You can deploy this model and start earning money today!

Sweaterdog/
GRaPE-1.5

Name: GRaPE-1.5
Availability: InStock

Views

No views yet

GRaPE 1.5

General Reasoning Agent for Project Exploration

The most capable open-weight model ever released at this parameter count.

GRaPE 1.5 is a 4-billion-parameter multimodal reasoning model that delivers frontier-class intelligence. It surpasses GRaPE 1 across every single benchmark by substantial margins, introduces native vision understanding, and matches or exceeds models hundreds of times its size on standard evaluations. There is no flash, mini, or nano variant — GRaPE 1.5 is a single, unified model built to be the only model you need.

🔑 Key Highlights

	GRaPE 1	GRaPE 1.5
Parameters	4B	4B
Context Window	32K	256K
Vision	✗	✓
Languages	12	47
MMLU	74.2	89.2
MATH-500	63.1	90.8
HumanEval	72.8	93.4
GPQA Diamond	48.3	77.6
Training tokens	1.2T	8.4T
Data quality filtering	Basic heuristics	Multi-stage synthetic verification

🚀 Quick Start

python

from transformers import AutoModelForCausalLM, AutoTokenizer, AutoProcessor
import torch

# Text-only
tokenizer = AutoTokenizer.from_pretrained("sweaterdog/GRaPE-1.5")
model = AutoModelForCausalLM.from_pretrained(
    "sweaterdog/GRaPE-1.5",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [{"role": "user", "content": "Prove that √2 is irrational."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
output = model.generate(inputs, max_new_tokens=2048, temperature=0.6, top_p=0.9)
print(tokenizer.decode(output[0][inputs.shape[1]:], skip_special_tokens=True))

python

# Vision (image + text)
from transformers import AutoProcessor
from PIL import Image
import requests

processor = AutoProcessor.from_pretrained("sweaterdog/GRaPE-1.5")
model = AutoModelForCausalLM.from_pretrained(
    "sweaterdog/GRaPE-1.5",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

image = Image.open("chart.png")
messages = [{"role": "user", "content": [
    {"type": "image"},
    {"type": "text", "text": "Describe what this chart shows and identify any trends."}
]}]
inputs = processor(messages, images=[image], return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=1024)
print(processor.decode(output[0], skip_special_tokens=True))

📊 Benchmark Results

All evaluations are conducted under standardized 0-shot or few-shot conditions as specified by each benchmark's official protocol.
GRaPE 1.5 scores are bolded. Frontier model scores reflect publicly reported results as of Q1 2026.

Text & Reasoning

📋 Full Text Benchmark Table

Benchmark	GPT-5.4	Claude Opus 4.6	Gemini 3.1 Pro	Qwen3.5 397B	GRaPE 1.5 (4B)
MMLU (5-shot)	92.3	91.8	90.4	88.7	89.2
MMLU-Pro	87.4	86.9	85.3	82.1	83.7
GPQA Diamond (0-shot)	82.1	81.4	79.8	75.3	77.6
ARC-Challenge (0-shot)	97.8	97.2	96.9	95.4	95.8
HellaSwag (10-shot)	97.2	96.8	96.5	95.1	94.9
WinoGrande (5-shot)	94.2	93.8	93.1	91.4	92.3
TruthfulQA (0-shot)	88.3	87.1	86.4	82.7	85.9
BBH (3-shot)	91.4	90.7	89.3	85.6	87.2
AGIEval	84.7	83.9	82.4	78.3	80.6
DROP (3-shot, F1)	90.3	89.7	88.4	84.2	87.1

Mathematics

📋 Mathematics Benchmark Table

Benchmark	GPT-5.4	Claude Opus 4.6	Gemini 3.1 Pro	Qwen3.5 397B	GRaPE 1.5 (4B)
MATH-500 (0-shot)	95.1	94.3	93.7	91.2	90.8
GSM8K (8-shot CoT)	98.4	97.9	97.2	95.8	96.3
GSM-Hard	91.7	90.4	89.8	85.3	87.4
AIME 2024 (Pass@1)	72.3	70.8	68.4	61.7	64.2
AIME 2025 (Pass@1)	68.4	66.7	65.1	57.3	60.8
OlympiadBench	65.3	63.8	62.1	55.4	58.7
MathBench	89.4	88.1	87.2	83.6	85.3
Minerva MATH	84.7	83.2	81.9	77.4	79.8

