DeepSeek V4.1 · Mixture of experts

40 layers › 4 groups of 4 layers › Re-indexing layer

Mixture of experts

learnedMoE(layer_id, args) · 854–904

Each position runs 6 of 384 routed experts and one shared expert.

Why

Kept. The shared expert and the many small routed experts are kept from DeepSeekMoE; the report gives no further reason for the counts. DeepSeek V4.1 report §2.1

y^{\mathrm{routed}}=\sum_{i\in\mathcal T}G_{i}\,E_{i}

\Delta_{\mathrm{ffn}}=y^{\mathrm{routed}}+y^{\mathrm{shared}}

All 22 steps

\text{Router:}

z=x_{\mathrm{ffn}}W_{{\htmlClass{vl-stream-context}{R}}}

s=\sqrt{\operatorname{softplus}(z)}

\mathcal T=\operatorname{top}_{6}(s+b)

G_{i}=1.5\,\frac{s_{i}}{\sum_{j\in\mathcal T}s_{j}}\;(i\in\mathcal T)

\text{Routed expert:}

g=x_{\mathrm{ffn}}W_{1}

u=x_{\mathrm{ffn}}W_{3}

\bar g=\min(g,\,10)

\bar u=\operatorname{clip}(u,\,-10,\,10)

a=\bar g\,\sigma(\bar g)

h=a\odot\bar u

E_i=hW_{2}

y^{\mathrm{routed}}=\sum_{i\in\mathcal T}G_{i}\,E_{i}

\text{Shared expert:}

g=x_{\mathrm{ffn}}W_{1}

u=x_{\mathrm{ffn}}W_{3}

\bar g=\min(g,\,10)

\bar u=\operatorname{clip}(u,\,-10,\,10)

a=\bar g\,\sigma(\bar g)

h=a\odot\bar u

y^{\mathrm{shared}}=hW_{2}

\Delta_{\mathrm{ffn}}=y^{\mathrm{routed}}+y^{\mathrm{shared}}

in: $x_{\mathrm{ffn}}$ [B batch, T positions, 5120 d]
out: $\Delta_{\mathrm{ffn}}$ [B batch, T positions, 5120 d]
parameters: 13,626,901,248 · × 4 groups = 54,507,604,992

Parameters

13,626,901,248 parameters per instance, across its learned components

Exact per-instance accounting
Component	Calculation	Parameters
Router	1,966,848 across its components	1,966,848
Routed expert	$35{,}389{,}440 \times 384$	13,589,544,960
Shared expert	35,389,440 across its components	35,389,440
Per-instance total		13,626,901,248

Across 4 independent groups: $13{,}626{,}901{,}248 \times 4 = 54{,}507{,}604{,}992$ parameters.

Derived from the pinned reference implementation and configuration; equal to the element counts of the checkpoint’s own tensors on the text path. The quantisation scales, the vision encoder and the DSpark blocks are not counted.

Explanation

Every one of the 40 layers has this mixture; none has a dense feed-forward (model.py 929).

Codemodel.py 792–904 · config.json 1–66

model.py 854–904

class MoE(nn.Module):
    """Top-k routed experts plus one shared expert every token goes through. Experts are split
    across ranks, so `self.experts` is None for those another rank owns."""

    def __init__(self, layer_id: int, args: ModelArgs):
        super().__init__()
        n_routed_experts, n_activated_experts = args.get_moe_config(layer_id)
        self.layer_id = layer_id
        self.dim = args.dim
        assert n_routed_experts % world_size == 0, (
            f"Number of experts must be divisible by world size (world_size={world_size})"
        )
        self.n_routed_experts = n_routed_experts
        self.n_local_experts = n_routed_experts // world_size
        self.n_activated_experts = n_activated_experts
        self.experts_start_idx = rank * self.n_local_experts
        self.experts_end_idx = self.experts_start_idx + self.n_local_experts
        self.gate = Gate(layer_id, args)
        expert_dtype = torch.float4_e2m1fn_x2 if args.expert_dtype == "fp4" else None
        self.experts = nn.ModuleList(
            [
                Expert(
                    args.dim,
                    args.moe_inter_dim,
                    dtype=expert_dtype,
                    swiglu_limit=args.swiglu_limit,
                )
                if self.experts_start_idx <= i < self.experts_end_idx
                else None
                for i in range(self.n_routed_experts)
            ]
        )
        assert args.n_shared_experts == 1
        self.shared_experts = Expert(args.dim, args.moe_inter_dim, swiglu_limit=args.swiglu_limit)

