---

base_model: AllSpark-Research/Iris-mini

library_name: gguf

tags:

- gguf

- apex

- custom-quantization

- unsloth-studio

- moe

- speculative-decoding

- mtp

- multi-token-prediction

- llama.cpp

- qwen35moe

license: apache-2.0

language:

- en

- es

- fr

- de

- pt

- it

- ru

- ja

- ko

- vi

- zh

pipeline_tag: text-generation

---

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

Handcrafted Speculative-Decoding Quantization · Native Multi-Token Prediction with Linear AVX2 Speed

Welcome to APEX-I-MiniPlus for AllSpark-Research/Iris-mini (Qwen3.5-MoE 35B language, math, and code architecture featuring native Multi-Token Prediction).

Most automated community releases strip or break the native Multi-Token Prediction head using --no-mtp. APEX-I-MiniPlus fully preserves and calibrates the native prediction block (blk.40), unlocking real-world self-speculative acceleration without an external draft model.

---

<a id="quick-navigation"></a>⚡ Quick Navigation Index

• 📦 Model Files & Specifications
• 🔬 Comparative Quantization Analysis (vs. Flat Quants & Generic APEX)
• ⚡ Native Multi-Token Prediction (MTP) Co-Pilot
• 💻 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)
• 🔬 Why Iris-mini Intentionally Uses Standard APEX (Not V2)
• 🛠️ Handcrafted Layer Architecture
• 📖 Recommended Configuration & Setup

---

<a id="model-specifications"></a>

📦 Model Files & Specifications

| File Name | File Size | Memory Footprint | BPW | Description |

| :--- | :--- | :--- | :--- | :--- |

| Iris-mini.APEX-I-MiniPlus.gguf | 14.84 GB | 13.82 GiB | 3.42 BPW | Handcrafted language, math, reasoning & native MTP draft head |

• Base Architecture: qwen35moe (35B total parameters, approx. 3.2B active per token).
• Speculative Decoding: Fully preserved native Multi-Token Prediction head (blk.40).
• Target Precision: Armored boundaries (Q3_K), calibrated core experts (IQ3_XXS), 6-bit uncompromised output head (Q6_K).

---

<a id="comparative-analysis"></a>

🔬 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) | Q8_0 (8.50 BPW) | Attention Head Stability: Attention gates modulate query-key routing across hybrid attention layers. Keeping them in 8-bit prevents attention crosstalk and hallucination over long contexts. |

| *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. |

| MTP Draft Block (blk.40) | Stripped with --no-mtp or broken | Crushed to IQ2_S / tier precision | Preserved in IQ3_S / Q3_K & IQ4_NL | Speculative Decoding Speedup: Maintains 58%–65% candidate acceptance rate, yielding 1.6x–1.75x real-world token speedup without speculative rejection waste. |

| Normalization & Biases | Often degraded | Standard | F32 uncompressed | Numerical Stability: Prevents cumulative floating-point underflow/overflow across deep 40-layer computation. |

---

<a id="mtp-acceleration"></a>

⚡ Native Multi-Token Prediction (MTP) Co-Pilot

Most automated community releases strip or break the native Multi-Token Prediction head using --no-mtp. APEX-I-MiniPlus fully preserves and calibrates the native prediction block (blk.40):

• Zero-Cost Speculative Acceleration: Unlike external draft models that consume separate VRAM and memory bandwidth, Iris-mini's native MTP head is integrated directly into the weights.
• Empirical Acceptance Rate: 58.8% to 65.5% of predicted candidate tokens are accepted on full GPU offload.
• Token Yield: Delivers 1.60 to 1.75 tokens per forward step on standard text, peaking at 2.0+ tokens/step during continuous code and prose generation.
• Net Speedup: Provides approx. 1.6x faster real-world generation without quality degradation.

---

<a id="laptop-benchmarks"></a>

💻 Everyday Laptop Benchmarks (23–26+ tok/s on DDR4)

• VRAM Allocation: 3.8 GB VRAM utilized on budget 4GB/6GB GPUs.
• System Memory: 32GB DDR4 holds the remaining layers.
• Prefill Speed: 300 to 410 tokens/second sustained across dense inputs.
• Generation Speed: 23 to 26+ tokens/second sustained on standard DDR4 RAM!

---

<a id="context-scaling"></a>

🔥 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 Feasibility |

| :--- | :--- | :--- | :--- | :--- | :--- |

| 32,512 (32k) | 13.82 GiB | 0.58 GiB | 1.80 GiB | 16.20 GiB | Full offload on 24GB; partial on 16GB |

| 64,512 (64k) | 13.82 GiB | 0.92 GiB | 1.95 GiB | 16.69 GiB | Effortless fit on 24GB GPUs |

| 128,640 (128k) | 13.82 GiB | 1.58 GiB | 2.22 GiB | 17.62 GiB | Effortless fit on 24GB GPUs |

| 262,144 (Full 256K)| 13.82 GiB | 2.92 GiB | 2.80 GiB | 19.54 GiB | 🔥 FULL 256K NATIVE IN VRAM! |

---

<a id="throughput-projections"></a>

🏎️ 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) | 110 – 135+ tok/s | 2,500 – 3,600+ tok/s | Blistering speculative execution speed |

| NVIDIA RTX 4090 (24GB GDDR6X) | Full GPU (-ngl 99) | 80 – 105+ tok/s | 1,800 – 2,600+ tok/s | Instantaneous multi-token prediction output |

| NVIDIA RTX 3090 (24GB GDDR6) | Full GPU (-ngl 99) | 66 – 80+ tok/s | 1,400 – 2,000+ tok/s | Full 256k native window in VRAM |

| Consumer Laptop (4GB GPU + DDR4) | Hybrid Offload | 20 – 24+ tok/s | 300 – 420+ tok/s | Smooth streaming from system RAM |

---

<a id="why-standard-apex"></a>

🔬 Why Iris-mini Intentionally Uses Standard APEX (Not V2)

Unlike Occamy and Apodex which use non-linear IQ codebooks, Iris-mini intentionally uses linear AVX2-vectorized Q_K blocks on boundaries:

• Linear Q3_K and Q4_K execute in single-cycle AVX2 instructions without table lookup overhead on CPUs.
• This allows Iris-mini to achieve 23–26+ tok/s on everyday DDR4 laptops, making it the fastest 35B speculative assistant available.

---

<a id="tensor-map"></a>

🛠️ 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 |

| MTP Draft Block (blk.40) | Speculative Head | Q3_K / Q4_K | High candidate acceptance rate |

| Normalization & Biases | All layers | F32 | Prevents cumulative floating point error |

---

<a id="recommended-setup"></a>

📖 Recommended Configuration & Setup

Unsloth Studio:

1. Load Iris-mini.APEX-I-MiniPlus.gguf.
2. Set Speculative Decoding to draft-mtp or set Draft Tokens to 1.
3. Configure KV Cache Dtype to q8_0 and Context Checkpoints to 1.

llama.cpp CLI:

llama-cli -m Iris-mini.APEX-I-MiniPlus.gguf \
  --spec-type draft-mtp \
  -ngl 99 \
  -c 32768