jashepp/Equinox-31B-i1-MXFP4-GGUF overview

---

license: apache-2.0

language:

• en

library_name: transformers

base_model_relation: quantized

pipeline_tag: text-generation

base_model:

• LatitudeGames/Equinox-31B
• bartowski/LatitudeGames_Equinox-31B-GGUF
• mradermacher/Equinox-31B-i1-GGUF

tags:

• gemma4
• text adventure
• roleplay
• imatrix
• GGUF
• conversational
• mxfp4
• quantized
• creative-writing

---

💎 Equinox-31B - Custom Mixed Precision GGUFs with Imatrix

> Equinox draws its name from the balance between extremes. Trained on a balanced blend of Wayfarer 2's unforgiving dark adventures and Hearthfire's quiet slice-of-life storytelling, Equinox is equally at home in perilous dungeons and candlelit conversations.\

> Equinox 31B was trained with two epochs of SFT (Supervised Fine-Tuning) on top of Gemma 4 31B Instruct.




This repository contains custom, highly optimized, multi-tier mixed precision GGUF weights for LatitudeGames/Equinox-31B.

📄 GGUF Files

In order of quality:

| Filename | Size | Quants |

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

| Equinox-31B-MXFP4_Q8_0_F16-Imatrix.gguf | 22.9 GB | MXFP4_MOE + Q8_0 + F16 |

| Equinox-31B-MXFP4_Q6_K_Q8_0-Imatrix.gguf | 19.5 GB | MXFP4_MOE + Q6_K + Q8_0 |

| Equinox-31B-MXFP4_Q8_0-Imatrix.gguf | 17.7 GB | MXFP4_MOE + Q8_0 |

| Equinox-31B-MXFP4_Q6_K-Imatrix.gguf | 17.1 GB | MXFP4_MOE + Q6_K |

| Equinox-31B-MXFP4-Only-Imatrix.gguf | 16.3 GB | *MXFP4_MOE Only*** |

Updated 2026-07-07:

• Used larger datasets for imatrix
• Added Equinox-31B-MXFP4-Only-Imatrix.gguf

Updated 2026-06-28:

• Used a new combined/blended imatrix
• MXFP4_Q8_0 & MXFP4_Q6_K

- Renamed to use model-quants-imatrix style naming

- Corrected --tensor-type regex

📊 Importance Matrix (Imatrix)

<details>

<summary>Expand to view datasets & details</summary>

The following datasets were used for the imatrix:

• I am currently remaking this imatrix with 2048 ctx, 3000 chunks, with smaller chunks & sets of text

• Old/Current: Custom Target Matrix (2048 ctx with 600 chunks), with all of the following:

- ~1.5MB of my own story/writing data

- eaddario/imatrix-calibration - tools_medium

- Crownelius/Opus-4.5-WritingStyle-1000x-formatted-fixed

- froggeric/imatrix - tiny.txt

- eaddario/imatrix-calibration - text_en_medium

- SimpleStories/SimpleStories

</details>

---

🔍 Precision Matrix & Flavor Variations

Standard global quantization presets (like stock MXFP4) compress the backbone layers uniformly, which degrades the quality of models.\

This repository provides multiple distinct manual configuration layouts to balance precision and memory constraints:

1. The Tri-Quant Hybrid Flavor (MXFP4 + Q8_0 + F16)

Equinox-31B-MXFP4_Q8_0_F16-Imatrix.gguf - Designed for maximum quality preservation, this layout implements a strict 3-Tier Precision Matrix:

• Tier 1 (Core - F16 Precision): token_embd.weight - Protects the critical input/output vocabulary mappings. Dramatically prevents text degradation.
• Tier 2 (Backbone - Q8_0 Precision): attn_q, attn_k, attn_v, attn_output, attn_qkv - All core attention layers. Keeps the model's primary attention mechanism high quality.
• Tier 3 (Dense FFN - MXFP4 Precision): ffn_down, ffn_gate, ffn_up - Downsamples the massive, heavy feed-forward network layers to hardware-optimized 4-bit microscaling blocks.

2. The Dual-Quant Hybrid Flavor (MXFP4 + Q8_0)

Equinox-31B-MXFP4_Q8_0-Imatrix.gguf - Designed for a slightly leaner memory profile, this layout utilizes 2-Tier Precision:

• Tier 1 (Backbone - Q8_0 Precision): Vocabulary embeddings (token_embd), final logit projections (output), layer normalizations, and structural layers fall back to the heavy Q8_0 format.
• Tier 2 (Compute Layers - MXFP4 Precision): attn_k, attn_q, attn_output, ffn_down, ffn_gate, ffn_up - The core attention projections and dense feed-forward network blocks are target-quantized directly to MXFP4 to leverage highly optimized hardware kernels.

