Qwen3.6-35B-A3B KMP Dev (MLX 2-bit)
A pruned Qwen3.6-35B-A3B model optimized for Kotlin Multiplatform development and agentic tool use, converted to MLX 2-bit format.
License: Apache 2.0. Based on Qwen3.6-35B-A3B by Alibaba Cloud (Apache 2.0). Modified via expert pruning.
What's kept
- Kotlin, Swift, Gradle, Coroutines, RxSwift, Jetpack Compose, Compose Multiplatform, SwiftUI, Decompose, Metro, Koin, Ktor, Room, Coil
- English and Russian
- Reasoning, instruction following, tool use / function calling
- General knowledge (humanities)
What's removed
- Other programming languages (Python, Java, Go, Rust, C, JS, PHP, Ruby, TS, HTML, bash, Qt)
- Science/tech/esoteric domains (medicine, law, biology, chemistry, astronomy, physics, esoterics, cooking, dietetics)
- All languages except English and Russian
Pruning method
- Smart pruning: kept top-150 most active experts per layer (out of 256) based on heat data
- Calibrated on 36 texts × 3 runs (Kotlin, Swift, RU/EN, SQL, architecture, reasoning, tool use, JSON/API) — 100% deterministic
- Experts that don't activate on keep-texts are physically removed
- Quantization: 2-bit (MLX)
Usage
MLX
pip install mlx-lm
from mlx_lm import load, generate
model, tokenizer = load("siendsi/Qwen3-6-KMP-Dev-MLX-2bit")
response = generate(model, tokenizer, "Write a Kotlin function to parse JSON")
print(response)
LM Studio
Search for siendsi/Qwen3-6-KMP-Dev-MLX-2bit in LM Studio.
Variants
| Variant | Format | Size | Quality |
|---|---|---|---|
| GGUF Q3_K_M | GGUF | 9.7GB | ~90% of original |
| GGUF Q2_K | GGUF | 7.6GB | ~75% of original |
| GGUF IQ1_M | GGUF | 4.9GB | ~60% of original, loops on tool use |
| MLX 2-bit | safetensors | 6.3GB | ~75% of original |
Architecture
- 40 layers, hybrid: attention (every 4th) + SSM (Gated DeltaNet) + MoE
- 150 experts/layer, 8 active, 1 shared
- embedding 2048, 16 heads, 2 KV heads, head_dim 256
- rope dim 64, mrope [11,11,10,0], freq_base 10M