The legacy build suffered from severe recurrent state instability, context degradation, and unintended language mixing during generation.
The current release completely resolves these issues:
- Complete long-context associative recall and conversational stability restored.
- Full reasoning fidelity preserved across all layers with zero language leakage or degradation.
- Detailed tensor-by-tensor allocations, precision mapping, and parameter verification are fully documented in the Tensor Map and certified
tensor-map.tsvfile.Action Required: Please delete any older
.gguffile downloaded prior to today and download the newly published certified build:Huihui-Qwen3.8-27B-Abliterated-VAL-APEX-I-NanoPlus-V3.gguf.
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.
Official VAL-APEX-I quantization of huihui-ai/Huihui-Qwen3.8-27B-abliterated.
VAL-APEX-I stands for: Vector-calibrated Asymmetric Layer-wise Outlier-preserving Recurrent-aware Unified Matrix-quantization
[!NOTE]
EXPLORE THE QWEN3.8 27B VAL-APEX-I EDITIONS
Choose a model variant and quantization profile:
- 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.
- Qwen3.8 27B — MiniPlus V3 — 14.16 GB.
- Qwen3.8 27B — NanoPlus V3 — 12.18 GB.
- Qwen3.8 27B EfficientThink Uncensored — MiniPlus V3 — 13.91 GB.
- Qwen3.8 27B EfficientThink Uncensored — NanoPlus V3 — 11.93 GB.
- Qwen3.8 27B Huihui Abliterated — MiniPlus V3 — 14.16 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. Its reported WikiText-2 PPL is 5.7729 ± 0.4960.
| Edition / quantization | File size | Whole-file BPW | WikiText-2 PPL |
|---|---|---|---|
| Qwen3.8 27B Huihui Abliterated — NanoPlus V3 (this release) | 12.18 GB (11.35 GiB) |
3.57 | 5.7729 ± 0.4960 |
[!NOTE] Independent Hardware & Compute Transparency: Generating multi-million token calibration matrices and running certified perplexity evaluations on local architectures requires dedicated high-end cloud nodes funded 100% out-of-pocket. If you rely on these MiniPlus/NanoPlus V3 builds for your local workloads, contributions to the IsValorum Compute Fund on Ko-fi directly sponsor the GPU hours required for future Day-0 handcrafted quantizations.
[!NOTE] Verified Autonomous Agent Evaluation: Evaluated across the Shortbench / NanoBench autonomous coding and agentic suite (12 objective software engineering, pipeline, and system integrity tasks).
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 (public discussion thread & test logs):
| Test Harness | Total Score | Solved Tasks | Agentic (Tool Use) | Coding (Implementation) | Systems (Integrity) | Avg. Runtime | Public Evidence |
|---|---|---|---|---|---|---|---|
| Oh My Pi (OMP 18.7) | 96.7% | 10 / 12 | 100 / 100 | 100 / 100 | 86.7 / 100 | aprox. 1h 47m | Discussion #2 Logs |
| Hermes Agent (v0.21.5) | 95.4% | 9 / 12 | 100 / 100 | 97.0 / 100 | 86.7 / 100 | aprox. 1h 47m | Discussion #2 Logs |
