Run 35B MoE models in ~2.5 GB of RAM. Weights stream from SSD; a trained router predicts loads a step ahead. Mac, iPhone, Android, Windows supported
Python
// Runtime routing facts that are not in config.json (config is only a cross-check).
#pragma once
#include <cstdint>
#include <string>
#include <vector>
namespace e0n {
enum class RouterFamily { kSoftmaxTopk, kSigmoidGroup };
struct TierRouting {
std::string tier; // "edge0-8b" / "edge0-35b" (= tier_id)
RouterFamily family;
int64_t num_experts;
int top_k; // runtime (35b=4, not config's 8)
bool norm_topk_prob; // true on both tiers
int n_group, topk_group;
double routed_scaling;
bool expert_bias; // live gate scores include bias; head path does not
int num_layers;
int first_head_consumer; // 35b 7 / 8b 8
int last_head_consumer; // 35b 38 / 8b 23 (35b layer 39 is not a consumer)
int first_owner, last_owner; // 35b 6/38 / 8b 7/22
int dense_prefix; // leading dense (no MoE) layers: 8b=1 (L0); 35b=0
int prerouter_hidden; // 512 on both live-head tiers
std::string head_file;
std::string lora_file;
bool heads_baked_in_model;
// Head-feature one-hot: false = executed set (qwen); true = teacher top-k (ling).
bool head_feature_teacher;
int prefill_chunk; // 2048 both wrappers
int hot_window; // 35b=4 / 8b=1
int prefill_hot; // 35b staged_k4=32 / 8b prod_k8=0
bool full_layer_prefill; // 35b=false / 8b=true
bool async_eval_prefill; // = prefill_multi and full_layer
std::vector<int> predicted() const;
std::vector<int> routed() const; // decode-time live-router layers, ascending
};
bool tier_routing(const std::string &tier_id, TierRouting &out);
// Empty string = match; non-empty = readable mismatch (caller decides whether to reject).
std::string routing_config_mismatch(const TierRouting &r, int64_t cfg_num_experts,
int64_t cfg_num_layers);
} // namespace e0n