Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
{
"model_type": "kimi_linear",
"architectures": [
"KimiLinearForCausalLM"
],
"hidden_size": 128,
"num_hidden_layers": 6,
"vocab_size": 320,
"first_k_dense_replace": 2,
"intermediate_size": 64,
"num_attention_heads": 4,
"num_key_value_heads": 4,
"q_lora_rank": 32,
"kv_lora_rank": 32,
"qk_nope_head_dim": 16,
"qk_rope_head_dim": 8,
"v_head_dim": 16,
"num_experts": 8,
"num_experts_per_token": 2,
"moe_intermediate_size": 32,
"routed_expert_hidden_size": 32,
"num_shared_experts": 1,
"attn_res_block_size": 2,
"hidden_act": "situ",
"activation_situ_beta": 1.0,
"activation_situ_linear_beta": 1.0,
"moe_router_activation_func": "sigmoid",
"topk_method": "noaux_tc",
"use_grouped_topk": true,
"num_expert_group": 1,
"topk_group": 1,
"moe_renormalize": true,
"moe_layer_freq": 1,
"routed_scaling_factor": 1.0,
"latent_moe_use_norm": true,
"mla_use_nope": true,
"mla_use_output_gate": true,
"rms_norm_eps": 1e-05,
"rope_theta": 10000.0,
"max_position_embeddings": 256,
"tie_word_embeddings": false,
"linear_attn_config": {
"num_heads": 2,
"head_dim": 16,
"short_conv_kernel_size": 4,
"kda_layers": [
1,
3
],
"full_attn_layers": [
2,
4,
5,
6
],
"use_full_rank_gate": true,
"gate_lower_bound": -5.0
},
"bos_token_id": 0,
"eos_token_id": 1
}