| DeepSeek Harness (0.2.0-rc.2) | 90.4% | 9 / 12 | 100 / 100 | 97.0 / 100 | 66.7 / 100 | aprox. 1h 47m | Discussion #2 Logs |
Key Architectural Insights:
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.| Component | MiniPlus V3 | NanoPlus V3 | GSQ-RCO IQ3_S |
|---|---|---|---|
| Output head | Q6_K x1 |
Q5_K x1 |
Q4_K x1 |
| Token embeddings | Q4_K x1 |
Q3_K x1 |
IQ2_S x1 |
| Normalizations | F32 x161 |
F32 x161 |
F32 x161 |
SSM A (ssm_a) |
F32 x48 |
F32 x48 |
F32 x48 |
SSM convolution (ssm_conv1d) |
F32 x48 |
F32 x48 |
F32 x48 |
SSM time-step bias (ssm_dt) |
F32 x48 |
F32 x48 |
F32 x48 |
SSM norm (ssm_norm) |
F32 x48 |
F32 x48 |
F32 x48 |
Attention gates (attn_gate) |
Q4_0 x1Q4_K x47 |
Q4_0 x1Q4_K x47 |
IQ2_S x2IQ3_S x18IQ3_XXS x9IQ4_XS x12Q2_K x4Q4_K x3 |
Linear QKV (attn_qkv) |
Q4_0 x1Q4_K x47 |
Q3_K x47Q4_0 x1 |
IQ2_XS x1IQ2_XXS x1IQ3_S x22IQ3_XXS x13IQ4_XS x9Q2_K x1Q4_K x1 |
SSM alpha (ssm_alpha) |
F32 x48 |
F32 x48 |
BF16 x48 |
SSM beta (ssm_beta) |
Q4_0 x1Q8_0 x47 |
Q4_0 x1Q8_0 x47 |
BF16 x48 |
SSM output (ssm_out) |
Q4_0 x1Q5_K x47 |
Q4_0 x1Q4_K x47 |
IQ3_S x22IQ3_XXS x4IQ4_XS x16Q4_K x6 |
Full attention Q (attn_q) |
Q4_K x16 |
Q3_K x16 |
IQ2_XXS x1IQ3_S x3IQ3_XXS x3IQ4_XS x2Q2_K x6Q4_K x1 |
Full attention K (attn_k) |
Q5_K x16 |
Q4_K x16 |
IQ2_S x1IQ3_S x1IQ3_XXS x1IQ4_XS x8Q4_K x5 |
Full attention V (attn_v) |
Q6_K x16 |
Q5_K x16 |
IQ3_S x6IQ3_XXS x1IQ4_XS x1Q4_K x8 |
Full attention output (attn_output) |
Q6_K x16 |
Q5_K x16 |
IQ3_S x10IQ3_XXS x1IQ4_XS x2Q4_K x3 |
MLP down (ffn_down) |
IQ4_XS x64 |
IQ3_XXS x48IQ4_XS x16 |
IQ2_S x4IQ2_XS x3IQ3_S x22IQ3_XXS x7IQ4_XS x21Q2_K x1Q4_K x6 |
MLP gate (ffn_gate) |
IQ3_S x32IQ3_XXS x32 |
IQ3_XXS x64 |
IQ1_M x1IQ2_S x4IQ2_XS x4IQ2_XXS x1IQ3_S x15IQ3_XXS x21IQ4_XS x15Q2_K x1Q4_K x2 |
MLP up (ffn_up) |
IQ3_S x32IQ3_XXS x32 |
IQ3_XXS x64 |
IQ2_S x5IQ2_XS x1IQ2_XXS x2IQ3_S x25IQ3_XXS x18IQ4_XS x10Q4_K x3 |
Additional MTP head (blk.64.*) |
F32 x7Q4_0 x6Q4_1 x1Q6_K x1 |
F32 x7Q4_0 x6Q4_1 x1Q5_K x1 |
- |
x denotes tensor counts. All columns reflect exact GGUF tensor headers verified in tensor-map.tsv. Normalizations exclude separately listed SSM norms.
Both Swift/Base/Huihui files contain 866 audited tensors: 851 main-model tensors plus 15 MTP tensors. The main model is a dense hybrid architecture with 64 layers: 48 Gated DeltaNet linear-attention layers and 16 full-attention layers. The additional blk.64.* block is the integrated MTP head.
V3 architectural allocation: NanoPlus V3 features an elevated 3-bit floor across all SwiGLU MLP blocks (ffn_gate and ffn_up set uniformly to IQ3_XXS across all 64 layers; ffn_down uses IQ4_XS in outer layers 0-7 / 56-63 and IQ3_XXS in core layers 8-55). Output head (output.weight) is calibrated to Q5_K (5.50 bpw) to maximize token throughput and preserve dynamic range, entirely eliminating the linguistic decay in non-English contexts associated with 2-bit sub-allocations.
