DeepSeek-R1-Distill-Llama-70B-lecore-mlx

MLX mixed-precision quantization of deepseek-ai/DeepSeek-R1-Distill-Llama-70B, with the bit width of every linear layer picked by leCore Unicron requantize (per-tensor bits chosen by measurement), ported to GPU for this model size.

How the bits were chosen

1. Each decoder linear tensor was quantized alone (MLX affine, group 64 during measurement) at 2, 3 and 4 bits; the damage was measured as KL(fp16 || quantized) on next-token distributions over a calibration mix (3066 WikiText-2 train tokens, 1022 Python-source tokens, and leCore's structured probe: prose, code, SQL, markdown).
2. Bits were assigned greedily: start at 4 bits everywhere, repeatedly take the downgrade with the least measured damage per byte saved until the predicted file size fit the budget. Embeddings 4-bit, lm_head 6-bit, group size 64.
3. Converted with mlx_lm.convert and a per-layer quant_predicate. Bit histogram (linear tensors): {'2': 251, '3': 205, '4': 102, '8': 0, '16': 2}.

Measured results (held-out: WikiText-2 test + a different Python-source slice, seq 512)

Perplexity was computed on the exact shipped weights (MLX tensors dequantized with mx.dequantize, run in bf16 on GPU). "bpw" = integer bits + 0.5 for the bf16 scale/bias per group of 64.

Model	Size (GB, weights)	bpw	WikiText-2 ppl (32704 tok)	Python-code ppl (8176 tok)
bf16 reference	–	16	8.2625	3.8528
this repo (leCore Unicron mix)	26.77	3.033	14.1309	6.8978
stock MLX 3-bit, group 64	30.87	3.500	155.2344	11.5845

Benchmarks (lm-eval-harness 0.4.9, run on the shipped weights dequantized to bf16 on GPU)

MMLU 5-shot on the first 10 questions of each of the 57 subjects; GSM8K 5-shot on the first 100 test problems (greedy, 320 new tokens, no chat template). These are subsets, so expect about ±2–4 points of sampling noise.

Model	MMLU (5-shot)	GSM8K strict	GSM8K flexible
leCore mix (this repo)	69.8%	83.0%	86.0%
stock 3-bit affine, group 64 (simulated)	25.3%	0.0%	0.0%

Only measured numbers are reported. No downstream-task benchmarks were run. Short-context perplexity does not tell you everything about quality, so check outputs for your own use case.

Usage (Apple Silicon)

pip install -U mlx-lm
mlx_lm.generate --model staccs/DeepSeek-R1-Distill-Llama-70B-lecore-mlx --prompt "Hello" --max-tokens 256 --max-kv-size 8192

Weights are ~26.77 GB. On a 36 GB Mac you may need to raise the GPU wired-memory limit, e.g. sudo sysctl iogpu.wired_limit_mb=30000.

License

Base model: deepseek-ai/DeepSeek-R1-Distill-Llama-70B, released by DeepSeek under the MIT License. That model is fine-tuned from Llama-3.3-70B-Instruct, so the Llama 3.3 Community License Agreement also applies (https://www.llama.com/llama3_3/license/), including its Acceptable Use Policy. Built with Llama. Only quantization was done here: no retraining.