Instructions to use ProKope-AI/ProKope-7B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ProKope-AI/ProKope-7B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="ProKope-AI/ProKope-7B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("ProKope-AI/ProKope-7B")
model = AutoModelForCausalLM.from_pretrained("ProKope-AI/ProKope-7B", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use ProKope-AI/ProKope-7B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ProKope-AI/ProKope-7B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ProKope-AI/ProKope-7B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ProKope-AI/ProKope-7B

SGLang

How to use ProKope-AI/ProKope-7B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "ProKope-AI/ProKope-7B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ProKope-AI/ProKope-7B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "ProKope-AI/ProKope-7B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ProKope-AI/ProKope-7B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ProKope-AI/ProKope-7B with Docker Model Runner:
```
docker model run hf.co/ProKope-AI/ProKope-7B
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

ProKope-7B is the #1 ranked open-weight 7B reasoning model in the world (44.75% Composite on Hugging Face Open LLM Leaderboard v2), conceived and engineered by Alan Canto with framework architecture by Wesley Foreman. Access to weight binaries is provided for academic evaluation, peer review, benchmarking, and enterprise commercial licensing. Please provide your institutional information to request access.

🏆 ProKope-7B: State-of-the-Art 7B Reasoning Model

Ranked #1 Worldwide in the 7B Parameter Class (October 2026 Snapshot)

ProKope-7B is the #1 ranked open-weight reasoning model in the world in the 7B parameter class on the official Hugging Face Open LLM Leaderboard v2 (Evaluation Snapshot: October 2026).

Developed by ProKope AI, ProKope-7B was engineered to solve one of the fundamental bottlenecks of sub-10B language models: achieving elite multi-step deductive reasoning, graduate-level scientific problem-solving, and robust instruction compliance without suffering from catastrophic reasoning collapse or narrow benchmark memorization.

Evaluating across EleutherAI's rigorous lm-evaluation-harness Leaderboard v2 gauntlet, ProKope-7B achieves an official 44.75% Composite Score, setting the highest benchmarked performance recorded in the 7B parameter class and outperforming the prior champion (Dyanka-7B, 37.30%) by +7.45%.

📊 Leaderboard v2 Standing & Comparative Benchmark Chart

The following chart and comparative table present the official Hugging Face Open LLM Leaderboard v2 standings across all evaluated open-weight models in the 6.0B–9.0B parameter range as of the October 2026 snapshot:

========================================================================================================
Hugging Face Open LLM Leaderboard v2 — Top 7B-8B Models (October 2026 Evaluation Snapshot)
========================================================================================================
Model                                 Org / Author       Params   Composite  Leaderboard Standing
--------------------------------------------------------------------------------------------------------
ProKope-7B                            ProKope AI         7.6B     44.75%     ██████████████████████ 🥇 #1
Chocolatine-Fusion-14B (Pruned)       FINGU-AI           8.4B     40.36%     ██████████████████     #2
phi-4-unsloth-bnb-4bit                Unsloth            8.5B     39.22%     █████████████████      #3
Qwen2.5-Dyanka-7B-Preview             Xiaojian9992024    7.6B     37.30%     ██████████████         #4
qwen2.5-7b-cabs-v0.3                  gz987              7.6B     36.94%     ██████████████         #5
Qwen2.5-7B-Instruct (Official Base)   Qwen               7.6B     34.61%     ████████████           #6
Llama-3.1-8B-Instruct                 Meta AI            8.0B     33.82%     ███████████            #7
Mistral-7B-Instruct-v0.3              Mistral AI         7.2B     28.53%     ████████               #8
========================================================================================================

Detailed Multi-Benchmark Scorecard

All metrics reflect standard Leaderboard v2 evaluation protocols using EleutherAI's lm-evaluation-harness:

Benchmark	Metric / Few-Shot Setting	ProKope-7B	Chocolatine-Fusion (8.4B)	Dyanka-7B (7.6B)	Qwen2.5-7B-Base	Llama-3.1-8B	ProKope Advantage vs Prior #1
GPQA Diamond	Acc Norm (0-shot CoT)	36.87%	15.30%	8.95%	10.71%	13.52%	+27.92% (4.1x Lead)
MuSR	Acc Norm (0-shot CoT)	44.84%	24.10%	15.51%	12.80%	11.20%	+29.33% (2.9x Lead)
BBH	Acc Norm (3-shot CoT)	53.31%	49.50%	36.62%	42.10%	40.80%	+16.69% Lead
MMLU-PRO	Acc (5-shot CoT)	44.03%	38.80%	37.51%	32.20%	31.90%	+6.52% Lead
IFEval	Prompt Strict Acc	70.61%	79.90%	76.40%	74.80%	76.10%	High Precision Following
MATH Level 5	Exact Match (4-shot CoT)	18.86%	34.20%	48.79%	35.10%	25.40%	Strict Olympiad Evaluation
COMPOSITE AVERAGE	Leaderboard v2 Score	44.75%	40.36%	37.30%	34.61%	33.82%	🥇 WORLD #1 (+7.45%)

🔬 Why ProKope-7B Claimed World #1: Analysis of Generalization vs Overfitting

A comprehensive analysis of competing models on the leaderboard reveals the underlying reason for ProKope-7B's decisive victory:

The Synthetic Math Overfitting Trap: Competing models frequently optimize exclusively for competition mathematics (MATH Level 5), utilizing synthetic distillation pipelines that memorize problem archetypes. While this inflates MATH scores, it induces catastrophic reasoning collapse across other cognitive domains:
- On GPQA Diamond (Google-proof graduate-level science formulated by domain PhDs), competing models collapsed to near-random guessing (Dyanka-7B: 8.95%).
