base_model: huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated base_model_relation: quantized quantized_by: IsValorum library_name: gguf license: apache-2.0 language:

• en
• zh
• es
• fr
• de
• pt
• it
• ja
• ko tags:
• gguf
• llama.cpp
• quantized
• quantization
• val-apex-i
• apex
• apex-quant
• apex-i-nanoplus
• nanoplus
• qwen
• qwen3.8
• coder
• coding
• agentic
• reasoning
• efficient-thinking
• abliterated
• uncensored
• imatrix pipeline_tag: text-generation

Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated VAL-APEX-I NanoPlus V3 GGUF

The authentic Coder390 + EfficientThink + Abliterated 27B release:

• Based on the upgraded Coder390 fine-tune (trained with multi-model SFT on Opus 5.5, GPT-6 Astra, Grok 4.7, DeepSeek V4 Pro, and RLOO).
• Features EfficientThink alignment, drastically cutting down thinking bloat and repetitive reasoning overhead while preserving rigorous problem-solving depth.
• Fully abliterated by Huihui-AI across layers 2 to 52, eliminating false-positive refusals on sensitive coding, cybersecurity, and reverse-engineering prompts.
• Handcrafted tensor-by-tensor allocation via the canonical VAL-APEX-I V3 standard with 866 verified tensors documented in the Tensor Map and certified tensor-map.tsv file.

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.

The Uncensored 27B Frontier Reasoning & Coding Engine · Coder390 SFT + RLOO · EfficientThink Low Overhead · Abliterated Base · Native 256K Context

Official VAL-APEX-I quantization of huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated.

VAL-APEX-I stands for: Vector-calibrated Asymmetric Layer-wise Outlier-preserving Recurrent-aware Unified Matrix-quantization

[!NOTE]

EXPLORE THE COMPANION EDITIONS

Choose your target quantization profile:

• Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated — MiniPlus V3 — 14.16 GB (Highest precision tier with Q6_K attention & outputs).
• Qwen3.8 27B Base — MiniPlus V3 — 14.16 GB.
• Qwen3.8 27B Base — NanoPlus V3 — 12.18 GB.
• Swift-1.5 Qwen3.8 27B Uncensored MTP — MiniPlus V3 — 14.16 GB.
• Swift-1.5 Qwen3.8 27B Uncensored MTP — NanoPlus V3 — 12.18 GB.

Collections: MiniPlus Abliterated 27B · MiniPlus Standard 27B · NanoPlus Abliterated 27B · NanoPlus Standard 27B

[!IMPORTANT]

THE DEFINITIVE SPECIFICATION IN THE 11 to 12 GB CEILING

This VAL-APEX-I NanoPlus V3 release is 12.18 GB, with 3.57 whole-file BPW and 866 audited tensors. It incorporates an elevated 3-bit floor across all MLP feed-forward blocks, avoiding destructive 2-bit sub-allocations.

Quick Navigation Index

1. Quantization Metrics & Tensor Map
2. Independent Community Benchmarks (NanoBench v0.4.0)
3. Model Files & Technical Specifications
4. Native Context & Runtime Memory
5. Recommended Configuration & Setup
6. Recommended Generation Parameters
7. CRITICAL: Coding Syntax & Repeat Penalty Advisory (Preventing Character Swapping)
8. Hardened Agentic Chat Template & Reasoning Effort
9. Community Compute Fund & Student Support

1. Quantization Metrics

Size & Quality Metrics

[!NOTE] Independent Hardware & Compute Transparency: Generating multi-million token calibration matrices and running certified evaluations on dense hybrid architectures requires dedicated high-end cloud nodes funded 100% out-of-pocket. If you rely on these MiniPlus/NanoPlus V3 builds for your daily coding workloads, contributions to the IsValorum Compute Fund on Ko-fi directly sponsor the GPU hours required for future Day-0 handcrafted quantizations.

Independent Community Benchmarks (NanoBench v0.4.0)

Independent multi-harness empirical evaluation conducted by community member @userr99 on a dedicated local workstation (24 GB VRAM / 64 GB RAM using llama.cpp). Evaluated across the 12-task autonomous coding and agentic suite on NanoBench v0.4.0 (public discussion thread & test logs):

Key Architectural Insights:

• Zero Tool-Calling Degradation: Achieved a unanimous 100/100 in Agentic tasks across all three harnesses, confirming that protecting output.weight in Q5_K and DeltaNet recurrent memory states (ssm_a in F32, ssm_beta in Q8_0) prevents JSON schema rejections and state drift over multi-step workflows.
• Universal Harness Compatibility: Surpassed the 90+ score barrier across all three test harnesses without entering infinite retry loops or syntax divergence.
• Verified Execution Validity: In the DeepSeek run, both partial Systems results (systems-02 scoring 60 and systems-03 scoring 40, alongside systems-01 at 100) were formally marked VALID by the runner verifier, confirming full tool-calling integrity and zero environment or syntax crashes.

