Instructions to use Ueuee82683/Puma-1.1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Ueuee82683/Puma-1.1 with Transformers:

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

pipe = pipeline("text-generation", model="Ueuee82683/Puma-1.1")
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("Ueuee82683/Puma-1.1")
model = AutoModelForCausalLM.from_pretrained("Ueuee82683/Puma-1.1", 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

llama.cpp

How to use Ueuee82683/Puma-1.1 with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Ueuee82683/Puma-1.1:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Ueuee82683/Puma-1.1:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Ueuee82683/Puma-1.1:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Ueuee82683/Puma-1.1:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf Ueuee82683/Puma-1.1:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf Ueuee82683/Puma-1.1:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf Ueuee82683/Puma-1.1:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Ueuee82683/Puma-1.1:Q4_K_M

Use Docker

docker model run hf.co/Ueuee82683/Puma-1.1:Q4_K_M

LM Studio
Jan

vLLM

How to use Ueuee82683/Puma-1.1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Ueuee82683/Puma-1.1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Ueuee82683/Puma-1.1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Ueuee82683/Puma-1.1:Q4_K_M

SGLang

How to use Ueuee82683/Puma-1.1 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 "Ueuee82683/Puma-1.1" \
    --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": "Ueuee82683/Puma-1.1",
		"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 "Ueuee82683/Puma-1.1" \
        --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": "Ueuee82683/Puma-1.1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use Ueuee82683/Puma-1.1 with Ollama:
```
ollama run hf.co/Ueuee82683/Puma-1.1:Q4_K_M
```
Unsloth Desktop
Docker Model Runner
How to use Ueuee82683/Puma-1.1 with Docker Model Runner:
```
docker model run hf.co/Ueuee82683/Puma-1.1:Q4_K_M
```

Lemonade

How to use Ueuee82683/Puma-1.1 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Ueuee82683/Puma-1.1:Q4_K_M

Run and chat with the model

lemonade run user.Puma-1.1-Q4_K_M

List all available models

lemonade list

Atomic Chat

Puma 1.1

Puma 1.1 is a 6.96-billion-parameter (6,955,113,984 parameters) open-weight chain-of-thought reasoning and software engineering model created by Merkium and released under the Apache 2.0 license. Built on a 28-layer LlamaForCausalLM decoder (hidden_size = 3584, intermediate_size = 19200, 7:1 Grouped-Query Attention, 32K RoPE) paired with the 49,152-token Puma byte-level BPE vocabulary, Puma 1.1 thinks step-by-step inside explicit <think>...</think> blocks before emitting verified code, mathematical derivations, and technical answers.

Ollama Registry: Merkium/puma-1.1
Hugging Face Repository: Ueuee82683/Puma-1.1
Lightweight Sibling (1.42B): Ueuee82683/Puma-1.0 / Merkium/puma-1.0

Architecture & Specifications

Parameter	Value
Architecture	Pre-norm Transformer decoder (`LlamaForCausalLM`, `arch: llama`)
Total Parameters	`6,955,113,984` (`6.96B`, untied input/output embeddings)
Reasoning Format	Structured `<think>...</think>` chain-of-thought traces prior to final output
Layers (`num_hidden_layers`)	`28`
Hidden Dimension (`hidden_size`)	`3584`
FFN Dimension (`intermediate_size`)	`19200` (`5.36x` SwiGLU MLP, `SiLU` gate, `256`-aligned super-blocks)
Attention Heads	`28` Query heads / `4` Key-Value heads (GQA 7:1, `head_dim = 128`)
Positional Encoding	Rotary Position Embeddings (`RoPE`, `theta = 1,000,000.0`)
Normalization	`RMSNorm` (`eps = 1e-6`)
Context Window	`32,768` tokens (`8,192` default in Ollama `Modelfile`)
Vocabulary (`vocab_size`)	`49,152` custom Puma byte-level BPE (`<
Weights Provided	Sharded `FP16` `.safetensors` (`13.9 GB`) and `puma-1.1-q4_k_m.gguf` (`Q4_K_M`, `~4.2 GB`)
License	Apache-2.0

Structured `<think>...</think>` Reasoning Mode

Puma 1.1 performs explicit internal analysis inside <think>...</think> tags before writing its final response:

Adaptive Reasoning Depth: Allocates concise planning steps on straightforward prompts and deeper multi-step decomposition (constraints, complexity analysis, edge-case checks, arithmetic verification) on algorithmic coding, debugging, and STEM problems.
Native Client Support: Ollama (0.9+), Open WebUI, LM Studio, and Continue.dev automatically parse <think>...</think> into a collapsible thinking block while keeping the final response clean.
Multi-Turn Context Hygiene: The bundled Jinja chat_template strips completed <think>...</think> blocks from earlier conversation turns so long multi-turn coding sessions preserve context window capacity.