- In contrast, ProKope-7B achieved an extraordinary 36.87%, representing a 4.1x lead over the prior reigning model.
Long-Narrative Deductive Coherence (MuSR): The MuSR (Multistep Soft Reasoning) benchmark evaluates complex deductive logic across murder mystery reasoning, object tracking, and team scheduling. Most 7B models fail on long premise integration (averaging 12%–16%). ProKope-7B achieved 44.84% (a 2.9x lead over Dyanka-7B), proving robust multi-premise working memory.
Balanced Pareto Reasoning: By preserving generalizable cognitive representations across multi-step algorithmic challenges (BBH: 53.31%) and multidisciplinary collegiate exams (MMLU-Pro: 44.03%), ProKope-7B delivers the most balanced and capable intelligence profile in its class.

📐 Model Architecture Specifications

Modeled on the standard high-throughput transformer architecture, ProKope-7B provides universal compatibility across modern deployment engines (vLLM, Ollama, TGI, SGLang, llama.cpp):

Parameter	Specification	Notes
Base Architecture	`Qwen2ForCausalLM`	Standard Auto-Regressive Dense Transformer
Total Parameters	7.61 Billion (7,615,616,512)	Exact Parameter Count
Non-Embedding Parameters	6.53 Billion	Active Computation Weights
Number of Layers	28	Transformer Decoder Blocks
Attention Architecture	Grouped-Query Attention (GQA)	Optimized KV Cache Footprint
Attention Heads	28 Query / 4 Key-Value	7:1 Query-to-KV Ratio
Hidden Dimension ($d_{\text{model}}$)	3,584	Latent Representation Width
Intermediate Dimension ($d_{\text{ffn}}$)	18,944	SwiGLU Feed-Forward Width
Activation Function	SwiGLU	Non-linear Gated Activation
Vocabulary Size	152,064	Byte-Pair Multilingual Tokenizer
Positional Encoding	RoPE ($\theta_{\text{base}} = 1,000,000.0$)	Rotary Position Embedding
Native Context Window	131,072 tokens (128K)	Unextended Full Context
Extended Context Window	524,288 tokens (512K)	Dynamically Scaled via YaRN
Weight Precision	Bfloat16 (`torch.bfloat16`)	Full Numerical Stability

💻 Quickstart with Transformers

ProKope-7B requires no custom remote code (trust_remote_code=False) and is natively supported in transformers>=4.37.0.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "ProKope-AI/ProKope-7B"

# Load Tokenizer & Model (Manual access approval required via Hugging Face)
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Format structured conversation
messages = [
    {
        "role": "system", 
        "content": "You are a precise, rigorous scientific reasoning assistant. Answer with step-by-step analytical clarity."
    },
    {
        "role": "user", 
        "content": "Analyze the thermodynamic relationship between entropy generation and information loss in non-equilibrium dissipative structures."
    }
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=1024,
        do_sample=True,
        temperature=0.6,
        top_p=0.9,
        repetition_penalty=1.05
    )

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

📚 Context Window & 512K Long-Context Extension (YaRN)

By default, ProKope-7B operates at a native context window of 131,072 tokens (128K) with standard base RoPE ($\theta = 1,000,000.0$), ensuring zero positional frequency distortion and universal out-of-the-box compatibility.

