Nurymanau/Nanbeige4.3-33B-MLX-4bit

Nanbeige4.3-33B — MLX 4-bit

Source code and reproduction guide on GitHub

Unofficial standard MLX conversion of Nanbeige/Nanbeige4.3-33B, pinned to c1093c501cc6460427c0ac634db7223d0450c0ee. Original model: Nanbeige LLM Lab. License: Apache 2.0.

Property	Value
Original parameters	33,118,977,024
Architecture	LlamaForCausalLM (llama)
Source dtype	BF16
Quantization	MLX affine 4-bit, group size 64
Weight files	18.631 GB, excluding tokenizer and runtime memory
Conversion	mlx 0.32.3; official quantizer equivalence checked against mlx-lm 0.32.0
Full-model generation	Not tested by this conversion

No custom architecture or remote Python code is required. This repository contains weights and the original tokenizer/chat template; it does not train or alter the model's behavior intentionally beyond quantization.

Run on Apple Silicon

pip install mlx-lm==0.32.0 mlx==0.32.3
mlx_lm.generate --model Nurymanau/Nanbeige4.3-33B-MLX-4bit \
  --prompt "Explain how a hash table works." \
  --max-tokens 512 --temp 0.6 --top-p 0.95

Weight files alone exceed 16 GB. A 32 GB or larger Apple Silicon Mac is a reasonable starting point, not a measured hardware guarantee. Runtime, KV cache, context length and other applications need additional memory. Long contexts can require substantially more memory.

Python with the original chat-template options:

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load("Nurymanau/Nanbeige4.3-33B-MLX-4bit")
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Explain how a hash table works."}],
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
    preserve_thinking=False,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512,
               sampler=make_sampler(temp=0.6, top_p=0.95)))

The upstream template supports enable_thinking, preserve_thinking, and XML/JSON tool-call formatting. It is copied byte-for-byte. Runtime integration with tool parsers, agent frameworks and long contexts has not been tested here.

Validation and limits

• Source shards verified against upstream SHA256; revision pinned above.
• Shardwise converter checked on a synthetic two-layer Llama: all 53 tensors match official mlx-lm quantization exactly; strict weight loading passes.
• Every converted file/header/shape/dtype/index entry is independently audited.
• The first two rows of all 366 quantized matrices are unpacked with NumPy and checked against the original BF16 rows.
• Original norms and tokenizer metadata are checked for exact preservation.
• The full pretrained model was not loaded for inference or generation. No generation quality, throughput, context-length, or full-model parity measurements are claimed. Quantization can change outputs and quality.

VALIDATION.json, CONVERSION.json and SHA256SUMS record the checks and provenance. Conversion/audit source is included in conversion/. Published benchmark results in the upstream card belong to the original authors and were not rerun for this MLX conversion.