference directly in the terminal:
llama cli -hf IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0

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/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0
# Run inference directly in the terminal:
./llama-cli -hf IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0

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/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0
# Run inference directly in the terminal:
./build/bin/llama-cli -hf IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0

Use Docker

docker model run hf.co/IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0

LM Studio
Jan

vLLM

How to use IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-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/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0

Ollama
How to use IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF with Ollama:
```
ollama run hf.co/IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0
```
Unsloth Desktop

How to use IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-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/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0

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/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF with Docker Model Runner:
```
docker model run hf.co/IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0
```

Lemonade

How to use IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0

Run and chat with the model

lemonade run user.Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF-Q8_0

List all available models

lemonade list

Hermes Agent

How to use IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-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/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0

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/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-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/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0

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/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF:Q8_0" \
  --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"

Apodex-1.1-mini APEX-I-MiniPlus-V2 (Multimodal Vision & Native MTP) GGUF

The Definitive 35B Flagship · 256 Fine-Grained Micro-Experts, Hybrid DeltaNet SSM, Vision & Native MTP

Quick Navigation Index:

📦 Model Files & Specifications

👑 The Flagship Architecture: Why V2 is Mandatory for Apodex

🔬 Live Empirical Benchmarks in Unsloth Studio

🎯 1 Draft Token vs. 2 Draft Tokens MTP Engineering Analysis

⚡ 16GB & 24GB VRAM Context Allocation (64k to 256k)

🏎️ Hardware Throughput Projections (RTX 30 / 40 / 50 Series)

🛠️ Surgical Tensor Quantization Map

📖 Recommended Configuration (Unsloth Studio & llama.cpp)

📦 Model Files & Specifications

File Name	File Size	Memory Footprint	BPW	Description
`Apodex-1.1-mini.APEX-I-MiniPlus-V2.gguf`	`15.01 GB`	`13.98 GiB`	3.38 BPW	Handcrafted language, math, DeltaNet SSM & native MTP draft head
`mmproj-apodex_Apodex-1.1-mini-Q8_0.gguf`	`610 MB`	`582 MiB`	8.50 BPW	Dedicated `Q8_0` vision projector for document OCR and image input

Base Architecture: qwen35moe (35B total parameters, fine-grained MoE).
Micro-Expert Granularity: 256 fine-grained micro-experts with intermediate dimension of 512.
Hybrid Attention: 30 DeltaNet SSM / Linear Attention layers + 10 full attention layers.
Speculative Decoding: Fully preserved native Multi-Token Prediction draft block (blk.40).

👑 The Flagship Architecture: Why V2 is Mandatory for Apodex

Apodex-1.1-mini combines every cutting-edge architectural advance in open-source AI, making it the most complex model in the lineup:

The 256 Micro-Expert Trap: Standard linear quants (Q3_K, Q4_K) introduce rounding drift that disrupts routing logits across 256 narrow micro-experts (intermediate dimension: 512). APEX-I-MiniPlus-V2 solves this by deploying non-linear IQ codebooks (IQ3_S, IQ3_XXS, IQ4_NL) calibrated against the importance matrix, preventing routing collapse.
DeltaNet SSM State Stability: With 30 linear recurrent layers, weight precision directly impacts state drift over long contexts. Critical projection and recurrence weights are preserved in non-linear IQ3_S and F32 normalization buffers.
Uncompromised 6-Bit Output Head (output.weight = Q6_K): Generic bot quants crush the vocabulary head to 3 bits, causing hallucination across the 248k vocabulary. We hand-assigned Q6_K to keep output token probabilities pristine.
Preserved Native MTP & Vision: Both the native Multi-Token Prediction draft block (blk.40) and the high-precision Q8_0 vision projector are fully preserved and bundled.

