[!NOTE]
OPTIMIZATION HISTORY — LEGACY EDITION
This repository hosts a previous iteration of our handcrafted MiniPlus architecture. While not our current specification, it remains an outstanding, high-fidelity quantization that significantly outperforms any flat 3-bit community quants (Q3_K_S / IQ3_S) and generic 2-bit APEX Mini community releases.
We preserve this repository publicly with 100% transparency as a verified engineering record of continuous optimization within the strict 13–14 GB envelope.
Current Definitive Specification (V2.1): Access the newly upgraded V2.1 release featuring zero AVX2 CPU stalls and maximum long-context stability directly at:
IsValorum/Nex-N2.5-mini-APEX-I-MiniPlus-V2.1-GGUF
Quick Navigation Index
- Model Files & Specifications
- Comparative Quantization Analysis (vs. Flat Quants & Generic APEX)
- Bundled Q8_0 High-Precision Vision Projector
- Everyday Laptop Benchmarks (23–26+ tok/s on DDR4)
- The 24GB Miracle: Full 256K Context Runs In VRAM!
- Hardware Throughput Projections (RTX 30 / 40 / 50)
- Handcrafted Layer Architecture
- Recommended Configuration & Setup
Model Files & Specifications
| File Name |
File Size |
Memory Footprint |
BPW |
Description |
Nex-N2.5-mini.APEX-I-MiniPlus.gguf |
14.56 GB (13.56 GiB) |
13.56 GiB |
3.36 BPW |
Main language, reasoning, tool-use & computer-use model |
mmproj-nex-agi_Nex-N2.5-mini-Q8_0.gguf |
610 MB (582 MiB) |
582 MiB |
8.50 BPW |
Dedicated Q8_0 vision projector for GUI parsing & high-res image input |
- Base Architecture:
qwen35moe (35.1B parameters, multimodal agentic MoE).
- Core Strengths: Autonomous computer-use, function calling, JSON schema compliance, high-resolution visual grounding.
Comparative Quantization Analysis (vs. Flat Quants & Generic APEX)
Also, don't confuse APEX-I-MiniPlus (Standard) with a generic baseline APEX-I-Mini. Traditional APEX-I-Mini drops core experts aggressively to 2-bit IQ2_S and leaves output.weight at 3-bit Q3_K_M, which creates a noticeable perplexity hit on complex reasoning tasks. Standard MiniPlus avoids that degradation floor while keeping boundary layers in linear Q3_K for single-cycle vectorized AVX2 CPU dequantization (hitting 23 to 26+ tok/s on DDR4 laptops), while protecting output in Q6_K and routers in F32.
To put the numbers in perspective: this cuts nearly 2 GB off a flat 3-bit quant (approx. 15.6 GB), and weighs only about approx. 1 GB more than a generic APEX-I-Mini (approx. 12.5 GB). For that single extra gigabyte of VRAM, you get a massive jump in reasoning and syntactic stability while maximizing CPU/RAM execution throughput.