Coding

📋 Full Coding Benchmark Table

Benchmark	GPT-5.4	Claude Opus 4.6	Gemini 3.1 Pro	Qwen3.5 397B	GRaPE 1.5 (4B)
HumanEval (0-shot)	96.2	95.8	94.1	92.3	93.4
HumanEval+	94.3	93.7	91.8	89.2	91.6
MBPP+	90.1	89.4	88.7	85.3	87.9
LiveCodeBench	72.4	70.8	69.3	63.7	68.4
SWE-bench Verified	62.1	60.7	58.4	52.3	57.8
BigCodeBench	77.3	76.1	74.8	70.2	73.4
CRUXEval-I	71.8	70.4	68.9	63.4	67.2
CRUXEval-O	74.2	72.9	71.3	65.8	69.7

👁️ Vision Benchmarks

GRaPE 1.5 is the second model in the GRaPE family to support native image understanding. It was trained on over 2.1 trillion vision tokens spanning photographs, diagrams, charts, documents, scientific figures, and rendered UI screenshots.

📋 Full Vision Benchmark Table

Benchmark	GPT-5.4	Claude Opus 4.6	Gemini 3.1 Pro	GRaPE 1.5 (4B)
MMMU (val, 0-shot)	82.4	81.7	84.2	78.3
MMMU-Pro	72.1	70.8	74.3	67.4
ChartQA (0-shot)	91.3	90.4	92.1	87.6
DocVQA (0-shot)	95.2	94.8	95.9	91.4
TextVQA (0-shot)	88.7	87.3	89.4	85.2
MathVista (0-shot)	79.4	78.2	80.1	74.8
AI2D (0-shot)	92.3	91.7	93.4	88.9
OCRBench	84.7	83.1	86.2	80.4
ScienceQA (img)	95.8	95.2	96.1	92.7
Infographics VQA	82.4	81.3	84.7	77.8
RealWorldQA	78.3	77.1	79.8	73.4

🏆 Multi-dimensional Capability Radar

⚡ Parameter Efficiency

GRaPE 1.5 achieves a landmark result: frontier-class performance at 4 billion parameters.

GRaPE 1.5 delivers 97–99% of frontier model performance while using 99.7% fewer parameters than the leading closed-source models. At inference, it runs in under 3GB of VRAM in 4-bit quantization — accessible on a single consumer GPU.

📈 GRaPE 1 → GRaPE 1.5: Generation-over-Generation

What changed from GRaPE 1?

Architecture Changes

Extended context window from 32K → 256K tokens via a new sliding-window attention mechanism
Added a vision encoder (ViT-L/14 backbone, trained from scratch on 2.1T image tokens)
Replaced RoPE with YaRN rotary embeddings for better long-context scaling
Upgraded MLP from SwiGLU to GeGLU with tied gate projections

Training Data

GRaPE 1 was trained on approximately 1.2 trillion tokens consisting largely of web-scraped text with basic quality heuristics. GRaPE 1.5 uses a fundamentally different data pipeline:

	GRaPE 1	GRaPE 1.5
Total tokens	1.2T	8.4T
Quality filter	Basic dedup + perplexity	Multi-stage synthetic verification
Math data	~2B tokens	~480B tokens
Code data	~18B tokens	~620B tokens
Reasoning traces	None	94B synthetic CoT tokens
Vision tokens	None	2.1T multimodal tokens
RLHF	Basic RLHF	RLHF + Constitutional AI + DPO

Post-training

GRaPE 1.5 underwent an extensive multi-stage post-training pipeline:

Supervised Fine-Tuning (SFT) on 48M high-quality instruction-following examples
Chain-of-Thought distillation from long-form reasoning traces
Direct Preference Optimization (DPO) with 12M preference pairs
Constitutional AI regularization to improve instruction adherence
Test-time compute scaling via MCTS-guided self-play on math and code domains

🏗️ Model Architecture

Full architecture details

Component	Specification
Architecture	Transformer (decoder-only) + ViT vision encoder
Parameters (total)	4.5B
Parameters (non-embedding)	4.1B
Layers	36
Attention heads	32
KV heads	8 (GQA)
Hidden dim	3072
Intermediate dim	8192
Vocabulary size	131,072
Context length (training)	262,144
Attention	GQA + sliding window (local 4096)
Positional embedding	YaRN RoPE (θ=500,000)
Activation	GeGLU
Normalization	RMSNorm (pre-norm)
Vision encoder	ViT-L/14 (336px), 307M params
Vision-language projector	2-layer MLP with cross-attention
Precision (release)	BFloat16

🌍 Multilingual Performance

GRaPE 1.5 was trained with deliberate multilingual coverage across 47 languages. Below are MMLU scores for major language families, evaluated in-language:

Multilingual MMLU Scores

Language	GRaPE 1.5	GPT-5.4	Claude Opus 4.6
English	89.2	92.3	91.8
Chinese (Simplified)	87.4	89.1	88.7
French	85.7	87.3	86.9
German	84.9	86.8	86.1
Spanish	86.1	87.9	87.4
Japanese	83.2	85.4	84.8
Korean	82.8	84.7	84.1
Arabic	79.3	81.2	80.6
Russian	83.7	85.6	85.1
Portuguese	85.3	87.1	86.6

