🌿 Nori-2: High-Potency Reasoning Model (Gemma 4 E4B)

Nori-2 is an advanced compact reasoning model built on top of google/gemma-4-E4B-it.

Unlike conventional reasoning models that undergo massive reinforcement learning (RL) runs or extensive instruction tuning directly on the target architecture, Nori-2 was produced without traditional training.

Instead, Nori-2 was created using **Isometric Manifold Transport (IMT)**—a framework that directly transplants the post-trained reasoning manifolds of advanced Qwen models into the geometric coordinate space of Gemma 4.

---

🔬 Architectural Upgrades

Nori-2 incorporates three mathematical upgrades:

1. Riemannian Null-Space Completion: Eliminates the 512-dimension orthogonal blind spot between Qwen (2,048) and Gemma (2,560) using covariance regression ($\mathcal{K}$), allowing reasoning updates to flow across 100% of Gemma's 2,560-dimensional residual stream.
2. Spectral Wiener Filtering (Rank 96): Replaces harsh SVD cutoffs with quadratic soft-shrinkage, dampening high-frequency coordinate phase noise while preserving dense math circuits.
3. Target-Norm Calibration (2.5%): Calibrates every layer’s update to exactly 2.5% of its native baseline Frobenius norm, preventing attention-head saturation.

---

📊 Benchmark Results

Evaluated across 13,524 total generation trajectories on the full test sets using vLLM with FP8 KV-caching:
temperature = 0.6, top_p = 0.95, max_tokens = 8192, N = 4 Rollouts.

Benchmark	Total Problems	Single-Pass (Pass@1)	Majority Vote (Consensus)	Search Ceiling (Pass@4)
AIME	90	40.56%	43.33% (39 / 90)	53.33% (48 / 90)
MATH-500	500	54.85%	57.00% (285 / 500)	59.20% (296 / 500)
ARC-Challenge	1,172	94.03%	94.71% (1,110 / 1,172)	97.27% (1,140 / 1,172)
SVAMP	300	84.50%	86.33% (259 / 300)	89.00% (267 / 300)
GSM8K	1,319	79.15%	79.30% (1,046 / 1,319)	84.00% (1,108 / 1,319)

---

⚔️ Benchmark Comparisons

1. Comparison Across the Nori Evolution & Base Gemma 4

Benchmark	Base Gemma 4 E4B	Nori-1	Nori-2	Net Gain over Base
MATH-500	32.50%	54.00%	57.00%	+24.50%
AIME (Consensus)	~25.00%	40.00%	43.33%	+18.33%
AIME (Pass@4)	~30.00%	—	53.33%	+23.33%
ARC-Challenge	84.20%	94.45%	94.71%	+10.51%
SVAMP	76.50%	82.67%	86.33%	+9.83%
GSM8K	76.30%	79.45%	79.30%	+3.00%

---

2. Comparison with Modern Reasoning Models (Majority Vote Consensus)

All models compared using their reported Majority Vote / Consensus metrics:

Model	Parameters	AIME (Consensus)	MATH-500	ARC-Challenge	SVAMP	GSM8K
Nori-2	~4B	43.33%	57.00%	94.71%	86.33%	79.30%
Nanbeige4.2-3B	3B	48.20%	67.60%	88.50%	85.00%	92.70%
Nanbeige4.1-3B	3B	38.60%	51.20%	86.40%	81.50%	85.90%
VibeThinker-3B	3B	94.30%*	83.40%	Not Reported	Not Reported	91.80%
Qwen-2.5-3B-Instruct	3B	3.30%	38.50%	78.50%	78.00%	79.20%
Gemma 4 E4B (Base)	~4B	~25.00%	32.50%	84.20%	76.50%	76.30%
Llama-3.2-3B-Instruct	3B	1.10%	26.20%	74.10%	72.50%	68.80%

*VibeThinker's 94.30% AIME score was achieved on AIME 2026 using its official sampling configuration.

---

🚀 Quickstart & Inference

Using vLLM (High Throughput)

from vllm import LLM, SamplingParams

llm = LLM(
    model="BIBLIOKLEPT/Nori-2",
    max_model_len=16384,
    dtype="bfloat16",
    kv_cache_dtype="fp8"
)

sampling_params = SamplingParams(
    temperature=0.6,
    top_p=0.95,
    max_tokens=8192,
    skip_special_tokens=False
)

outputs = llm.generate(
    ["A farmer has 17 sheep, and all but 9 die. How many sheep are left?"], 
    sampling_params
)

print(outputs[0].outputs[0].text)

Using Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "BIBLIOKLEPT/Nori-2"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

prompt = "A farmer has 17 sheep, and all but 9 die. How many sheep are left?"
conv = [{"role": "user", "content": prompt}]

formatted = tokenizer.apply_chat_template(
    conv, 
    tokenize=False, 
    add_generation_prompt=True, 
    enable_thinking=True
)
inputs = tokenizer(formatted, return_tensors="pt").to("cuda")

outputs = model.generate(
    **inputs,
    max_new_tokens=2048,
    temperature=0.6,
    top_p=0.95,
    do_sample=True
)

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=False)
print(response)

---

📜 Citation & Attribution

@misc{nori2_2026,
  author = {BIBLIOKLEPT},
  title = {Nori-2: High-Potency Reasoning Model via Isometric Manifold Transport and Null-Space Completion},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face Hub},
  howpublished = {https://huggingface.co/BIBLIOKLEPT/Nori-2}
}