🔬 Live Empirical Benchmarks in Unsloth Studio

Tested by the Creator on Consumer Hardware (Intel Core i5-12450H, 4GB Laptop GPU, 32GB DDR4)

Lightning Ingestion (Prefill): 431.53 tokens/second sustained across dense text! Ingested an entire academic thesis of 24,870 tokens in just 57.63 seconds (peaking at 452.30 tokens/second).
Streaming Generation Speed: 18.5 to 21 tokens/second in standard context (settles at 14.5 to 18.5 tok/s even with 25,000+ tokens loaded in DDR4 RAM).
Native MTP Draft Acceptance: 57.55% to 59.66% empirical acceptance rate (yielding 2.19 tokens per forward step).
Linear Attention Advantage: DeltaNet SSM layers keep KV cache memory growth almost flat, eliminating quadratic context bottlenecks.

🎯 1 Draft Token vs. 2 Draft Tokens MTP Engineering Analysis

Configuration	First Token Acceptance	Effective Token Yield	Draft Cache VRAM (256k)	Total VRAM at Full 256K	Best Hardware Target
1 Draft Token (`--spec-draft-n-max 1`)	~68% to 75%	1.68 to 1.75 tok/step	~0.38 GiB (Saves ~0.40 GiB)	`18.06 GiB`	16GB GPUs (at 64k) & 24GB GPUs
2 Draft Tokens (`--spec-draft-n-max 2`)	~59.7% combined	2.19 tok/step	~0.78 GiB	`18.46 GiB`	Workstations & 24GB GPUs (RTX 3090/4090/5090)

With 1 Draft Token: Captures the highest-confidence prediction candidate, saving 0.20 to 0.40 GiB VRAM and reducing memory bus verification latency on DDR4 laptops.
With 2 Draft Tokens: Unlocks maximum raw token output (2.19 tokens/step) on dedicated high-bandwidth GPUs.

⚡ 16GB & 24GB VRAM Context Allocation (64k to 256k)

Empirically Verified in Unsloth Studio by the Creator

At 13.98 GiB, APEX-I-MiniPlus-V2 fits into 16GB GPUs at 64k context and fits the full native 256k context into 24GB GPUs:

Context Length	KV Cache (`q8_0`)	MTP Setting	Vision (`Q8_0`)	Total GPU VRAM	Hardware Verdict
64,512 (64k)	`q8_0`	Draft 1 Token	Disabled	`15.05 GiB`	🔥 Ultra-Safe fit on 16GB GPUs (RTX 4080 / 5080 16GB)
64,512 (64k)	`q8_0`	Draft 2 Tokens	Disabled	`15.24 GiB`	Full offload on 16GB GPUs
64,512 (64k)	`q8_0`	Draft 2 Tokens	Enabled	`16.03 GiB`	Full offload on 16GB / 24GB GPUs
128,512 (128k)	`q8_0`	Draft 1 Token	Disabled	`16.03 GiB`	Effortless fit on 24GB GPUs
128,512 (128k)	`q8_0`	Draft 2 Tokens	Disabled	`16.28 GiB`	Effortless fit on 24GB GPUs
192,512 (192k)	`q8_0`	Draft 1 Token	Disabled	`17.01 GiB`	Effortless fit on 24GB GPUs
192,512 (192k)	`q8_0`	Draft 2 Tokens	Disabled	`17.33 GiB`	Effortless fit on 24GB GPUs
262,144 (Full 256K)	`q8_0`	Draft 1 Token	Disabled	`18.06 GiB`	🔥 FULL 256K IN VRAM (Saves ~0.40 GiB VRAM!)
262,144 (Full 256K)	`q8_0`	Draft 2 Tokens	Disabled	`18.46 GiB`	🔥 FULL 256K IN VRAM! (~5.5 GiB headroom left)
262,144 (Full 256K)	`q8_0`	Draft 2 Tokens	Enabled	`19.25 GiB`	🔥 Full 256K + Vision in 24GB VRAM! (~4.75 GiB headroom)

🏎️ Hardware Throughput Projections (RTX 30 / 40 / 50 Series)