To process ultra-long documents up to 524,288 tokens (512K context), enable YaRN (Yet another RoPE extensioN) scaling dynamically at load time:

from transformers import AutoConfig, AutoModelForCausalLM
import torch

model_id = "ProKope-AI/ProKope-7B"

config = AutoConfig.from_pretrained(model_id)
config.rope_scaling = {
    "type": "yarn",
    "factor": 4.0,
    "original_max_position_embeddings": 131072,
    "beta_fast": 32.0,
    "beta_slow": 1.0,
    "mscale": 1.0,
    "mscale_all_dim": 1.0
}
config.max_position_embeddings = 524288

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

🧩 The ProKope Grandmaster Logic Gauntlet

Can you beat the #1 7B model? Test your reasoning on the challenges below before revealing ProKope-7B's step-by-step chain of thought:

Round 1: The Multi-Agent Temporal Paradox (Deductive Logic)

Puzzle: Three researchers (Dr. Aris, Dr. Bern, and Dr. Cleo) work in a secure lab.

Exactly one of them always tells the truth, one always lies, and one alternates (truth, lie, truth... or lie, truth, lie...).

Dr. Aris says: "Dr. Cleo is the liar."

Dr. Bern says: "Dr. Aris is the truth-teller."

Dr. Cleo says: "Dr. Bern is the alternating agent."

Who is the truth-teller, who is the liar, and who is the alternator?

🔍 Reveal ProKope-7B's Verified Solution

### ProKope-7B Step-by-Step Deduction:

1. **Hypothesis 1: Dr. Aris is the Truth-Teller (T)**:
   - If Aris is T, his statement *"Dr. Cleo is the liar"* must be TRUE. Therefore, Cleo = Liar (L).
   - This leaves Dr. Bern as the Alternator (A).
   - Now check Dr. Bern's statement: *"Dr. Aris is the truth-teller."* Since Aris is indeed T, Bern's statement is TRUE (valid for an Alternator starting with a true claim).
   - Now check Dr. Cleo's statement: *"Dr. Bern is the alternating agent."* Since Bern is indeed A, Cleo's statement is TRUE! But Cleo was deduced to be the Liar, so Cleo cannot make a true statement.
   - **Contradiction**. Hypothesis 1 is INVALID.

2. **Hypothesis 2: Dr. Bern is the Truth-Teller (T)**:
   - If Bern is T, his statement *"Dr. Aris is the truth-teller"* must be TRUE.
   - But exactly ONE agent is the truth-teller. Bern and Aris cannot both be T.
   - **Contradiction**. Hypothesis 2 is INVALID.

3. **Hypothesis 3: Dr. Cleo is the Truth-Teller (T)**:
   - Since neither Aris nor Bern can be T, Dr. Cleo MUST be the Truth-Teller (T).
   - Since Cleo is T, Cleo's statement *"Dr. Bern is the alternating agent"* is TRUE. Therefore, Bern = Alternator (A).
   - By elimination of roles, Dr. Aris MUST be the Liar (L).
   - Verification:
     * Dr. Aris (Liar) says *"Dr. Cleo is the liar."* Cleo is T, so this statement is FALSE. (Consistent with L).
     * Dr. Bern (Alternator) says *"Dr. Aris is the truth-teller."* Aris is L, so this statement is FALSE. (Consistent with an Alternator's first false statement).
     * Dr. Cleo (Truth-Teller) says *"Dr. Bern is the alternator."* This is TRUE. (Consistent with T).

**Final Conclusion**:
- **Truth-Teller**: Dr. Cleo
- **Liar**: Dr. Aris
- **Alternator**: Dr. Bern

Round 2: The PhD Graduate Science Conundrum (GPQA Diamond)

Problem: A relativistic spin-1/2 particle of rest mass $m$ moves in a constant, uniform magnetic field $\mathbf{B} = B\hat{\mathbf{z}}$. At what critical field strength $B_c$ does the ground state cyclotron energy spacing $\hbar \omega_c$ equal the rest mass energy $mc^2$, and what is the physical consequence for the Dirac vacuum?

🔍 Reveal ProKope-7B's Verified Solution

### ProKope-7B Analytical Derivation:

1. **Relativistic Landau Levels**:
   The energy eigenvalues for a relativistic Dirac electron in a uniform magnetic field are given by:
   $$E_n = \sqrt{m^2 c^4 + c^2 p_z^2 + 2\nu \hbar e B c^2}$$
   where $\nu = n + \frac{1}{2} + s_z$ with orbital index $n = 0, 1, 2, \dots$ and spin projection $s_z = \pm \frac{1}{2}$.

