37 tok/s from Qwen3-Next-80B (48 GB) on a 32 GB laptop, 24 without drafts vs 2.8 for llama.cpp; reading 62 vs 16 tok/s. Faster than llama.cpp on all 13 models compared in Q4_K_M. LLMs bigger than your RAM, pure C, CPU or GPU.
C avx2ccpu-inferenceggufinference-enginelinuxllmlocal-llmmixture-of-expertsmoenvmequantizationqwen3simdspeculative-decodingvulkan
/* SPDX-License-Identifier: GPL-3.0-or-later
Copyright (C) 2026 Maurizio Cammalleri */
/*
* arch.h - what sets one model architecture apart from another, as data.
*
* The forward pass is written once; each architecture says which of its
* parts it has. A new family is a row of the table in arch.c, plus the code
* of whatever it has that no other family had.
*/
#ifndef JANAS_LLM_ARCH_H
#define JANAS_LLM_ARCH_H
/* The recurrent layers, if any: Gated DeltaNet in two layouts. */
enum janas_rec {
JANAS_REC_NONE, /* attention in every layer */
JANAS_REC_NEXT, /* qwen3next: q|k|v|z and b|a from two fused matrices */
JANAS_REC_Q35, /* qwen35: q|k|v, z, alpha, beta from four */
JANAS_REC_MAMBA2 /* nemotron_h_moe: Mamba-2 (model_mamba.c) */
};
struct janas_arch {
const char *name; /* general.architecture */
enum janas_rec rec; /* DeltaNet layers, full_attention_interval apart */
int attn_gate; /* the query projection carries an output gate */
int shared_expert; /* a shared expert with a sigmoid gate, per layer */
int post_attn_norm; /* the FFN norm is post_attention_norm, not ffn_norm */
int vmap_mod; /* value head h shares key head h % n_kh, else h / rf */
int nextn_in_file; /* the file's last layers are the MTP block */
/* Gemma 4 */
int gelu; /* the feed-forward's activation is GELU (tanh), not SiLU */
int sandwich; /* attention and FFN outputs normalized before the
residual (post_attention_norm, post_ffw_norm), the
embedding scaled by sqrt(d_model), each layer's output
by layer_output_scale if the file has it */
int swa; /* sliding-window layers beside full ones, each kind with
its head size, KV heads and RoPE; V normalized without
weights, V = K where attn_v is missing, scores not
scaled (the q and k norms do it) */
int kv16; /* keys and values kept at sixteen bits, not eight */
/* Llama 3: no norm on the heads of q and k; the file's queries and keys
in the checkpoint's row order (gguf2jns), rotated by halves */
int no_qk_norm;
/* Qwen3.8-Flash-Next (qwen4exp, model_hc.c): hyper-connections - the
residual in several streams, mixed into each sublayer's input by a
gate of low rank, in place of its norm, and each sublayer's output
added to every stream with a weight of its own - n-gram embeddings
at the layers the file names, the DeltaNet's output gated by a
sigmoid instead of SiLU, and the attention's sparse selection */
int hc;
/* Nemotron-H (nemotron_h_moe, model_nh.c): each layer one part, its
norm and residual of its own - a Mamba-2 mixer, an attention without
RoPE or a mixture - the mixture's experts in a latent space (down
from d_model before them, up after), relu^2 without a gate, routed
by sigmoids with a bias for the choice */
int nh;
};
/* The architecture called name, or NULL if Janas does not know it. */
const struct janas_arch *janas_arch_find(const char *name, unsigned len);
#endif