peppi314/Bonsai-27B-dSpark-MLX-4bit

Bonsai 27B dSpark — MLX 4-bit sidecar

This repository contains the dSpark speculative-decoding drafter only for Prism ML's Bonsai-27B, converted to an MLX-compatible 4-bit safetensors sidecar for Higgs.

It is not a standalone language model. Pair it with prism-ml/Bonsai-27B-mlx-1bit as the target model.

What is included

Property	Value
Drafter architecture	dSpark / DFlash, 6 transformer layers
Trained draft block	4 positions
Target hidden-state taps	[1, 16, 31, 46, 61]
Hidden size	5120
Attention heads / KV heads	40 / 4
Markov rank	256
Vocabulary	248,320 tokens
Quantization	MLX affine 4-bit, group size 32
Tensor sidecar size	814,062,178 bytes (776.35 MiB)

This is the lean variant. It reuses the resident target model's bit-identical Q1 token embedding and its Q1 language-model head instead of storing duplicate copies. config.json records this as reuse_target_head: true.

Conversion provenance

The tensors were converted from Bonsai-27B-dspark-Q4_1.gguf at source revision 0cf7e3d21581b169b4df1de8bf01316000e2fbb7.

• Source GGUF SHA-256: 25e73f9f7ab5d1f7f1336711496dbc12da674e639ec88d579dc8683045befb1b
• Converted model.safetensors SHA-256: ae3eeb6db71dc22bd62fea0c0e7c217263389ef44a9e33f550c665536f8e9822
• config.json SHA-256: 4a285dec60023a17df63373123d6b8b8b0d46c55e68b7d0678c6bd109d324a2f

The conversion directly and exactly repacks the source GGUF Q4_1 blocks into MLX affine Q4/group-32 storage. It does not dequantize and requantize the Q4 weights. Representative and exhaustive tensor checks matched the GGUF source after dequantization.

Usage with Higgs

Download both the target and this drafter sidecar:

hf download prism-ml/Bonsai-27B-mlx-1bit \
  --local-dir ./Bonsai-27B-mlx-1bit

hf download peppi314/Bonsai-27B-dSpark-MLX-4bit \
  --local-dir ./Bonsai-27B-dSpark-MLX-4bit

Point Higgs at the drafter directory and load Bonsai as the target through your normal Higgs entry point:

export HIGGS_DFLASH_PATH="$PWD/Bonsai-27B-dSpark-MLX-4bit"

The dSpark runtime validates the target hidden size, vocabulary size, and tap layers at load time. This artifact requires a Higgs build containing Bonsai Q1 and dSpark/DFlash recurrent-tape support.

Fidelity note

The lean sidecar deliberately uses the target's packed Q1 language-model head for proposals. That Q1 head is very close to, but not numerically identical to, the Q4 head shipped inside Prism ML's original dSpark GGUF. This can change which draft tokens are proposed and their acceptance rate. Higgs still verifies every proposal with the Bonsai target, so greedy generation remains target- correct; the drafter cannot substitute unverified tokens into the output.

If exact proposal fidelity to the original Prism ML Q4 head is required, use a full conversion that retains the GGUF output head instead of this lean release.

License and attribution

The source model and this converted artifact are distributed under the Apache License 2.0. See LICENSE.txt and NOTICE.txt.

Created using Bonsai by Prism ML. Bonsai is built from Qwen3.6-27B by Alibaba Cloud. Please see the upstream Bonsai-27B-gguf model card for the original model description, benchmarks, limitations, and citation.