Tensor Precision Map

x denotes tensor counts. All columns reflect exact GGUF tensor headers verified in tensor-map.tsv. Normalizations exclude separately listed SSM norms.

The model contains 866 audited tensors: 851 main-model tensors plus 15 MTP tensors. The architecture is a dense hybrid with 64 layers: 48 Gated DeltaNet linear-attention layers and 16 periodic full-attention layers.

Complete tensor-by-tensor audit: names, formats, shapes and parameter counts.

2. Model Files & Technical Specifications

• Base Model: huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated (Abliterated fine-tune of nerkyor's Coder390 SFT + RLOO + EfficientThink checkpoint)
• Parameters: 27B total dense hybrid
• Architecture: 64 main hybrid layers (48 Gated DeltaNet SSM linear attention layers + 16 periodic full quadratic attention layers)
• Context Length: 262,144 tokens (native 256K)
• Quantization Standard: VAL-APEX-I NanoPlus V3 calibrated against mradermacher's certified importance matrix (imatrix.gguf)
• Abliteration Status: True directional weight orthogonalization across layers 2-52, removing refusal vectors while maintaining full analytical capability.

3. Native Context & Runtime Memory

The model supports a native context window of 262,144 tokens. Its 64-layer architecture comprises 48 Gated DeltaNet layers and 16 full-attention layers. Runtime memory usage depends on the backend, KV-cache precision, batch size, compute buffers, and offload configuration.

4. Recommended Configuration & Setup

llama.cpp Server (OpenAI-Compatible API)

llama-server \
  -m Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated.VAL-APEX-I-NanoPlus-V3.gguf \
  --jinja \
  -ngl 99 \
  -c 32768 \
  --host 0.0.0.0 \
  --port 8080

llama.cpp CLI

llama-cli \
  -m Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated.VAL-APEX-I-NanoPlus-V3.gguf \
  --jinja \
  -ngl 99 \
  -c 32768 \
  -p "<|im_start|>system\nYou are an expert autonomous software engineer.<|im_end|>\n<|im_start|>user\nWrite an optimized concurrent worker pool in Go with backpressure handling.<|im_end|>\n<|im_start|>assistant\n<think>\n"

LM Studio / Ollama

1. Import the .gguf file directly into your local library.
2. Ensure GPU acceleration is set to Maximum / 100% offload.
3. Set Context Length to 32768.
4. Verify chat template is set to Qwen ChatML with <think> delimiter support.

5. Recommended Generation Parameters

6. CRITICAL: Coding Syntax & Repeat Penalty Advisory (Preventing Character Swapping)

[!IMPORTANT]

PREVENTING SYNTAX & TOKEN SWAPPING IN CODE WORKFLOWS

In programming code, brackets ({, }), assignment operators (=), and indentation whitespace repeat constantly across multi-line structures.

Common Issue: Many local frontends (such as LM Studio defaults, Ollama, or web interfaces) ship with repeat_penalty set to 1.1 or 1.15. While this prevents loops in creative prose, applying repeat penalties to code artificially penalizes necessary syntax tokens. When the logit of { drops, the model is forced to emit the next closest mathematical token (= or [), resulting in character swapping or dropped/doubled whitespace.

Eliminating Character Swapping:

1. Disable Repeat Penalties (Required for Code):
   • repeat_penalty: 1.0 (strictly disabled)
   • presence_penalty: 0.0
   • frequency_penalty: 0.0
2. Use Min-P Filtering:
   • Set min_p: 0.05 with temperature: 0.6. This eliminates low-probability hallucinations without distorting syntax distribution.

7. Hardened Agentic Chat Template & Reasoning Effort

The repository includes the official, syntax-verified chat_template.jinja compatible with llama.cpp (llama-server), vLLM, and SGLang.

Reasoning Effort Guidance

You can control the internal reasoning budget by passing reasoning_effort in your generation payload:

• "low" / "minimal": Instructs the model to keep internal thought chains brief and proceed directly to code execution.
• "medium": Standard balanced analytical depth.
• "high" / "xhigh": Exhaustive, multi-step edge-case exploration and systematic planning.

8. Community Compute Fund & Student Support

Creating custom multi-tensor precision recipes, running multi-million token importance matrix sweeps, and conducting independent multi-harness benchmarks requires dedicated high-VRAM cloud instances funded 100% out-of-pocket.

If these quantizations are valuable to your local software development, research, or agentic pipelines, your support on Ko-fi makes a direct difference:

• Sponsoring cloud compute time for upcoming model architectures.
• Maintaining zero-compromise Day-0 releases for the open-source community.
• Directly supporting my engineering university studies.

Thank you to everyone in the community who tests, reports feedback, and supports this work!