Take a look at the tensor-by-tensor comparison table below to inspect the exact architectural differences and see why this specific allocation is optimal. That's specifically what this was built for:
| Architectural Component |
Generic Automated Quants (Flat Q3_K_S / IQ3_S) |
Generic APEX-I-Mini (Baseline Recipe) |
Our Handcrafted APEX-I-MiniPlus (Standard / IsValorum) |
Perceived Quality & Real-World Impact |
Output Head (output.weight) |
Flat IQ3_S / Q3_K_S (approx. 3.44 BPW) |
Inherits base type Q3_K_M (approx. 3.44 BPW unarmored) |
Q6_K (approx. 6.56 BPW uncompromised) |
Eliminates Syntax & Vocabulary Hallucinations: Low-bit output heads cause tokenizer classification noise, breaking code indentation, brackets ({}, []), math symbols, and domain terms. Q6_K preserves near-FP16 output classification. |
Expert Routers (ffn_gate_inp.weight) |
Blindly quantized to 3-bit / unoptimized |
Inherits base type Q3_K_M (approx. 3.44 BPW compressed) |
F32 uncompressed (32.0 BPW, 2 MB/layer) |
Zero Router Drift: In micro-expert models, even minuscule quantization errors in router logits misdirect tokens to wrong experts. Retaining uncompressed F32 guarantees 100% routing fidelity with virtually zero memory overhead (approx. 80 MB total). |
Attention & Language (attn_output, attn_qkv) |
Flat IQ3_S / Q3_K_S |
Q3_K on 34 middle layers (L3–36), Q4_K on 6 edge layers |
Q6_K for attn_output, Q3_K / Q4_K + imatrix |
Contextual Precision & CPU Throughput: Combines uncompromised Q6_K for the output projection with fast vectorized linear blocks for attention, balancing retrieval accuracy with maximum token streaming speed on CPU/RAM. |
Attention Gates (attn_gate.weight) |
Blindly compressed to 3-bit |
Compressed to Q3_K (middle) / Q4_K (edges) |
Q4_K / Q8_0 (linear high-precision) |
Attention Routing Dynamics: High-precision linear gating modulating query-key projections without CPU dequantization latency. |
Shared Foundation Expert (ffn_*_shexp) |
Flat IQ3_S / Q3_K_S (3.44 BPW) |
Linear Q4_K (middle) / Q5_K (edges) |
Linear Q4_K (middle) / Q5_K (edges) + imatrix |
Foundational Knowledge Stability: Keeps the universal pathway in high-fidelity linear blocks, eliminating quantization drift while maintaining rapid single-cycle dequantization. |
| Core MoE Layers (Middle: 10–29) |
Flat IQ3_S / Q3_K_S (uniform bit-rate across all layers) |
Aggressive IQ2_S (2.50 BPW) |
IQ3_XXS (3.06 BPW) + calibrated imatrix |
Above the Quality Threshold: Generic 2-bit IQ2_S baselines drop below the critical quality floor for 35B MoEs, resulting in perplexity spikes on reasoning tasks. Our IQ3_XXS with imatrix achieves deep compression (272 MiB → 98 MiB per block) without sacrificing logic. |
| Edge MoE Layers (Layers 0–9 & 30–39) |
Flat IQ3_S / Q3_K_S (no layer-wise gradient) |
Q3_K (limited to first/last 5 layers only: L0–4, L35–39) |
Q3_K (expanded to 10 input & 10 output layers) |
AVX2 Single-Cycle Speed: Expanded 10+10 layer protection using linear Q3_K blocks enables single-cycle vectorized AVX2 CPU dequantization, unlocking 23 to 26+ tok/s on budget DDR4 laptops. |
Multimodal Vision (mmproj) |
Often omitted, or left as uncompressed FP16 (approx. 900 MB) |
Often omitted or separate uncompressed FP16 |
Bundled Q8_0 (582 MB) with 27 critical F32/F16 fallbacks |
Saves approx. 320 MB VRAM with Zero Loss: Handcrafted quantization preserves normalization and bias tensors in F32/F16, ensuring razor-sharp OCR, DOM viewport reading, and coordinate detection without visual noise. |
| Normalization & Biases |
Often degraded |
Standard |
F32 uncompressed |
Numerical Stability: Prevents cumulative floating-point underflow/overflow across deep 40-layer computation. |
Bundled Q8_0 High-Precision Vision Projector
Unlike text-only MoEs, Nex-N2.5-mini is designed for computer use, visual grounding, and multi-modal interaction.
- Rather than leaving users to search for external FP16 projectors (approx. 900 MB), this repository bundles the official projector quantized to Q8_0 (610 MB / 582 MiB).
- Delivers near-lossless visual recognition while saving VRAM.
Everyday Laptop Benchmarks (23–26+ tok/s on DDR4)
Empirically Verified in Unsloth Studio
- GPU VRAM Offload: Uses only 3.8 GB VRAM (fits effortlessly on budget 4GB and 6GB laptop GPUs like the RTX 4050, 3050, or older 1660 Ti/2060).