3. The Tri-Quant Hybrid Flavor (MXFP4 + Q6_K + Q8_0)

Equinox-31B-MXFP4_Q6_K_Q8_0-Imatrix.gguf - Designed for a tighter VRAM footprint while preserving quality, this layout implements a strict 3-Tier Precision Matrix:

• Tier 1 (Core - Q8_0 Precision): token_embd.weight - Keeps the critical input/output vocabulary mappings. Lessens the text degradation.
• Tier 2 (Backbone - Q6_K Precision): attn_q, attn_k, attn_v, attn_output, attn_qkv - All core attention layers. Keeps the model's primary attention mechanism at decent quality.
• Tier 3 (Dense FFN - MXFP4 Precision): ffn_down, ffn_gate, ffn_up - Downsamples the massive, heavy feed-forward network layers to hardware-optimized 4-bit microscaling blocks.

4. The Dual-Quant Hybrid Flavor (MXFP4 + Q6_K)

Equinox-31B-MXFP4_Q6_K-Imatrix.gguf - Designed for an even tighter VRAM footprint while preserving the mathematical depth of the logical layers, this layout utilizes 2-Tier Precision:

• Tier 1 (Backbone - Q6_K Precision): Vocabulary embeddings (token_embd), final logit projections (output), layer normalizations, and structural layers fall back to the heavy Q6_K format.
• Tier 2 (Compute Layers - MXFP4 Precision): attn_k, attn_q, attn_output, ffn_down, ffn_gate, ffn_up - The core attention projections and dense feed-forward network blocks are target-quantized directly to MXFP4 to leverage highly optimized hardware kernels.

5. Bonus Single-Quant (MXFP4)

Equinox-31B-MXFP4-Only-Imatrix.gguf - Using only MXFP4, this shrinks the model down to 16.3 GB. The quality is not the best, but it can still do decent work.

• Single Tier (All Layers - MXFP4 Precision): All layers are target-quantized directly to MXFP4, for speed and a low VRAM footprint.

---

📝 Exact Conversion Details

These files were converted via llama-quantize utilizing the following manual recipe parameters:

Generate Tri-Quant MXFP4_MOE + Q8_0 + F16:

llama-quantize \
  --tensor-type "token_embd\.weight=F16" \
  --tensor-type "blk\..*\.(ffn_down|ffn_gate|ffn_up)\.weight=MXFP4" \
  --imatrix "imatrix.gguf" \
  "Equinox-31B_Unified_BF16.gguf" \
  "Equinox-31B-MXFP4_Q8_0_F16-Imatrix.gguf" \
  Q8_0

Generate Dual-Quant MXFP4_MOE + Q8_0:

llama-quantize \
  --tensor-type "blk\..*\.(attn_k|attn_q|attn_output|ffn_down|ffn_gate|ffn_up)\.weight=MXFP4" \
  --imatrix "Equinox-31B-combined-imatrix.gguf" \
  "Equinox-31B_Unified_BF16.gguf" \
  "Equinox-31B-MXFP4_Q8_0-Imatrix.gguf" \
  Q8_0

Generate Tri-Quant MXFP4_MOE + Q6_K + Q8_0:

llama-quantize \
  --tensor-type "token_embd\.weight=Q8_0" \
  --tensor-type "blk\..*\.(ffn_down|ffn_gate|ffn_up)\.weight=MXFP4" \
  --imatrix "imatrix.gguf" \
  "Equinox-31B_Unified_BF16.gguf" \
  "Equinox-31B-MXFP4_Q6_K_Q8_0-Imatrix.gguf" \
  Q6_K

Generate Dual-Quant MXFP4_MOE + Q6_K:

llama-quantize \
  --tensor-type "blk\..*\.(attn_k|attn_q|attn_output|ffn_down|ffn_gate|ffn_up)\.weight=MXFP4" \
  --imatrix "imatrix.gguf" \
  "Equinox-31B_Unified_BF16.gguf" \
  "Equinox-31B-MXFP4_Q6_K-Imatrix.gguf" \
  Q6_K

Generate Single-Quant MXFP4_MOE:

llama-quantize \
  --tensor-type "token_embd\.weight=MXFP4" \
  --tensor-type "blk\..*\.(attn_k|attn_q|attn_v|attn_output|ffn_down|ffn_gate|ffn_up)\.weight=MXFP4" \
  --imatrix "imatrix.gguf" \
  "Equinox-31B_Unified_BF16.gguf" \
  "Equinox-31B-MXFP4-Only-Imatrix.gguf" \
  MXFP4_MOE

---

ℹ️ Misc Details

I'm doing this as a side hobby, with my AMD 5900X, 64GB DDR4, RTX 3060 12GB & RTX 5060 Ti 16GB.

I've found this model works best up to 35k context before it starts hallucinating, with it worsening when it reaches 45k, 64k, etc.

It works well for SillyTavern with CharMemory and MessageLimit.\

It also works well enough for story-writing analysis.

---

🤝 Support the Journey

As a passionate developer, I'm always programming, automating, or experimenting with new ideas.\

I love building open-source tools, trying out new web tech, and creating things that don't yet exist, including local AI & quantizing models.

I love sharing these creations to give back to the community.\

If my projects have saved you time or helped you out, consider supporting my work below!

👉 Support me on Ko-fi

---

✨ Acknowledgments

• LatitudeGames for the exceptional Equinox-31B base model.
• bartowski for the BF16 GGUF quants of Equinox-31B.
• mradermacher for the high-quality importance matrix data anchors.