kapitsaprime/Clover-35B-A3B-Preview-MLX-4bit

Clover-35B-A3B-Preview, 4-bit for Kapitsa

DiffuSpace's Clover-35B-A3B-Preview (a block-diffusion MoE tuned from Qwen3.6-35B-A3B, 3B active per token) quantized for Kapitsa, the Metal inference engine for Apple silicon. 19.5 GB.

kapitsa serve --model clover

Kapitsa downloads this pack with Qwen3.6-35B-A3B's DFlash2 drafter (incoai/Qwen3.6-35B-A3B-DFlash2) beside it.

What is in it

• MLX affine 4-bit, group 64, for the experts, attention, DeltaNet, embedding and output layer; 8-bit router and shared expert gate; norms, the DeltaNet's A_log, dt_bias and convolutions, and Clover's own right convolution and softmax controls kept in bf16 as stored. Made from the bf16 checkpoint by tools/convert_clover.py in Kapitsa's repository; conversion.json lists every file's sha256.
• Tensor names under language_model. as in MLX's Qwen3.6 packs. mlx-lm has no clover_moe model, so this pack is for Kapitsa.
• config.json, generation_config.json (Clover's repetition, length and </think> controls), the tokenizer and chat template, LICENSE and ATTRIBUTIONS.md as DiffuSpace ships them.

Decoding in Kapitsa

NI_CLOVER_DECODE picks how Kapitsa decodes Clover:

Mode	How	Decode, M5 Max (T 0.6 / greedy)
auto (default)	Clover's causal pathway, drafted by Qwen3.6-35B-A3B's DFlash2 drafter	319 / 332 tok/s
spec	DiffuSpace's self-speculative decoding: Clover drafts 7 tokens by denoising, then verifies them	214 / 210 tok/s
diffusion	Hugging Face's block diffusion (LowConfidence, block 32, threshold 0.95)	46 tok/s greedy

Prefill: 6,002 tok/s at 4K tokens. Measured with kapitsa bench on a MacBook Pro M5 Max (128 GB), Kapitsa 0.2.24.

License

Apache-2.0, as Clover and Qwen3.6-35B-A3B. Clover is DiffuSpace's work, derived from Qwen3.6-35B-A3B (Copyright 2026 Alibaba Cloud); see LICENSE and ATTRIBUTIONS.md. This repository changes only the weights' precision and layout.