- System Memory: Standard 32 GB DDR4 @ 3200 MHz holds the rest of the model.
- Estimated Generation Speed: 23 to 26+ tokens/second sustained output!
- Estimated Document Ingestion (Prefill): 300 to 410+ tokens/second.
The 24GB Miracle: Full 256K Context Runs In VRAM!
| Context Length |
Model Weights (Est.) |
KV Cache (q8_0, 4 slots) |
Compute Buffers |
Total GPU VRAM (Est.) |
Hardware Verdict |
| 32,512 (32k) |
13.56 GiB |
0.57 GiB |
1.79 GiB |
15.92 GiB |
Full offload on 24GB; 38/40 layers on 16GB |
| 64,512 (64k) |
13.56 GiB |
0.90 GiB |
1.93 GiB |
16.39 GiB |
Effortless fit on 24GB GPUs |
| 128,640 (128k) |
13.56 GiB |
1.55 GiB |
2.20 GiB |
17.31 GiB |
Effortless fit on 24GB GPUs |
| 262,144 (Full 256K) |
13.56 GiB |
2.90 GiB |
2.78 GiB |
19.24 GiB |
FULL 256K NATIVE CONTEXT IN VRAM! |
Hardware Throughput Projections (RTX 30 / 40 / 50)
| Hardware Target |
Offload Mode |
Generation Speed (Est.) |
Prompt Prefill Speed (Est.) |
Highlights |
| NVIDIA RTX 5080 / 5090 (Blackwell) |
Full GPU (-ngl 99) + mmproj |
105 – 130+ tok/s |
2,400 – 3,500+ tok/s |
Blistering agentic GUI interaction throughput |
| NVIDIA RTX 4090 (24GB GDDR6X) |
Full GPU (-ngl 99) + mmproj |
75 – 100+ tok/s |
1,700 – 2,500+ tok/s |
Real-time computer-use screen analysis & tool calling |
| NVIDIA RTX 3090 (24GB GDDR6) |
Full GPU (-ngl 99) + mmproj |
62 – 78+ tok/s |
1,350 – 1,950+ tok/s |
Full 256k multi-modal context in dedicated VRAM |
| Consumer Laptop (4GB GPU + 32GB RAM) |
Hybrid Offload |
20 – 24+ tok/s |
300 – 420+ tok/s |
Smooth streaming from system DDR4/DDR5 RAM |
Handcrafted 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 |
Expert Routers (ffn_gate_inp) |
All layers (0–39) |
F32 |
Uncompressed 32-bit floating point; 100% exact expert selection without routing noise |
Attention Output (attn_output) |
All layers |
Q6_K |
Uncompromised 6-bit attention projection across all layers |
| Attention QKV & SSM States |
All layers |
Q3_K / Q4_K |
Fast vectorized AVX2 linear dequantization for tool-use responsiveness |
| 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 |
Q4_K |
Armors foundational reasoning |
| Normalization & Biases |
All layers |
F32 |
Prevents cumulative floating point error |
Vision Projector (mmproj) |
Visual adapter |
Q8_0 |
Ultra-high fidelity visual comprehension without FP16 bloat |
Recommended Configuration & Setup
Unsloth Studio:
- Load
Nex-N2.5-mini.APEX-I-MiniPlus.gguf.
- Select
mmproj-nex-agi_Nex-N2.5-mini-Q8_0.gguf as the vision projector.
- Configure KV Cache Dtype to
q8_0 and Context Checkpoints to 1.
- Set GPU Offload to 100% (
-ngl 99) on 24GB GPUs.
llama.cpp CLI:
llama-cli -m Nex-N2.5-mini.APEX-I-MiniPlus.gguf \
--mmproj mmproj-nex-agi_Nex-N2.5-mini-Q8_0.gguf \
-ngl 99 \
-c 32768