    def forward(self, x: torch.Tensor, image_mask: torch.Tensor | None = None) -> torch.Tensor:
        shape = x.size()
        x = x.view(-1, self.dim)
        weights, indices = self.gate(x, None if image_mask is None else image_mask.flatten())
        y = torch.zeros_like(x, dtype=torch.float32)
        counts = torch.bincount(indices.flatten(), minlength=self.n_routed_experts).tolist()
        for i in range(self.experts_start_idx, self.experts_end_idx):
            if counts[i] == 0:
                continue
            expert = self.experts[i]
            idx, top = torch.where(indices == i)
            y[idx] += expert(x[idx], weights[idx, top, None])
        if world_size > 1:
            dist.all_reduce(y)
        y += self.shared_experts(x)
        return y.type_as(x).view(shape)

model.py 792–827

class Gate(nn.Module):
    """MoE gating. The correction bias steers expert selection only; the routing weights come from the
    unbiased scores. Image-span tokens use a separate bias (training `noaux_tc_for_vl`)."""

    def __init__(self, layer_id: int, args: ModelArgs):
        super().__init__()
        n_routed_experts, n_activated_experts = args.get_moe_config(layer_id)
        self.dim = args.dim
        self.topk = n_activated_experts
        self.score_func = args.score_func
        self.gate_temp = args.gate_temp
        self.norm_topk_prob = args.norm_topk_prob
        self.route_scale = args.route_scale
        self.weight = nn.Parameter(torch.empty(n_routed_experts, args.dim))
        self.bias = nn.Parameter(torch.empty(n_routed_experts, dtype=torch.float32))
        self.bias_vl = nn.Parameter(torch.empty(n_routed_experts, dtype=torch.float32)) if args.vision_enabled else None

    def forward(self, x: torch.Tensor, image_mask: torch.Tensor | None = None) -> tuple[torch.Tensor, torch.Tensor]:
        """x: [n, dim]; image_mask: [n] bool, True for tokens inside an image span."""
        scores = linear(x.float(), self.weight.float()) / self.gate_temp
        if self.score_func == "softmax":
            scores = scores.softmax(dim=-1)
        elif self.score_func == "sigmoid":
            scores = scores.sigmoid()
        else:
            scores = F.softplus(scores).sqrt()
        bias = self.bias
        if image_mask is not None and self.bias_vl is not None:
            bias = torch.where(image_mask.unsqueeze(-1), self.bias_vl, bias)
        # the bias picks experts but does not scale them: weights come from the raw scores
        indices = (scores + bias).topk(self.topk, dim=-1)[1]
        weights = scores.gather(1, indices)
        if self.norm_topk_prob and self.topk > 1:
            weights /= weights.sum(dim=-1, keepdim=True) + 1e-20  # not norm_eps, matches training
        weights *= self.route_scale
        return weights, indices

model.py 830–851

class Expert(nn.Module):
    """One SwiGLU FFN. The clamps come straight from training, where they keep fp8/fp4 activations in
    range: the up branch is clamped on both sides, the gate branch only from above."""