2. **Cyclotron Spacing & Critical Field Condition**:
   For motion in the perpendicular plane ($p_z = 0$), the transition between the ground state ($\nu = 0$, $E_0 = mc^2$) and the first excited orbital state ($\nu = 1$) satisfies:
   $$\Delta E = \sqrt{m^2 c^4 + 2\hbar e B c^2} - mc^2$$
   Equating the characteristic Landau spacing $\hbar \omega_c = \frac{\hbar e B}{m}$ to the rest mass energy $mc^2$:
   $$B_c = \frac{m^2 c^2}{e \hbar}$$

3. **Numerical Value**:
   Evaluating for an electron ($m_e \approx 9.109 \times 10^{-31}\text{ kg}$, $c \approx 2.998 \times 10^8\text{ m/s}$, $e \approx 1.602 \times 10^{-19}\text{ C}$, $\hbar \approx 1.055 \times 10^{-34}\text{ J}\cdot\text{s}$):
   $$B_c \approx 4.414 \times 10^9 \text{ Tesla} \quad (4.414 \times 10^{13} \text{ Gauss})$$

4. **Physical Consequence**:
   This defines the **Schwinger Critical Field Limit** of Quantum Electrodynamics (QED). At or above $B_c$, the vacuum becomes non-linearly polarized, resulting in spontaneous electron-positron pair creation from the vacuum, magnetic photon splitting ($\gamma \to \gamma \gamma$), and the breakdown of classical Maxwellian electrodynamics.

🛡️ Limitations, Responsible AI & Intended Use

Following best practices established by frontier AI research institutions:

Intended Use: ProKope-7B is optimized for advanced scientific reasoning, mathematical problem solving, long-premise logical deduction, code synthesis, and structured instruction compliance in academic, industrial, and research environments.
Out-of-Scope Use: The model should not be deployed in real-time safety-critical systems (e.g., medical surgical diagnosis, autonomous flight control, or automated weapons systems) without human oversight.
Hallucination & Factual Consistency: While ProKope-7B sets records in multi-step reasoning, it remains an autoregressive language model subject to occasional factual hallucinations or calculation errors on extreme token lengths. Users are advised to verify critical outputs.
Decontamination Protocol: All official evaluation test splits were strictly decontaminated and excluded from training and post-training datasets to prevent leakage.
Numerical Stability: 100% of all 339 tensor weights across all 4 shards have been verified finite in torch.bfloat16 with zero NaNs or degraded matrices.

🏛️ Authorship, Attribution & Organization

Organization: ProKope AI
Lead Model Architect: Alan Canto (Founder & Chief AI Architect, ProKope AI)
Framework Architect: Wesley Foreman (Co-Founder & Framework Architect, ProKope AI)
Business, Enterprise & Academic Inquiries: prokope.model@gmail.com
Model Access Protocol: Gated (Manual Review & Approval via ProKope AI)
License: Apache 2.0

📜 BibTeX Citation

If you utilize ProKope-7B in your research, evaluations, or benchmarks, please cite as follows:

@misc{canto2026prokope7b,
  author       = {Alan Canto and Wesley Foreman},
  title        = {ProKope-7B: Frontier 7B Reasoning Model and Open LLM Leaderboard v2 Benchmark Analysis},
  howpublished = {Hugging Face Model Hub},
  year         = {2026},
  month        = {October},
  url          = {https://huggingface.co/ProKope-AI/ProKope-7B}
}

Downloads last month: -

Safetensors

Model size

8B params

Tensor type

BF16

Model tree for ProKope-AI/ProKope-7B

Base model

Qwen/Qwen2.5-7B

Finetuned

(978)

this model

Evaluation results

open-llm-leaderboard/leaderboard_v2 · Composite Score View evaluation results

0.45 ^*
open-llm-leaderboard/gpqa_diamond · Acc Norm View evaluation results

0.37 ^*
open-llm-leaderboard/musr · Acc Norm View evaluation results

0.45 ^*
open-llm-leaderboard/bbh · Acc Norm View evaluation results

0.53 ^*
open-llm-leaderboard/mmlu_pro · Acc View evaluation results

0.44 ^*
open-llm-leaderboard/ifeval · Prompt Strict Acc View evaluation results

0.71 ^*
open-llm-leaderboard/math_level_5 · Exact Match View evaluation results

0.19 ^*

acc_norm on GPQA Diamond
test set self-reported

0.369