Quick Navigation Index

Independently computed using dedicated cloud compute clusters, funded 100% out-of-pocket.

If you consider these releases to be of high quality and they help you in your day-to-day workflow, consider supporting on Ko-fi. Every contribution directly sustains future cloud compute hours and helps support my university studies.

1. APEX MLX Variants
2. Fidelity Verification
3. Quality Spectrum: APEX MLX MiniPlus vs. NanoPlus
4. Model Files & Technical Specifications
5. Surgical Tensor Quantization Map
6. Apple Silicon Unified Memory & Context
7. Recommended Configuration & Setup
8. Recommended Generation Parameters
9. CRITICAL: Coding Syntax & Repeat Penalty Advisory
10. Native Chat Template
11. 7. Community Compute Fund & Student Support

Qwen3.8-35B-A3B-Distill MLX APEX-MiniPlus Abliterated (Uncensored)

Q5-Class Frontier MoE · Native MLX Mixed Precision · Apple Silicon · Under 20 GB

[!WARNING]

Known Upstream Long-Generation Limitation

The upstream Qwen3.8-35B-A3B-Distill checkpoint can repeat during very long responses, including after the <think> block. This behavior originates upstream and is not specific to the MLX APEX quantization.

[!NOTE]

🪶 LOOKING FOR THE SMALLER MLX RELEASE?

If footprint matters more than the extra precision budget, use the sister NanoPlus release: 🔗 Qwen3.8-35B-A3B-Distill MLX APEX-NanoPlus Abliterated — Q4-class practical quality at 18.82 GB of Safetensors weights.

[!TIP]

🔓 ABLITERATED / UNCENSORED MLX EDITION

This is the Abliterated / Uncensored MLX APEX-MiniPlus edition of Qwen3.8-35B-A3B-Distill. It is the higher-fidelity member of the two-model MLX APEX family and is intended for users who deliberately prefer reduced refusal behavior.

[!IMPORTANT]

THE UNDER-20-GB SPECIFICATION: Q5-CLASS PRACTICAL QUALITY

MLX APEX-MiniPlus uses a native 4-bit G64 MLX base with carefully protected FP32 paths and higher-precision 5/6/8-bit components. The published Safetensors weights occupy 19.9485 GB (18.58 GiB) with an average quantization of 4.604 bpw, targeting Q5-class practical quality without crossing the 20 GB weight ceiling.

[!TIP]

🏆 THE MLX MINIPLUS SWEET SPOT: Q5-CLASS QUALITY BELOW 20 GB

The extra footprint over NanoPlus is concentrated in always-active and sensitivity-critical paths:

• LM head: 6-bit G64.
• Shared experts: 8-bit G64.
• Linear-attention in_proj_z: 8-bit G64.
• Norms, MoE routers and recurrent state/control tensors: protected in FP32.
• Linear-attention in_proj_b and out_proj: 5-bit G64.
• All routed experts: 4-bit G64 across all 40 layers.

[!TIP]

🏆 FIDELITY RESULT

Release	Weight Size	Quality Class	ΔPPL	%PPL
MLX APEX-MiniPlus Abliterated (CURRENT)	19.9485 GB	Q5-class	~+0.42	~+4.8%

[!WARNING]

DO NOT CONFUSE APEX MLX MINIPLUS / NANOPLUS WITH GENERIC APEX OR APEX-I GGUF RELEASES!

The MLX APEX family is a separate native MLX mixed-precision specification. It is not a renamed GGUF APEX-I-Mini, APEX-I-MiniPlus, K-quant or IQ-quant.

• Generic / GGUF APEX releases: use llama.cpp/GGUF quantization types such as Q*_K and IQ*, with their own tensor layouts, kernels and bit accounting.
• MLX APEX NanoPlus / MiniPlus: use MLX affine quantization with per-module bit width and group-size overrides, plus protected FP32 paths. There are no IQ quants here, which is why these releases use APEX, not APEX-I.
• Quality-class labels: Q4-class and Q5-class describe practical fidelity positioning; they do not claim that the checkpoints are uniform flat 4-bit or 5-bit models.

