kapitsaprime/Nanbeige4.3-33B-DFlash2

Nanbeige4.3-33B · DFlash2 drafter

A speculative-decoding draft model for Nanbeige/Nanbeige4.3-33B, kept here as the drafter for Kapitsa on Apple silicon. It is not a standalone language model: it proposes tokens from the 33B's own hidden states, and the 33B checks every one of them, so the text is what the 33B alone would write.

These are ProCreations' weights, unchanged. The model was trained and released by ProCreations/Nanbeige4.3-33B-DFlash2 and repackaged by Luni/Nanbeige-4.3-33B-DFlash2. model.safetensors is byte-identical to both (SHA-256 b031ede26614f8800110e062f81fbaf3d1974dfc43bb32f346e6de941dc0efff), and config.json is Luni's (ProCreations' with tap_shift: 0 written out). All credit for the model goes to them.

What it is

Target	Nanbeige4.3-33B (52 layers, hidden 7168), revision c1093c50
Drafter	DFlash2 block diffusion: 3 layers, ~0.98B parameters, hidden 7168, 28 / 4 heads
Block	8 tokens per proposal (one verified anchor + 7 guesses)
Reads	the target's layer outputs 1, 25 and 49, fused by a learned projection
Shares	the target's embedding and output head (not stored here)
Format	BF16 safetensors, 1.96 GB

One drafter for every pack of the target. It reads the 33B's hidden states, so it pairs with the bf16 model and with the 8-bit and 4-bit MLX packs alike. Acceptance may be slightly lower on a quantized target, and the output is always the target's own.

Why BF16 for Apple silicon. The upstream EXL3 and FP8 branches are NVIDIA formats. This BF16 file is the hardware-neutral one; Kapitsa quantizes the drafter to 4-bit when it loads it, as it does its other DFlash2 drafters.

Status in Kapitsa

Support is in progress, not released. Kapitsa runs DFlash2 drafters for Qwen today; wiring this one to Nanbeige's backend (this drafter's selector shape, its head layout, the 8-row verify on Nanbeige's kernels) is being built and will be measured on Apple silicon before it ships. This card will carry the Mac numbers then.

Upstream measurements (NVIDIA, not Apple silicon)

From ProCreations' and Luni's cards, batch 1:

Domain	Accepted tokens per step	Speedup over plain decoding
Agentic coding	2.64	2.16x
Code	2.21	1.79x
Math and reasoning	2.21	1.78x
Chat	1.52	1.22x
Chinese	1.45	1.18x

ProCreations reports 1.76x greedy / 1.63x at T 1 with SGLang on a BF16 target (RTX PRO 6000). How much of this carries over depends on what an 8-row verify pass costs on the Mac.

Training

Trained by ProCreations with SpecForge on 19.3M response tokens that Nanbeige4.3-33B generated itself (46% agentic, 21% code, the rest math, chat, Chinese, multilingual, structured and long-context prompts). The data sources, their licenses and revisions are in NOTICE and dataset-source-policy.json; the full recipe and provenance are on the ProCreations card.

License

Apache 2.0, as the base model and the upstream drafter. Weights by ProCreations; packaged by Luni; mirrored here for Kapitsa by Nace AI.