Capabilities & Intended Scope

What Puma 1.1 Excels At

Algorithmic & Multi-Language Software Engineering: Designs, implements, and debugs non-trivial code across Python, TypeScript/JavaScript, Rust, C++, SQL, and Bash—explicitly checking boundary conditions and algorithmic complexity inside <think>...</think> before emitting code.
Mathematical & Logical Problem Solving: Breaks down multi-step algebra, calculus, combinatorics, probability, and invariants step by step before stating the final result.
Architecture & Code Review: Diagnoses concurrency bugs, off-by-one errors, SQL query bottlenecks, and API schema mismatches with clear root-cause analysis.

Limitations

Hardware Footprint (~4.8 GB RAM in Q4_K_M): Designed for laptops, desktops, workstations, and 8 GB+ Apple Silicon / modern mobile devices. For ultra-constrained 4 GB phones or Raspberry Pi deployments, use Puma 1.0 (1.42B).
Offline Knowledge Cutoff: Does not browse the web natively; supply documentation excerpts or RAG context when querying library APIs released after training.
Reasoning Latency Trade-Off: Because <think>...</think> generates intermediate tokens before the final answer, time-to-final-answer is proportionally longer than non-thinking models.

Running Puma 1.1 Locally (Desktop & Laptop CPU / GPU)

Option 1: Ollama (Recommended — One Command)

ollama run Merkium/puma-1.1

Or run directly from Hugging Face via Ollama:

ollama run hf.co/Ueuee82683/Puma-1.1

Expected throughput (Q4_K_M, ~4.2 GB):

NVIDIA RTX 3060 / 4060 / T4 or newer (8 GB+ VRAM): 55 - 110 tokens/s
Apple Silicon (M1 / M2 / M3 / M4, 8 GB+ unified memory): 28 - 60 tokens/s
Modern x86_64 Desktop/Laptop CPU (AVX2, 8-16 threads, 8 GB+ RAM): 9 - 18 tokens/s

Option 2: `llama.cpp` CLI / OpenAI-Compatible Server

huggingface-cli download Ueuee82683/Puma-1.1 puma-1.1-q4_k_m.gguf --local-dir .

# Interactive CLI chat with reasoning support
./llama-cli -m puma-1.1-q4_k_m.gguf -cnv -c 8192 -t 8 --temp 0.5

# Local HTTP API server on port 8080
./llama-server -m puma-1.1-q4_k_m.gguf -c 8192 -t 8 --port 8080

Running on High-Memory Mobile / Tablet Devices (8 GB+ RAM)

Devices with 8 GB+ RAM (iPhone 15 Pro / 16 / 17 series, iPad M-series, Snapdragon 8 Gen 2/3/Elite Android flagships) can run puma-1.1-q4_k_m.gguf offline:

In PocketPal AI or Layla, add Ueuee82683/Puma-1.1 from Hugging Face, download puma-1.1-q4_k_m.gguf (~4.2 GB), set context length to 4096, and configure stop strings <|eos|> and <|user|>.
On 4 GB - 6 GB phones, use Ueuee82683/Puma-1.0 (1.44 GB).

Using Puma 1.1 in Python (Ollama API & Hugging Face `transformers`)

Streaming Chat via Ollama Local API (Separating Thought & Answer)

from openai import OpenAI

client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")

resp = client.chat.completions.create(
    model="Merkium/puma-1.1",
    messages=[
        {"role": "user", "content": "Write an O(n log n) Python function for longest increasing subsequence and prove why binary search on tails works."}
    ],
    temperature=0.5,
)
print(resp.choices[0].message.content)

Direct Hugging Face `transformers` Inference (`FP16` / `4-bit`)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Ueuee82683/Puma-1.1"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Find all integer solutions to x^2 - y^2 = 45."}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False, return_token_type_ids=False).to(model.device)

stop_ids = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("<|user|>"),
    tokenizer.convert_tokens_to_ids("<|system|>"),
]

with torch.no_grad():
    out = model.generate(
        **inputs,
        max_new_tokens=1024,
        temperature=0.5,
        top_p=0.92,
        do_sample=True,
        repetition_penalty=1.06,
        eos_token_id=stop_ids,
        pad_token_id=tokenizer.pad_token_id,
    )

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

Prompt Format

<|bos|><|system|>
You are Puma 1.1, an open-weight 6.96B-parameter reasoning and coding model created by Merkium and released under Apache 2.0. Think step-by-step inside <think>...</think> before giving your final response.<|eos|>
<|user|>
{user_message}<|eos|>
<|assistant|>
<think>
{step_by_step_reasoning}
</think>

{final_response}<|eos|>

License

Released under the Apache License 2.0.

Downloads last month: 58

Safetensors

Model size

7B params

Tensor type

F16