    def __init__(self, dim: int, inter_dim: int, dtype=None, swiglu_limit=0.0):
        super().__init__()
        self.w1 = Linear(dim, inter_dim, dtype=dtype)
        self.w2 = Linear(inter_dim, dim, dtype=dtype)
        self.w3 = Linear(dim, inter_dim, dtype=dtype)
        self.swiglu_limit = swiglu_limit

    def forward(self, x: torch.Tensor, weights: torch.Tensor | None = None) -> torch.Tensor:
        dtype = x.dtype
        gate = self.w1(x).float()
        up = self.w3(x).float()
        if self.swiglu_limit > 0:
            up = torch.clamp(up, min=-self.swiglu_limit, max=self.swiglu_limit)
            gate = torch.clamp(gate, max=self.swiglu_limit)
        x = F.silu(gate) * up
        if weights is not None:
            x = weights * x
        return self.w2(x.to(dtype))

config.json 1–66

{
    "vocab_size": 129280,
    "dim": 5120,
    "moe_inter_dim": 2304,
    "n_layers": 40,
    "n_mtp_layers": 3,
    "dspark_block_size": 5,
    "dspark_noise_token_id": 128799,
    "dspark_target_layer_ids": [37, 38, 39],
    "dspark_markov_rank": 256,
    "dspark_n_routed_experts": 128,
    "dspark_n_activated_experts": 3,
    "n_heads": 64,
    "n_routed_experts": 384,
    "n_shared_experts": 1,
    "n_activated_experts": 6,
    "score_func": "sqrtsoftplus",
    "route_scale": 1.5,
    "swiglu_limit": 10.0,
    "q_lora_rank": 1280,
    "head_dim": 512,
    "rope_head_dim": 64,
    "norm_eps": 1e-20,
    "o_groups": 8,
    "o_lora_rank": 1024,
    "window_size": 128,
    "kv_source_layers": [2, 8, 14, 20],
    "index_source_layers": [2, 8, 14, 20, 24, 28, 32, 36],
    "original_seq_len": 65536,
    "rope_theta": 10000,
    "rope_factor": 16,
    "beta_fast": 32,
    "beta_slow": 1,
    "index_n_heads": 32,
    "index_head_dim": 128,
    "index_topk": 512,
    "candidate_source_layer": 20,
    "candidate_topk_blocks": 2048,
    "candidate_block_size": 8,
    "hc_mult": 4,
    "hc_sinkhorn_iters": 20,
    "hc_eps": 1e-06,
    "engram_layer_ids": [1, 14],
    "engram_vocab_size": 16000000,
    "engram_num_embeddings": [384006168, 384016682],
    "engram_max_ngram_size": 4,
    "engram_pad_id": 2,
    "engram_compressed_vocab_size": 99092,
    "dtype": "fp8",
    "expert_dtype": "fp4",
    "compress_rope_theta": 160000,
    "compress_ratios": [0, 0, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
    "vision_n_layers": 32,
    "vision_dim": 1024,
    "vision_n_heads": 16,
    "vision_inter_dim": 2816,
    "vision_patch_size": 14,
    "vision_downsample_ratio": 3,
    "vision_max_n_token": 1024,
    "vision_min_pixels": 295936,
    "vision_max_wh_ratio": null,
    "image_token_id": 129264,
    "engram_n_heads": 8,
    "engram_head_dim": 256,
    "vision_rope_theta": 10000
}

Sources4

Direct evidence

The mixture adds the selected routed experts’ outputs and the shared expert’s; routed experts are stored in FP4.

moe · deepseek-ai/DeepSeek-V4.1-Flash / inference/model.py
Pinned revision 2cba9e42aa02 ↗ · lines 854–904

Read code excerpt

class MoE(nn.Module):
    """Top-k routed experts plus one shared expert every token goes through. Experts are split
    across ranks, so `self.experts` is None for those another rank owns."""

    def __init__(self, layer_id: int, args: ModelArgs):
        super().__init__()
        n_routed_experts, n_activated_experts = args.get_moe_config(layer_id)
        self.layer_id = layer_id
        self.dim = args.dim
        assert n_routed_experts % world_size == 0, (
            f"Number of experts must be divisible by world size (world_size={world_size})"
        )
        self.n_routed_experts = n_routed_experts
        self.n_local_experts = n_routed_experts // world_size
        self.n_activated_experts = n_activated_experts
        self.experts_start_idx = rank * self.n_local_experts
        self.experts_end_idx = self.experts_start_idx + self.n_local_experts
        self.gate = Gate(layer_id, args)
        expert_dtype = torch.float4_e2m1fn_x2 if args.expert_dtype == "fp4" else None
        self.experts = nn.ModuleList(
            [
                Expert(
                    args.dim,
                    args.moe_inter_dim,
                    dtype=expert_dtype,
                    swiglu_limit=args.swiglu_limit,
                )
                if self.experts_start_idx <= i < self.experts_end_idx
                else None
                for i in range(self.n_routed_experts)
            ]
        )
        assert args.n_shared_experts == 1
        self.shared_experts = Expert(args.dim, args.moe_inter_dim, swiglu_limit=args.swiglu_limit)

