Project Goal & Real-World Results
What This Project Set Out to Solve
Standard Qwen3.8-27B variants often suffer from circular <think> loops, mid-generation coordinate/logic drift, or unfinished multi-system code on complex single-shot prompts—while losing their native MTP speculative decoding head when merged naively. The goal was to combine three complementary coding and reasoning fine-tunes into a single high-accuracy model without sacrificing native MTP speed.
What Was Accomplished
- One-Shot Complex App & Game Execution: Consistently completes demanding single-file builds—such as Choplifter (helicopter tilt/inertia physics, destructible barracks, rescue state machines, and parallax terrain), Space Invaders, and multi-tool utilities—in a single pass where base
Qwen3.8-27B (even at xhigh thinking) and Qwen 3.8 Flash Next frequently fail or misalign physics and graphics.
- More Structured, Human-Like Reasoning: Uses ~20–30% more
<think> tokens than heavily pruned fine-tunes, but spends that budget on clean, top-down architectural planning rather than repetitive self-doubt.
- Out-of-the-Box Visual Polish: Generates clean, well-proportioned HTML/CSS/JS interfaces without needing prompt-engineering tricks like "make it look professional."
- 40–60 t/s with Native MTP: Preserving the
mtp.* head inside the checkpoint enables 40–60 tokens/sec on HTML/code generation (and 25–40 tokens/sec on general text) in llama.cpp (Q8_0 imatrix, BF16 calibration PPL = 3.4195 ± 0.02407) with zero separate draft model required.
Merge Details
Merge Method
This model was merged using the DARE-TIES (dare_ties) merge method using Qwen/Qwen3.8-27B as a base.
Models Merged & Individual Contributions
The following models were included in the merge:
- Jackrong/Qwopus3.8-27B-Coder-Flash (
density: 0.75, weight: 0.60)
- Individual Strengths: Frontier-distilled coding and structural reasoning model known for strong architectural planning, edge-case awareness, and clean front-end/UI aesthetics.
- Contribution to Merge: Serves as the primary coding and spatial design "brain." Retaining 75% of its parameter deltas preserved subtle, low-magnitude creative weights—giving the merge its human-like step-by-step
<think> structure, accurate 2D physics/spatial math, and polished visual design out of the box.
- ukisai/Swift-1.5-Qwen3.8-27b (
density: 0.45, weight: 0.25)
- Individual Strengths: Reinforcement-learning distilled variant engineered to eliminate the repetitive self-correction loops, stalls, and early logic failures common in base
Qwen3.8-27B.
- Contribution to Merge: Acts as an execution stabilizer. Blended at 25% weight, it keeps Qwopus's deep reasoning linear and decisive—preventing mid-file coordinate drift or stalled loops on long single-shot code generations.
- bottlecapai/ThinkingCap-Qwen3.8-27B (
density: 0.35, weight: 0.15)
- Individual Strengths: Fine-tuned to identify and prune runaway chain-of-thought bloat while preserving core logic and benchmark accuracy.
- Contribution to Merge: Serves as a light regularizer (15% weight, 35% density). Kept intentionally low so it never truncates complex architectural planning early, it prevents infinite reasoning spirals while allowing the model to think 20–30% deeper when a multi-system problem genuinely demands it.
Post-Merge Steps Applied
- ChatML Template Injection: Added Qwen's ChatML +
<think> Jinja template into tokenizer_config.json.
- MTP Head Transplant: Extracted all 15
mtp.* tensors from Qwen/Qwen3.8-27B into model-mtp-head.safetensors and registered them in model.safetensors.index.json.
Configuration
The following YAML configuration was used to produce this model:
merge_method: dare_tiesbase_model: Qwen/Qwen3.8-27B # Base model acts as the neutral anchormodels: - model: Jackrong/Qwopus3.8-27B-Coder-Flash parameters: density: 0.75 weight: 0.60 - model: ukisai/Swift-1.5-Qwen3.8-27b parameters: density: 0.45 weight: 0.25 - model: bottlecapai/ThinkingCap-Qwen3.8-27B parameters: density: 0.35 weight: 0.15parameters: normalize: true int8_mask: truedtype: bfloat16