Tiel-Coder-35B-A3B Splash Package
Community-built Splash inference package for
peculiar-ragdoll/Tiel-Coder-35B-A3B-MLX-oQ4e-MTP,
converted to the splash-packed-q4-moe format so it runs through the Splash
engine inside LM Studio.
Compatibility
Splash 1.3.0 only. Splash 1.3.1 removed support for the legacy packaged format, so this model will not load on 1.3.1 or newer. If LM Studio has automatic runtime updates enabled, turn that off before installing.
What's inside
target/— Tiel-Coder-35B-A3B weights repacked from the MLX checkpoint into 4-bit affine Splash sections (40 layers, head, embedding).draft/— DFlash2 speculative-decoding draft model, reused unchanged from incoai/Qwen3.6-35B-A3B-Splash.vision/— vision tower, reused unchanged from the official 35B package.tokenizer/— tokenizer files for the Qwen3.5-MoE vocabulary (248,320 tokens).
Conversion notes
The source MLX uses oMLX's dynamic oQ quantization (4/5/6/8-bit per tensor). Tensors that were not already 4-bit were dequantized to BF16 and requantized to 4-bit affine (group size 64) following MLX's quantization scheme. The 8-bit router gate was quantized from BF16. Expect small quality differences versus the source MLX checkpoint.
The DFlash2 draft was trained against stock Qwen3.6-35B-A3B, not this fine-tune. Speculative acceptance rates on Tiel-Coder are unverified.
Attribution
- Base model: Tiel-Coder-35B-A3B by peculiar-ragdoll (fine-tune of Ornith-1.5-35B-A3B for coding)
- Package format: Splash by IncoAI
- Draft/vision binaries: incoai/Qwen3.6-35B-A3B-Splash
This is an unofficial community conversion, not affiliated with peculiar-ragdoll or IncoAI.