    def forward(self, x: torch.Tensor, image_mask: torch.Tensor | None = None) -> torch.Tensor:
        shape = x.size()
        x = x.view(-1, self.dim)
        weights, indices = self.gate(x, None if image_mask is None else image_mask.flatten())
        y = torch.zeros_like(x, dtype=torch.float32)
        counts = torch.bincount(indices.flatten(), minlength=self.n_routed_experts).tolist()
        for i in range(self.experts_start_idx, self.experts_end_idx):
            if counts[i] == 0:
                continue
            expert = self.experts[i]
            idx, top = torch.where(indices == i)
            y[idx] += expert(x[idx], weights[idx, top, None])
        if world_size > 1:
            dist.all_reduce(y)
        y += self.shared_experts(x)
        return y.type_as(x).view(shape)

Direct evidence

The router: the square root of softplus of a projection, a selection bias (a second one at image positions), the best six, their unbiased scores normalised and scaled.

gate · deepseek-ai/DeepSeek-V4.1-Flash / inference/model.py
Pinned revision 2cba9e42aa02 ↗ · lines 792–827

Read code excerpt

class Gate(nn.Module):
    """MoE gating. The correction bias steers expert selection only; the routing weights come from the
    unbiased scores. Image-span tokens use a separate bias (training `noaux_tc_for_vl`)."""

    def __init__(self, layer_id: int, args: ModelArgs):
        super().__init__()
        n_routed_experts, n_activated_experts = args.get_moe_config(layer_id)
        self.dim = args.dim
        self.topk = n_activated_experts
        self.score_func = args.score_func
        self.gate_temp = args.gate_temp
        self.norm_topk_prob = args.norm_topk_prob
        self.route_scale = args.route_scale
        self.weight = nn.Parameter(torch.empty(n_routed_experts, args.dim))
        self.bias = nn.Parameter(torch.empty(n_routed_experts, dtype=torch.float32))
        self.bias_vl = nn.Parameter(torch.empty(n_routed_experts, dtype=torch.float32)) if args.vision_enabled else None

    def forward(self, x: torch.Tensor, image_mask: torch.Tensor | None = None) -> tuple[torch.Tensor, torch.Tensor]:
        """x: [n, dim]; image_mask: [n] bool, True for tokens inside an image span."""
        scores = linear(x.float(), self.weight.float()) / self.gate_temp
        if self.score_func == "softmax":
            scores = scores.softmax(dim=-1)
        elif self.score_func == "sigmoid":
            scores = scores.sigmoid()
        else:
            scores = F.softplus(scores).sqrt()
        bias = self.bias
        if image_mask is not None and self.bias_vl is not None:
            bias = torch.where(image_mask.unsqueeze(-1), self.bias_vl, bias)
        # the bias picks experts but does not scale them: weights come from the raw scores
        indices = (scores + bias).topk(self.topk, dim=-1)[1]
        weights = scores.gather(1, indices)
        if self.norm_topk_prob and self.topk > 1:
            weights /= weights.sum(dim=-1, keepdim=True) + 1e-20  # not norm_eps, matches training
        weights *= self.route_scale
        return weights, indices

Direct evidence

An expert is SwiGLU with its gate and up branches clamped; the routing weight multiplies the hidden features before the down projection.

expert · deepseek-ai/DeepSeek-V4.1-Flash / inference/model.py
Pinned revision 2cba9e42aa02 ↗ · lines 830–851

Read code excerpt

class Expert(nn.Module):
    """One SwiGLU FFN. The clamps come straight from training, where they keep fp8/fp4 activations in
    range: the up branch is clamped on both sides, the gate branch only from above."""