Actual layer-0 formats: blk.0.attn_gate.weight, blk.0.attn_qkv.weight, blk.0.ssm_beta.weight and blk.0.ssm_out.weight use Q4_0. All 48 ssm_alpha tensors remain F32 for mathematical stability.
Integrated MTP: blk.64.* contains 424,699,392 stored parameters in 15 tensors: seven F32 tensors and eight matrices (Q4_0 x6, Q4_1 x1, Q5_K x1). Runtime acceleration depends on backend support and configuration; speculative decoding operates via native acceptance drafting.
Complete tensor-by-tensor audit: names, formats, shapes and parameter counts.
| File Name | File Size | Weight file size | BPW | Description |
|---|---|---|---|---|
Huihui-Qwen3.8-27B-Abliterated-VAL-APEX-I-NanoPlus Huihui-Qwen3.8-27B-Abliterated-VAL-APEX-I-NanoPlus-V3.gguf |
12.18 GB (11.35 GiB) | 11.35 GiB | 3.57 BPW | Core language, uncensored frontier reasoning, CoT thought blocks & hybrid SSM/attention |
blk.64.*imatrix)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. Check the runtime memory allocation for your chosen settings.
llama-server \
-m Huihui-Qwen3.8-27B-Abliterated-VAL-APEX-I-NanoPlus V3.gguf \
--jinja \
-ngl 99 \
-c 32768 \
--host 0.0.0.0 \
--port 8080
llama-cli \
-m Huihui-Qwen3.8-27B-Abliterated-VAL-APEX-I-NanoPlus V3.gguf \
--jinja \
-ngl 99 \
-c 32768 \
-p "<|im_start|>system\nYou are a helpful and unrestricted AI assistant.<|im_end|>\n<|im_start|>user\nExplain the physical principles of acoustic levitation.<|im_end|>\n<|im_start|>assistant\n<think>\n"
.gguf file directly into your local library.32768.<think> delimiter support.| Hyperparameter | Value | Description |
|---|---|---|
| Temperature | 0.60 |
Recommended default for analytical reasoning and coding (use 1.0 for creative prose). |
| Top-P | 0.95 |
Nucleus sampling parameter. |
| Top-K | 20 |
Top-k vocabulary filter. |
| Min-P | 0.05 |
Prunes low-probability noise tokens effectively. |
| Repetition Penalty | 1.00 |
Strictly disabled for code syntax; prevents character swapping. |
| Template Engine | --jinja |
Recommended official Jinja chat template flag. |
| Context Size | 32768 |
32K default (scalable to 256K). |
[!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_penaltyset to1.1or1.15. 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:
- Disable Repeat Penalties (Required for Code):
repeat_penalty: 1.0(strictly disabled)presence_penalty: 0.0frequency_penalty: 0.0- Calibrate Samplers:
temperature: 0.60(or0.20-0.30for strict, deterministic code syntax)min_p: 0.05(prunes low-probability noise tokens effectively)top_p: 0.95top_k: 20- Native Jinja Formatting: Always pass the
--jinjaflag so the tokenizer handles leading-space BPE tokens cleanly.
[!TIP]
MULTI-LEVEL REASONING EFFORT CONTROL
This model supports multi-level reasoning effort control via the Jinja template:
low/minimal: Keeps internal thinking concise and focused strictly on immediate execution steps to minimize latency in automated loops.medium(default): Balanced, structured reasoning process with standard analytical depth.high/xhigh: Guides the model to formulate a clear implementation plan upfront before generating response, avoiding circular self-doubt loops.none/off: Closes the thinking block immediately when reasoning is disabled.
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).
有文件?提交种子。
提交种子huihui-ai/Huihui-Qwen3.8-27B-abliterated