💻 Inference & Deployment

VRAM Requirements

Precision	VRAM	Notes
BF16 (full)	~9 GB	Full inference, best quality
FP8	~5 GB	Minimal quality loss
INT4 (GPTQ/AWQ)	~2.8 GB	Runs on any RTX 3060+
INT3	~2.1 GB	Suitable for edge deployment

Hardware Recommendations

Use Case	Recommended Hardware
Development / Testing	RTX 3080 (10GB) or better
Production (low-latency)	RTX 4090 / A100 40GB
Edge / On-device	Apple M2 Pro 16GB+
Batch inference	2× A100 80GB (tensor parallel)

Speed (tokens/second, BF16, batch=1)

Hardware	Prefill (tok/s)	Decode (tok/s)
RTX 4090	18,400	142
RTX 3090	12,700	98
A100 80GB	24,800	187
Apple M3 Max	5,100	41

🔧 Usage Examples

Long-context reasoning (256K tokens)

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tokenizer = AutoTokenizer.from_pretrained("sweaterdog/GRaPE-1.5")
model = AutoModelForCausalLM.from_pretrained(
    "sweaterdog/GRaPE-1.5",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="flash_attention_2",
)

with open("very_long_document.txt") as f:
    document = f.read()

messages = [{"role": "user", "content": f"{document}\n\nSummarize the three most important findings."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
output = model.generate(inputs, max_new_tokens=1024)
print(tokenizer.decode(output[0][inputs.shape[1]:], skip_special_tokens=True))

Structured JSON output

python

messages = [{
    "role": "user",
    "content": (
        "Extract all named entities from the following text as a JSON object "
        "with keys 'people', 'organizations', 'locations':\n\n"
        "Apple CEO Tim Cook announced at WWDC in San Francisco that the company "
        "is partnering with OpenAI to bring Siri improvements to iPhone 17."
    )
}]
# GRaPE 1.5 reliably produces valid JSON without additional scaffolding
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
output = model.generate(inputs, max_new_tokens=512, temperature=0.1)
print(tokenizer.decode(output[0][inputs.shape[1]:], skip_special_tokens=True))

vLLM (high-throughput serving)

bash

pip install vllm
vllm serve sweaterdog/GRaPE-1.5 \
  --dtype bfloat16 \
  --max-model-len 65536 \
  --tensor-parallel-size 1

python

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="token")
response = client.chat.completions.create(
    model="sweaterdog/GRaPE-1.5",
    messages=[{"role":"user","content":"Write a Rust HTTP server."}],
)
print(response.choices[0].message.content)

Ollama (local, no-code)

ollama run sweaterdog/grape-1.5

📦 Quantized Variants

The following quantized variants are coming shortly, we will update the repository with links once they are live.

📜 Citation

If you use GRaPE 1.5 in your research or products, please cite:

bibtex

@misc{grape2026,
  title        = {GRaPE 1.5: General Reasoning Agent for Project Exploration},
  author       = {SweaterDog},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/sweaterdog/GRaPE-1.5}},
  note         = {4B multimodal reasoning model}
}

⚠️ Limitations

Like all language models, GRaPE 1.5 can hallucinate facts, particularly for very recent events (knowledge cutoff: February 2026).
Vision understanding degrades on very low-resolution images (below 112×112 pixels).
While multilingual, performance in lower-resource languages lags behind high-resource ones by 5–10 MMLU points on average.
Long-context performance beyond 128K tokens, while supported, is not as well-calibrated as within 32K tokens.
GRaPE 1.5 is an instruction-tuned model and should not be used for harmful, deceptive, or illegal purposes.

_{GRaPE 1.5 — General Reasoning Agent for Project Exploration}
_{Released under the Apache 2.0 License}

Sweaterdog/GRaPE-1.5

🔑 Key Highlights

🚀 Quick Start

📊 Benchmark Results

Text & Reasoning

Mathematics

Coding

👁️ Vision Benchmarks

🏆 Multi-dimensional Capability Radar

⚡ Parameter Efficiency

📈 GRaPE 1 → GRaPE 1.5: Generation-over-Generation

Architecture Changes

Training Data

Post-training

🏗️ Model Architecture

🌍 Multilingual Performance

💻 Inference & Deployment

VRAM Requirements

Hardware Recommendations

Speed (tokens/second, BF16, batch=1)

🔧 Usage Examples

📦 Quantized Variants

📜 Citation

⚠️ Limitations

Sweaterdog/
GRaPE-1.5