    def __init__(self, dim: int, inter_dim: int, dtype=None, swiglu_limit=0.0):
        super().__init__()
        self.w1 = Linear(dim, inter_dim, dtype=dtype)
        self.w2 = Linear(inter_dim, dim, dtype=dtype)
        self.w3 = Linear(dim, inter_dim, dtype=dtype)
        self.swiglu_limit = swiglu_limit

    def forward(self, x: torch.Tensor, weights: torch.Tensor | None = None) -> torch.Tensor:
        dtype = x.dtype
        gate = self.w1(x).float()
        up = self.w3(x).float()
        if self.swiglu_limit > 0:
            up = torch.clamp(up, min=-self.swiglu_limit, max=self.swiglu_limit)
            gate = torch.clamp(gate, max=self.swiglu_limit)
        x = F.silu(gate) * up
        if weights is not None:
            x = weights * x
        return self.w2(x.to(dtype))

Direct evidence

The reference configuration: 40 layers of width 5,120, 64 heads of 512 features with a 64-feature rotary part, a 1,280-feature query latent, a grouped output of 8 groups at rank 1,024, a window of 128, the per-layer compression ratios and the two source lists, the indexer and its candidate stage, hyper-connections of 4 streams, 384 routed experts of which 6 are used with 1 shared, Engram before layers 1 and 14, and three DSpark blocks.

config · deepseek-ai/DeepSeek-V4.1-Flash / inference/config.json
Pinned revision 2cba9e42aa02 ↗ · lines 1–66

Read code excerpt

{
    "vocab_size": 129280,
    "dim": 5120,
    "moe_inter_dim": 2304,
    "n_layers": 40,
    "n_mtp_layers": 3,
    "dspark_block_size": 5,
    "dspark_noise_token_id": 128799,
    "dspark_target_layer_ids": [37, 38, 39],
    "dspark_markov_rank": 256,
    "dspark_n_routed_experts": 128,
    "dspark_n_activated_experts": 3,
    "n_heads": 64,
    "n_routed_experts": 384,
    "n_shared_experts": 1,
    "n_activated_experts": 6,
    "score_func": "sqrtsoftplus",
    "route_scale": 1.5,
    "swiglu_limit": 10.0,
    "q_lora_rank": 1280,
    "head_dim": 512,
    "rope_head_dim": 64,
    "norm_eps": 1e-20,
    "o_groups": 8,
    "o_lora_rank": 1024,
    "window_size": 128,
    "kv_source_layers": [2, 8, 14, 20],
    "index_source_layers": [2, 8, 14, 20, 24, 28, 32, 36],
    "original_seq_len": 65536,
    "rope_theta": 10000,
    "rope_factor": 16,
    "beta_fast": 32,
    "beta_slow": 1,
    "index_n_heads": 32,
    "index_head_dim": 128,
    "index_topk": 512,
    "candidate_source_layer": 20,
    "candidate_topk_blocks": 2048,
    "candidate_block_size": 8,
    "hc_mult": 4,
    "hc_sinkhorn_iters": 20,
    "hc_eps": 1e-06,
    "engram_layer_ids": [1, 14],
    "engram_vocab_size": 16000000,
    "engram_num_embeddings": [384006168, 384016682],
    "engram_max_ngram_size": 4,
    "engram_pad_id": 2,
    "engram_compressed_vocab_size": 99092,
    "dtype": "fp8",
    "expert_dtype": "fp4",
    "compress_rope_theta": 160000,
    "compress_ratios": [0, 0, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
    "vision_n_layers": 32,
    "vision_dim": 1024,
    "vision_n_heads": 16,
    "vision_inter_dim": 2816,
    "vision_patch_size": 14,
    "vision_downsample_ratio": 3,
    "vision_max_n_token": 1024,
    "vision_min_pixels": 295936,
    "vision_max_wh_ratio": null,
    "image_token_id": 129264,
    "engram_n_heads": 8,
    "engram_head_dim": 256,
    "vision_rope_theta": 10000
}