GPU / Hardware Setup	VRAM & Bandwidth	Standard Speed	Generation with MTP Co-Pilot	Prompt Prefill Speed
NVIDIA RTX 5080 / 5090 (Blackwell)	24GB / 32GB GDDR7 (>1,500 GB/s)	115 to 140+ tok/s	🔥 180 to 220+ tok/s!	2,500 to 3,600+ tok/s
NVIDIA RTX 4090	24GB GDDR6X (1,008 GB/s)	82 to 105 tok/s	🔥 135 to 165+ tok/s!	1,900 to 2,800+ tok/s
NVIDIA RTX 4080 (16GB) (up to 64k)	16GB GDDR6X (717 GB/s)	75 to 88 tok/s	🔥 120 to 145+ tok/s!	1,700 to 2,300+ tok/s
NVIDIA RTX 3090 / 3090 Ti	24GB GDDR6 (936 GB/s)	68 to 80 tok/s	🔥 110 to 135+ tok/s!	1,500 to 2,100+ tok/s
Everyday Laptop (RTX + DDR4 RAM)	3.8GB VRAM + DDR4 System RAM	18.5 to 21 tok/s	14.5 to 18.5 tok/s	431 to 452 tok/s (Empirical)

🛠️ Surgical Tensor Quantization Map

Tensor Pattern	Target Layer Scope	Quant Type	BPW	Rationale
`output.weight`	Final projection	`Q6_K`	~6.56	Zero lexical degradation across 248,320 vocabulary tokens
`token_embd`	Input projection	`IQ3_S`	3.44	Non-linear codebook guided by imatrix; protects semantic input space
`ffn_gate_inp`	Expert Routers	`F32`	32.0	Uncompressed 32-bit floating point; 100% exact expert selection without routing drift
`attn_gate`	Attention Gates	`Q8_0`	8.50	Near-lossless 8-bit precision for attention modulation
`attn_output`	Attention Out	`Q6_K`	~6.56	High-precision 6-bit attention projection across all 40 layers
`attn_qkv`, SSM Out	QKV & DeltaNet	`IQ3_S`	3.44	High-fidelity attention heads and recurrent state projections
`ffn_down/up/gate_exps`	Edges (0–9, 30–39 & 40)	`IQ3_S`	3.44	Preserves boundary ingestion and response synthesis
`ffn_down/up/gate_exps`	Core (10–29, 256 MoE)	`IQ3_XXS`	3.06	Maximum compression while strictly respecting >3 BPW quality barrier
*`ffn__shexp`**	Shared Experts	`IQ4_NL`	4.50	Non-linear 4-bit precision for the pathway that processes 100% of tokens
`blk.40.nextn.eh_proj`	MTP Projection	`IQ4_NL`	4.50	Dedicated 4-bit non-linear precision for native draft token prediction
Norms, Biases & States	Norms, conv1d, biases	`F32`	32.0	Pristine numerical stability across all recurrent and attention steps

📖 Recommended Configuration (Unsloth Studio & llama.cpp)

Unsloth Studio:

Load Apodex-1.1-mini.APEX-I-MiniPlus-V2.gguf.
Select mmproj-apodex_Apodex-1.1-mini-Q8_0.gguf as the vision projector.
Configure KV Cache Dtype to q8_0 and Context Checkpoints to 1.
Set Speculative Decoding to MTP (Draft Tokens: 1 or 2).

LM Studio / Ollama / llama.cpp:

llama-cli -m Apodex-1.1-mini.APEX-I-MiniPlus-V2.gguf --mmproj mmproj-apodex_Apodex-1.1-mini-Q8_0.gguf --draft-mtp -c 64512 -ngl 99

Credits

Base Model: apodex/Apodex-1.1-mini
Quantization: Custom Handcrafted APEX-I-MiniPlus-V2 by IsValorum

Downloads last month: 225

GGUF

Model size

36B params

Architecture

qwen35moe

Hardware compatibility

We're not able to determine the quantization variants.

View all variants

Model tree for IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF

Base model

Qwen/Qwen3.5-35B-A3B-Base

Finetuned

Qwen/Qwen3.5-35B-A3B

Finetuned

apodex/Apodex-1.1-mini

Quantized

(22)

this model

Collection including IsValorum/Apodex-1.1-mini-APEX-I-MiniPlus-V2-GGUF

APEX-I Custom Quants

Collection

Handcrafted APEX-I GGUFs optimized for 4GB-24GB VRAM setups with high throughput. • 4 items • Updated about 19 hours ago