[!TIP]

APPLE SILICON UNIFIED MEMORY

MLX uses Apple Silicon unified memory rather than the discrete GPU-VRAM/system-RAM split used by CUDA + llama.cpp. MiniPlus weights alone occupy about 18.58 GiB; runtime state, activations, macOS and the active context require additional unified memory. Context headroom therefore depends on the Mac and its available memory at inference time.

---

APEX MLX Variants

MiniPlus uses the larger precision budget for higher practical fidelity. NanoPlus remains the compact alternative for users who need to save roughly 1.13 GB of weight storage and unified-memory footprint.

---

Fidelity Verification

• ΔPPL: ~+0.42
• %PPL: ~+4.8%

Quality Spectrum: APEX MLX MiniPlus vs. NanoPlus

The quality tiers are practical positioning labels across two different MLX mixed-precision maps, not claims of uniform flat-bit equivalence.

---

Model Files & Technical Specifications

• Base Model: empero-ai/Qwen3.8-35B-A3B-Distill
• Architecture: Qwen3.8 / Qwen3.5-MoE family, 40 layers, 256 micro-experts with hybrid linear attention / recurrent state and periodic full-attention layers
• Runtime: MLX / MLX-LM on Apple Silicon
• Context Length: 262,144 tokens (native 256K); usable context is additionally constrained by available unified memory

---

Surgical Tensor Quantization Map

This is the MLX-native mixed-precision map represented by the published checkpoint:

Published quantized-module counts: 181 × 4-bit G64, 60 × 5-bit G64, 11 × 6-bit G64, 150 × 8-bit G64, plus 110 protected FP32 modules.

---

Apple Silicon Unified Memory & Context

Unlike a GGUF CUDA deployment, MLX does not split the model into -ngl GPU layers and system-RAM offload. The CPU and GPU share the same unified-memory pool.

No Mac throughput figure is claimed here without a corresponding measured Apple-Silicon result.

---

Recommended Configuration & Setup

pip install -U mlx-lm

mlx_lm.generate \
  --model IsValorum/Qwen3.8-35B-A3B-Distill-MLX-APEX-MiniPlus-Abliterated \
  --prompt "Hello"

Server:

mlx_lm.server --model IsValorum/Qwen3.8-35B-A3B-Distill-MLX-APEX-MiniPlus-Abliterated

⚙️ Recommended Generation Parameters (empero-ai Official)

Official generation guidance used by empero-ai/Qwen3.8-35B-A3B-Distill:

[!IMPORTANT]

🔍 Model Behavior vs. Quantization Fidelity

Sampling, prompting and upstream model behavior can materially affect the output. The Q5-class label describes this release's practical fidelity target; PPL is only one signal of quantization fidelity and does not guarantee identical real-world behavior to BF16.

[!IMPORTANT]

CRITICAL ADVISORY FOR CODING WORKFLOWS: PREVENTING SYNTAX & TOKEN SWAPPING

Programming code necessarily repeats brackets, operators and indentation. A repeat penalty above 1.0 can suppress legitimate syntax tokens and cause character substitutions.

Recommended coding profile:

• repeat_penalty: 1.0
• presence_penalty: 0.0
• frequency_penalty: 0.0
• temperature: 0.60 (or lower for stricter deterministic code)
• top_p: 0.95
• top_k: 20

[!TIP]

NATIVE CHAT TEMPLATE

chat_template.jinja and the tokenizer configuration are included in this repository. Use the repository's native template with MLX-LM so system, user, assistant, reasoning and tool-call formatting follow the model's expected control-token structure.

7. Community Compute Fund & Student Support

All IsValorum quantizations will always remain completely free and open to the public without paywalls.

However, cloud GPU compute is expensive and funded entirely out-of-pocket. If you find these handcrafted builds to be of high quality and they benefit your day-to-day workflow, you can support future Day-0 releases on Ko-fi. Every contribution directly funds the cloud compute hours required for new model runs and helps support my university education:

(When supporting on Ko-fi, feel free to leave a note with your Hugging Face handle and the specific model you would like prioritized).