Swift-1.5-Qwen3.8-27B-Uncensored-MTP VAL-APEX-I MiniPlus 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.

Uncensored Frontier 27B Dense Hybrid Reasoning Model with Intact Vision, Multilingual Syntax Preservation, and Native MTP Self-Speculation

Official VAL-APEX-I MiniPlus V3 quantization of ajgazin/Swift-1.5-Qwen3.8-27B-Uncensored-MTP. This edition quantizes the uncensored foundation model built on UkisAI's Swift-1.5 reasoning fine-tune of Qwen3.8-27B with the official canonical V3 recipe (strictly enforcing IQ4_XS and protecting attention/recurrent matrices).

• Multilingual Source Model: Preserves the source model's multilingual tokenizer and native chat template.
• Native MTP Speculative Head Included: Features the built-in Multi-Token Prediction auxiliary head (blk.64.*, 15 tensors), for MTP speculative decoding in compatible backends without an external draft model.
• Multimodal Optical Vision: Retains the high-precision vision encoder (mmproj-Q8_0.gguf) for visual and document reasoning.
• ISTA-DASLab Calibrated: Quantized with the official ISTA-DASLab importance matrix (imatrix-qwen3.8-27b.gguf) computed across diverse tokens to safeguard critical recurrent and projection weights.

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:

• Qwen3.8 27B — MiniPlus V3 — 14.16 GB.
• Qwen3.8 27B — NanoPlus V3 — 12.18 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

---

Quick Navigation Index

1. Quantization Metrics & Tensor Map
2. Independent Community Benchmarks (Shortbench / NanoBench #06 Global - 100%)
3. Model Files & Technical Specifications
4. Recommended Configuration & Setup
5. Community Compute Fund & Priority Model Requests

---

1. Quantization Metrics

Size & Quality Metrics

This MiniPlus edition was evaluated on WikiText-2 under the same conditions as its previous recipe, as confirmed by the release maintainer. Its reported PPL changed from 6.6669 to 6.2501 (6.25% lower). The current evaluation uses llama-perplexity with context 512, batch 512, four chunks and GPU offload (-ngl 99), from llama.cpp build 11512 (a11f57ba9).

Whole-file BPW is calculated from the complete GGUF file size and its 27,320,697,856 stored parameters, including MTP.

Evaluation runs: current canonical V3 job · previous recipe job.

🏅 Independent Community Benchmarks (Shortbench / NanoBench v0.5.0)

[!NOTE] Verified Autonomous Agent Evaluation: Evaluated on the Shortbench / NanoBench v0.5.0 autonomous coding and agentic benchmark (12 objective software engineering, pipeline, and system integrity tasks).

In verified evaluations on the public Shortbench Leaderboard, Swift-1.5-Qwen3.8-27B-Uncensored-MTP-VAL-APEX-I-MiniPlus-V3 achieved a flawless 100% resolution score, ranking in the Top 6 globally:

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 Q6_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

× denotes tensor counts. All three columns were checked against the released GGUF tensor headers. Normalizations exclude the separately listed SSM norms and MTP tensors.

Both Swift files contain 866 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 MTP head, not another main layer.

V3 layer allocation: MiniPlus MLP gate/up use IQ3_S in layers 0-15 and 48-63, and IQ3_XXS in layers 16-47; all 64 main MLP down tensors use IQ4_XS. NanoPlus MLP gate/up use IQ3_XXS in layers 0-7 and 56-63, and IQ2_XXS in layers 8-55; MLP down uses IQ4_XS in the outer 16 layers and IQ3_XXS in the remaining 48.

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, including blk.0.ssm_alpha.weight, remain F32; this is the single difference from the recipe files, which request Q4_0 for layer-0 alpha.

Integrated MTP: blk.64.* contains 424,699,392 stored parameters in 15 tensors: seven F32 tensors and eight matrices (Q4_0 ×6, Q4_1 ×1, Q6_K ×1). Runtime acceleration depends on backend support and configuration; the WikiText-2 results above do not measure speculative-decoding speed.

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

GSQ-RCO tensor-map reference GGUF.

---

2. Model Files & Technical Specifications

• Architecture: qwen35, dense hybrid Gated DeltaNet / full attention.
• Native context: 262,144 tokens (256K).
• Stored parameters: 27,320,697,856, including 424,699,392 in the MTP head.
• Vision projector: 334 tensors; Q8_0 ×83, F16 ×27 and F32 ×224. Its filename describes the quantization profile, not a uniform format for every tensor.

---

3. Recommended Configuration & Setup

llama-server (OpenAI-Compatible API with Native MTP Self-Speculation)

llama-server \
  -m Swift-1.5-Qwen3.8-27B-Uncensored-VAL-APEX-I-MiniPlus-V3.gguf \
  --mmproj mmproj-Q8_0.gguf \
  --port 8080 \
  --spec-type draft-mtp \
  --spec-draft-n-max 2 \
  -ngl 99 \
  -c 16384

---

5. Community Compute Fund & Priority Model Requests

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

If you find these builds valuable and would like to support the project or sponsor GPU compute time, you can contribute via Ko-fi: