Kimi K2 · Down projection · OpenModelAtlas

61 transformer blocks › 60 MoE blocks › Mixture of experts › Shared expert

Down projection

learnedself.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))

Map 2,048 gated features back to 7,168 features.

{\htmlData{light=output down}{y^{\mathrm{shared}}}}={\htmlData{light=h product}{h}}W_{2}

in: $h$ [B batch, T positions, 2048 shared d_ff]
out: $y^{\mathrm{shared}}$ [B batch, T positions, 7168 d]
parameters: 14,680,064 · × 60 MoE blocks = 880,803,840

Parameters

14,680,064 parameters per instance.

Exact per-instance accounting
Component	Calculation	Parameters
Weight matrix	$2{,}048 \times 7{,}168$	14,680,064
Per-instance total		14,680,064

This projection has no bias parameter.

The reference module holds the weight output by input, [out, in].

Across 60 independent MoE blocks: $14{,}680{,}064 \times 60 = 880{,}803{,}840$ parameters.

Derived from the configured matrix dimensions; the model total matches the checkpoint’s tensors.

Codemodeling_deepseek.py 374–390 · config.json 38–46

modeling_deepseek.py 374–390

class DeepseekV3MLP(nn.Module):
    def __init__(self, config, hidden_size=None, intermediate_size=None):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size if hidden_size is None else hidden_size
        self.intermediate_size = (
            config.intermediate_size if intermediate_size is None else intermediate_size
        )

        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x):
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj

config.json 38–46

  "quantization_config": {
    "activation_scheme": "dynamic",
    "fmt": "e4m3",
    "quant_method": "fp8",
    "weight_block_size": [
      128,
      128
    ]
  },

Sources2

Direct evidence

The dense feed-forward network is a SwiGLU MLP: down(silu(gate x) ⊙ up x).

mlp · moonshotai/Kimi-K2-Instruct / modeling_deepseek.py
Pinned revision fd1984e2b7a3 ↗ · lines 374–390

Read code excerpt

class DeepseekV3MLP(nn.Module):
    def __init__(self, config, hidden_size=None, intermediate_size=None):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size if hidden_size is None else hidden_size
        self.intermediate_size = (
            config.intermediate_size if intermediate_size is None else intermediate_size
        )

        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x):
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj

Direct evidence

Weights are stored in FP8 (e4m3) with one scale per 128 × 128 block; the scales are quantization metadata and are excluded from parameter counts.

linear · moonshotai/Kimi-K2-Instruct / config.json
Pinned revision fd1984e2b7a3 ↗ · lines 38–46

Read code excerpt

  "quantization_config": {
    "activation_scheme": "dynamic",
    "fmt": "e4m3",
    "quant_method": "fp8",
    "weight_block_size": [
      128,
      128
    ]
  },