APEX-2

A from-scratch Mixture-of-Experts LLM: 3.87B total / 1.45B active parameters, pretrained on 86.5B tokens and instruction-tuned with SFT.

• Full code, architecture writeup and training diary: github.com/DW-dev-UE/LLM-from-scratch.
• The full benchmark writeup this card summarizes is BENCHMARK v3.
• Base model (for continued pretraining or your own fine-tuning): YOON1v/Apex-2-Base.

This checkpoint is the SFT release (pretrain → 2-stage SFT). A DPO stage was tried and dropped because it lowered code, math and instruction-following scores (see below).

Model details

Architecture	Decoder-only MoE Transformer, every layer MoE. GQA (16Q/4KV, head_dim 128) + QK-RMSNorm + RoPE (θ=10K) + SwiGLU experts + RMSNorm, weight tying
Parameters	3,869.1M total · 1,453.2M active per token (1,142.0M non-embedding active), measured
Layers · d_model	32 · 2048
Experts	16 per layer, top-4 routing with renormalized probabilities, expert width 1024
Router losses	load-balancing aux loss 0.01 (global-batch statistics) + router z-loss 1e-3
Context length	4096
Vocab	151,936 (Qwen3 tokenizer), tied embeddings
Weights	stored as bfloat16 (~7.7GB), cast from the float32 training export
Pretrain	54,250 steps · 86.5B tokens (web · code in 11 languages · math · curated), 1M-token batches; the last 20.4B tokens are an LR decay phase (web 30 / code 35 / math 15 / curated 20); parts trained with DiLoCo on 2× GH200
Post-training	SFT stage 1: 2.21B tokens (general 29 / code 44 / math 27). SFT stage 2: 0.28B tokens × 2 epochs (execution-verified code, competition math, precise instruction following)
HF format	Weights map 1:1 onto Qwen3MoeForCausalLM (the training model uses fused expert tensors; the export splits them per expert). Export check: logits relative difference 1.0e-6 and 100% next-token agreement against the training checkpoint

Usage

The model uses the ChatML template (<|im_start|>role\n...<|im_end|>) shipped in tokenizer_config.json; generation ends at <|im_end|>.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("YOON1v/Apex-2", dtype=torch.bfloat16, device_map="cuda")
tokenizer = AutoTokenizer.from_pretrained("YOON1v/Apex-2")

messages = [{"role": "user", "content": "Write a Python function that checks if a number is prime."}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to("cuda")
out = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.7, top_p=0.9)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

vLLM loads it directly as well: LLM(model="YOON1v/Apex-2", dtype="bfloat16", max_model_len=4096).

Benchmarks

Measured by us with vLLM greedy decoding and the chat template (prompt + answer ≤ 4096 tokens). Model-written code was executed in a sandbox.

Area	Benchmark	Apex-2 (SFT)	Base
Code	HumanEval	43.9	36.6
Code	HumanEval+	41.5	32.9
Code	MBPP	56.3	54.8
Code	MBPP+	48.9	46.3
Code	MultiPL-E HumanEval C++	36.0	—
Code	MultiPL-E MBPP C++	41.6	—
Code	LiveCodeBench v5–v6	3.2	—
Code	└ LiveCodeBench easy (84)	11.9	—
Code	└ LiveCodeBench medium (104)	1.0	—
Code	└ LiveCodeBench hard (154)	0.0	—
Code	CRUXEval-O	8.5	—
Code	CRUXEval-I	3.2	—
Math	GSM8K	32.4 (0-shot CoT)	15.1 (8-shot)
Math	MATH-500 (0-shot CoT)	21.0	—
Instruction following	IFEval prompt strict	44.7	—
Instruction following	IFEval instruction strict	56.6	—
Knowledge	MMLU (5-shot)	28.6	28.2
Commonsense	HellaSwag (acc_norm)	62.2	60.8
Commonsense	ARC-e (acc_norm)	64.1	66.7
Commonsense	ARC-c (acc_norm)	38.4	39.7
Commonsense	PIQA (acc_norm)	74.8	73.9
Commonsense	WinoGrande (acc)	61.8	60.1
Commonsense	LAMBADA (acc)	52.8	54.5

