Solar-Open-100B-MXFP4

W4A4 (MXFP4) quantized version of upstage/Solar-Open-100B, with both weights and activations quantized to OCP MXFP4 format.

Solar Open is a dense-MoE hybrid large language model developed by Upstage, featuring 100B total parameters with shared expert and routed expert architecture.

Quantization Details

Property	Value
Base Model	upstage/Solar-Open-100B (100B params, dense-MoE hybrid)
Precision	W4A4 (MXFP4 Weight + MXFP4 Activation)
Weight Quantization	OCP MXFP4 (E2M1), Static, group_size=32, E8M0 shared scales
Activation Quantization	OCP MXFP4 (E2M1), Dynamic, group_size=32, E8M0 shared scales
Quantization Tool	quanto + AMD Quark 0.11.1 (file-to-file)
Algorithm	RTN (Round-To-Nearest)
Original Size	192 GB
Quantized Size	63.4 GB
Compression Ratio	3.0x

Benchmark Results

Evaluated using vLLM (ROCm) with 5-shot prompting on AMD MI355 (gfx950).

MMLU (5-shot)

Category	Baseline (BF16)	MXFP4	Delta
Overall	77.58%	76.78%	-0.80%
Humanities	71.09%	70.48%	-0.61%
Social Sciences	86.90%	86.32%	-0.58%
STEM	74.44%	73.04%	-1.40%
Other	81.36%	80.69%	-0.67%

KMMLU (5-shot, Korean)

Category	Baseline (BF16)	MXFP4	Delta
Overall	57.38%	56.45%	-0.93%
Applied Science	53.28%	52.41%	-0.87%
HUMSS	66.43%	66.00%	-0.43%
STEM	58.13%	57.29%	-0.84%
Other	56.64%	55.19%	-1.45%

Excluded Layers

This model uses the attn-excl strategy: all attention projections and all shared expert projections are excluded from quantization, leaving only MoE routed expert weights quantized to MXFP4.

• Self-Attention (all 48 layers): q_proj, k_proj, v_proj, o_proj
• Shared Expert MLP (all 48 layers): gate_proj, up_proj, down_proj
• MoE Router Gates: All mlp.gate layers
• Standard exclusions: lm_head, embed_tokens, all norm layers

Total: 340 excluded layers. This approach achieves less than 1% accuracy degradation vs BF16 while maintaining 3.0x compression.

Hardware

• Quantized on: AMD MI355 (gfx950), 288 GB VRAM
• Tested with: vLLM v0.18.2 (ROCm), TP=1
• Compatible with: vLLM with Quark quantization support (quant_method: "quark")

Usage

vLLM

vllm serve haanjack/Solar-Open-100B-MXFP4 \
  --trust-remote-code \
  --tensor-parallel-size 1 \
  --max-model-len 4096

Quantization Reproduction

from quanto import UnifiedQuantizer, UnifiedConfig

NUM_LAYERS = 48
exclude = ["lm_head", "*embed*", "*norm*", "*.gate"]
for i in range(NUM_LAYERS):
    for proj in ["q_proj", "k_proj", "v_proj", "o_proj"]:
        exclude.append(f"model.layers.{i}.self_attn.{proj}")
    for proj in ["gate_proj", "up_proj", "down_proj"]:
        exclude.append(f"model.layers.{i}.mlp.shared_experts.{proj}")

config = UnifiedConfig(
    model_path="upstage/Solar-Open-100B",
    output_dir="./Solar-Open-100B-MXFP4",
    precision="mxfp4",
    sensitivity_analysis=False,
    exclude_layers=exclude,
    trust_remote_code=True,
)
UnifiedQuantizer(config).run()

Credits

• Base Model: Upstage — Solar Open
• Quantization: quanto with AMD Quark
• Hardware: AMD MI355 (gfx950), 288 GB VRAM

License

This model inherits the Apache 2.0 License from the base model.