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

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.tsv file.

Action Required: Please delete any older .gguf file downloaded prior to today and download the newly published certified build: Qwen3.8-27B-EfficientThink-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.

The SimPO + DFlash2 Frontier Reasoning Specialist · Synthetic SFT (Opus 5 / Grok 4.6 / GPT 5.6 Sol) · Native 256K Context

Official VAL-APEX-I quantization of nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2.

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 — MiniPlus V3 — 13.91 GB.
• Qwen3.8 27B Huihui Abliterated — MiniPlus V3 — 14.16 GB.
• Qwen3.8 27B Huihui Abliterated — 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 11.18 GB, with 3.27 whole-file BPW and 866 audited tensors. Its reported WikiText-2 PPL is 5.6504 ± 0.4832.

---

Quick Navigation Index

1. Quantization Metrics & Tensor Map
2. Independent Community Benchmarks (NanoBench v0.4.0)
3. Upstream Lineage & Ecosystem Compatibility (DFlash2, SGLang, vLLM, Lynn Agent)
4. Model Files & Technical Specifications
5. Native Context & Runtime Memory
6. Recommended Configuration & Setup
7. Recommended Generation Parameters
8. CRITICAL: Coding Syntax & Repeat Penalty Advisory (Preventing Character Swapping)
9. Hardened Agentic Chat Template & Reasoning Effort
10. 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 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.

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 Q4_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.

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.

This release contains 866 audited tensors: the 851 main-model tensors listed above plus 15 MTP tensors in blk.64.* (block_count = 65), seven of them F32 and eight matrices (Q4_0 x6, Q4_1 x1, Q4_K x1). A separate DFlash2 draft model is also provided as a companion file.

V3 architectural allocation: NanoPlus V3 keeps a 3-bit floor across the SwiGLU MLP blocks: ffn_gate and ffn_up use IQ3_XXS on 48 layers and IQ2_S on layers 24-39, and ffn_down uses IQ3_S on outer layers 0-7 and 56-63 and IQ3_XXS on core layers 8-55. The output head (output.weight) and the 16 attn_output projections are Q4_K; the attention gates are IQ3_S on 28 layers and IQ2_S on 19, and the linear-attention outputs are IQ3_S on 47 layers, with layer 0 in Q4_0 in both cases.

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.

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

---

2. Upstream Lineage & Ecosystem Compatibility

This model is derived from nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 and inherits its full ecosystem compatibility:

1. Training Lineage & Data Mixture

• Pretrained Base: Qwen3.8-27B hybrid linear-quadratic architecture.

• Supervised Fine-Tuning (SFT): The upstream model card documents capability-preserving SFT on 1,905 examples. The release identifier names K3 / Opus 5 / Grok 4.6 / GPT 5.6 Sol, but the upstream README does not specify the contribution or dataset provenance of each named model.

• Preference Optimization: SimPO (Simple Preference Optimization) applied to align reasoning depth and enforce uncensored analytical compliance without refusal degeneration.

• Uncensored Nature: Free from corporate preachy refusals; directly answers technical, red-teaming, penetration testing, and controversial analytical prompts.

2. DFlash2 Multi-Token Prediction (MTP) Speculative Decoding

• Native Integrated MTP Head: The main GGUF includes the 15 MTP prediction tensors in blk.64.* directly within the model (block_count = 65). A separate DFlash2 companion draft model (speculative/dflash2-qwen38-27b-Q4_K_M.gguf) is also available for alternative speculative decoding.
• Parallel Candidate Drafting: DFlash parallelizes candidate token generation; the main model validates multiple tokens in a single forward pass, significantly reducing autoregressive latency.
• Compatibility with llama.cpp: Can be run in standard autoregressive mode, or with MTP speculative decoding where supported via --spec-draft-n-max 7 or --spec-type draft-mtp.

3. Serving Engine Support

• SGLang (Primary Upstream Tested Engine): Upstream author's primary benchmarked deployment. Supports SGLang DFlash2 verify block size 8.
• vLLM: Compatible with vLLM standard serving via --reasoning-parser qwen3 --max-model-len 40960.
• Lynn Agent (v0.87.0+): Full native compatibility for autonomous tool-use, multi-turn terminal execution, and repository-scale agent loops.
• llama.cpp / llama-server: 100% plug-and-play with official llama.cpp builds (b11000+) supporting hybrid DeltaNet SSM architectures.

---

3. Model Files & Technical Specifications

• Base Model: nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2
• Parameters: 27B total dense hybrid
• Architecture: 64 hybrid base layers (48 Gated DeltaNet SSM linear attention layers + 16 periodic full quadratic attention layers) + 1 native MTP prediction layer (layer 64) for a total of 65 blocks
• Context Length: 262,144 tokens (native 256K)
• Quantization Standard: VAL-APEX-I NanoPlus V3 calibrated against official Qwen3.8-27B importance matrix (imatrix)
• Ecosystem: SGLang, vLLM, Lynn Agent v0.87.0, llama.cpp

---

4. 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. Check the runtime memory allocation for your chosen settings.

---

5. Recommended Configuration & Setup

llama.cpp Server (OpenAI-Compatible API)

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

Multimodal Vision Inference (llama.cpp)

Pass the companion mmproj file using the --mmproj flag:

llama-server \
  -m Qwen3.8-27B-EfficientThink-VAL-APEX-I-NanoPlus V3.gguf \
  --mmproj mmproj-Qwen3.8-27B-Q4_K_M.gguf \
  --jinja \
  -ngl 99 \
  -c 32768 \
  --host 0.0.0.0 \
  --port 8080

Speculative Decoding / Multi-Token Prediction (MTP)

Accelerate generation throughput using the companion DFlash2 draft model (or run native single-stream with the integrated MTP head):

llama-server \
  -m Qwen3.8-27B-EfficientThink-VAL-APEX-I-NanoPlus V3.gguf \
  -md speculative/dflash2-qwen38-27b-Q4_K_M.gguf \
  --draft-max 7 \
  --jinja \
  -ngl 99 \
  -c 32768 \
  --host 0.0.0.0 \
  --port 8080

llama.cpp CLI

llama-cli \
  -m Qwen3.8-27B-EfficientThink-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 mathematical proof of the Riemann mapping theorem.<|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.

---

6. Recommended Generation Parameters

---

7. 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. Calibrate Samplers:
   • temperature: 0.60 (or 0.20 - 0.30 for strict, deterministic code syntax)
   • min_p: 0.05 (prunes low-probability noise tokens effectively)
   • top_p: 0.95
   • top_k: 20
3. Native Jinja Formatting: Always pass the --jinja flag so the tokenizer handles leading-space BPE tokens cleanly.

---

8. Hardened Agentic Chat Template & Reasoning Effort

[!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.

---

10. 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).