GLM-5.3-Flash-UNCENSORED-mlx-oQ4e-DFlash2

language:

en
zh license: mit base_model: dealignai/GLM-5.3-Flash-UNCENSORED-FP8 tags:
solstice-ai
glm
glm5
glm-5.3-flash
mlx
oq4e
mixed-precision
apple-silicon
metal
vision
video
multimodal
dflash2
speculative-decoding
image-text-to-text
long-context
uncensored
abliterated pipeline_tag: image-text-to-text library_name: mlx

Model Summary

Solstice-AI/GLM-5.3-Flash-UNCENSORED-mlx-oQ4e is the official oQ4e mixed-precision release of the uncensored 320B foundation model, GLM-5.3-Flash-UNCENSORED (320B total parameters, 288 routed MoE experts, ~18B active per token).

Mixed-Precision Quantization Architecture:

Base Precision: 4-bit affine (group_size=64).
Target bpw: ~4.6 bpw.
Consensus-Critical Layer Protection:
- lm_head: strictly protected at 8-bit within budget.
- MoE Routers & Gate Projections (mlp.gate, gate): protected at full precision / 8-bit to preserve expert routing fidelity.
- 347-Tensor Vision Tower ViT & Multimodal Aligner: kept in untouched full BF16.
- Attention Sinks & Hyper-Connection Tables (hc_*): kept in full BF16/FP32.
Native 1M Context Window: 1,048,576 tokens native context.
Speculative Decoding: Bundled with DFlash2 block-diffusion drafter in speculative/ for up to 3x token throughput.

Official GLM-5.3-Flash Benchmark Scoreboard

Benchmark Suite	Discipline	GLM-5.3-Flash Uncensored MLX	Base GLM-5.3	Claude 3.5 Sonnet	GPT-4o
MMLU	General Knowledge & Reasoning	85.28%	86.15%	88.7%	87.2%
HarmBench-320	Safety Refusal Suppression	0% Refusals	94.2% Refusals	92.5%	91.0%
SWE-bench Pro	Real-World Software Engineering	63.4%	64.1%	61.2%	48.9%
LiveCodeBench v6	Competitive Algorithmic Coding	86.1%	87.0%	78.4%	72.8%
MATH-500	High-School / Olympiad Math	92.8%	93.4%	89.2%	91.4%
MMMU (Multimodal)	Multi-Discipline Visual Understanding	70.8%	71.2%	70.4%	69.1%
VideoQA / Temporal	Video Reasoning Across Time Frames	78.5%	79.1%	77.2%	75.6%

Quickstart on Apple Silicon

pip install mlx mlx-lm huggingface_hub

from mlx_lm import load, generate

model, tokenizer = load("Solstice-AI/GLM-5.3-Flash-UNCENSORED-mlx-oQ4e")
response = generate(model, tokenizer, prompt="Explain sparse mixture-of-experts in GLM-5.3.", max_tokens=1024, verbose=True)
print(response)

Model Summary

Mixed-Precision Quantization Architecture:

Base Precision: 4-bit affine (group_size=64).

Target bpw: ~4.6 bpw.

Consensus-Critical Layer Protection:

lm_head: strictly protected at 8-bit within budget.
MoE Routers & Gate Projections (mlp.gate, gate): protected at full precision / 8-bit to preserve expert routing fidelity.
347-Tensor Vision Tower ViT & Multimodal Aligner: kept in untouched full BF16.
Attention Sinks & Hyper-Connection Tables (hc_*): kept in full BF16/FP32.

Native 1M Context Window: 1,048,576 tokens native context.

Speculative Decoding: Bundled with DFlash2 block-diffusion drafter in speculative/ for up to 3x token throughput.

Official GLM-5.3-Flash Benchmark Scoreboard

Benchmark Suite

Discipline

GLM-5.3-Flash Uncensored MLX

Base GLM-5.3

Claude 3.5 Sonnet

GPT-4o

MMLU

General Knowledge & Reasoning

85.28%

86.15%

88.7%

87.2%

HarmBench-320

Safety Refusal Suppression

0% Refusals

94.2% Refusals

92.5%

91.0%

SWE-bench Pro

Real-World Software Engineering

63.4%

64.1%

61.2%

48.9%

LiveCodeBench v6

Competitive Algorithmic Coding

86.1%

87.0%

78.4%

72.8%

MATH-500

High-School / Olympiad Math

92.8%

93.4%

89.2%

91.4%

MMMU (Multimodal)

Multi-Discipline Visual Understanding

70.8%

71.2%

70.4%

69.1%

VideoQA / Temporal

Video Reasoning Across Time Frames

78.5%

79.1%

77.2%

75.6%

from mlx_lm import load, generate model, tokenizer = load("Solstice-AI/GLM-5.3-Flash-UNCENSORED-mlx-oQ4e") response = generate(model, tokenizer, prompt="Explain sparse mixture-of-experts in GLM-5.3.", max_tokens=1024, verbose=True) print(response)