Compared with similar-size models (official numbers; protocols differ, so this is a rough comparison):

Model	Pretraining tokens	HumanEval	HumanEval+	MBPP	MBPP+	GSM8K	IFEval	MMLU
Apex-2 (3.87B · 1.45B active)	0.087T	43.9	41.5	56.3	48.9	32.4	44.7	28.6
Apex-1 DPO (1.1B dense)	0.02T	8.5	—	5.2	—	1.9	—	24.9
Qwen2.5-1.5B-Instruct	18T	61.6	—	63.2	—	73.2	42.5	50.7
Qwen2.5-Coder-1.5B-Instruct	5.5T	70.7	66.5	69.2	59.4	—	—	—
Qwen3-1.7B (non-thinking)	36T	—	—	—	—	—	68.2	64.4
Llama-3.2-1B-Instruct	9T	—	—	—	—	44.4	59.5*	49.3
Gemma-3-1B-it	2T	41.5	—	35.2	—	62.8	80.2*	38.8
OLMoE-1B-7B (1.3B active)	5.1T	62.3	54.4	—	—	72.4	66.4*	55.1
DeepSeek-Coder-1.3B	2T	65.9	60.4	65.3	54.8	—	—	—

_{DeepSeek-Coder numbers are the Qwen2.5-Coder report's re-evaluation. * Different IFEval metric: Apex-2 and Qwen report prompt-level strict; the others report an average of metrics or do not specify.}

With 1/23 to 1/400 of the pretraining data of these models, Apex-2's base model matches Qwen2.5-1.5B (18T tokens) on HumanEval+ (32.9 vs 32.9). The big gaps are knowledge (MMLU) and math, which mainly track pretraining scale.

SFT data was 13-gram decontaminated against HumanEval(+), MBPP(+), GSM8K test, MATH, MMLU, ARC, HellaSwag, PIQA, WinoGrande, IFEval and LAMBADA. For LiveCodeBench only the newer v5–v6 problems (after 2024-08) were used, to reduce contamination risk.

Why no DPO: DPO on allenai/Dolci-Instruct-DPO (220K pairs) made answers 2.3× longer (320 → 733 tokens on average). Chosen-answer likelihood also fell during training. In evaluation HumanEval+ went 41.5 → 32.3, MBPP+ 48.9 → 37.3, GSM8K 32.4 → 14.3 and IFEval 44.7 → 35.7, so this SFT checkpoint is the release.

Known limitations

• English-centric. Other languages, including Korean, barely work (multilingual data was excluded from SFT).
• Limited world knowledge (MMLU ~29%), so it often states wrong facts with confidence.
• Weak at competitive programming (LiveCodeBench medium ~1%) and at predicting code execution (CRUXEval).
• 4096-token context. No tool use, no thinking mode.

Training data

• Pretrain: 40 deduplicated sources, including:
  • web: FineWeb-Edu, DCLM, FinePDFs
  • code: StarCoder-family and The Stack-derived code in 11 languages, OpenCoder corpora, synthetic code
  • math: FineMath, InfiWebMath, Nemotron-CC-Math
  • curated: Wikipedia, StackExchange, peS2o, Cosmopedia, Gutenberg
• SFT:
  • general: allenai/Dolci-Instruct-SFT (tool-use, multilingual and identity subsets removed)
  • code: OpenCoder opc-sft-stage1/2, nvidia/OpenCodeInstruct (unit-test pass rate ≥ 0.9), bigcode self-oss-instruct, m-a-p/Code-Feedback
  • math: nvidia/OpenMathInstruct-2, AI-MO/NuminaMath-CoT

Some SFT sources contain synthetic data generated by third-party models (e.g. GPT-4-class, Llama-3.1-405B, Qwen2.5-Coder); check each dataset's terms for your use case.

License

Apache 2.0.