Merkium/ puma-1.0

6 Downloads Updated 20 hours ago

1.42B open-weight Llama-architecture language and coding model with a custom 49k BPE tokenizer and 32K context window for fast CPU and mobile inference.

ollama run Merkium/puma-1.0

curl http://localhost:11434/api/chat \
  -d '{
    "model": "Merkium/puma-1.0",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='Merkium/puma-1.0',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'Merkium/puma-1.0',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Models

Name

1 model

Size / Usage

Context

Input

puma-1.0:latest

1.5GB · 32K context window · Text · 20 hours ago

puma-1.0:latest

1.5GB

32K

Text

Readme

Puma 1.0

Puma 1.0 is a 1.42-billion-parameter (1,418,911,232 parameters) open-weight language and code generation model created by Merkium and released under the Apache 2.0 license. Built on a 28-layer LlamaForCausalLM decoder (hidden_size = 1536, intermediate_size = 9216, 6:1 Grouped-Query Attention, 32K RoPE) paired with a custom 49,152-token byte-level BPE vocabulary, it runs locally on desktop CPUs, laptops, and smartphones in ~1.44 GB (Q8_0 GGUF).

Quickstart

ollama run Merkium/puma-1.0

Architecture

Specification	Value
Architecture	28-layer pre-norm Transformer decoder (`LlamaForCausalLM`, `arch: llama`)
Parameters	`1,418,911,232` (`1.42B`, tied embeddings)
Dimensions	`hidden_size = 1536`, `intermediate_size = 9216` (`6.0x` SwiGLU), `12` Q / `2` KV heads
Context Window	`32,768` tokens (`num_ctx = 8192` default)
Vocabulary	`49,152` custom byte-level BPE (`<
Quantization	`Q8_0` (`1.44 GB`)
License	Apache-2.0

Capabilities

Software Engineering & Scripting: Generates clean Python, SQL, Bash, regexes, data structures, and unit tests (90% Python AST validity, 70% strict unit-test pass rate on held-out algorithmic tasks).
Multi-Turn Local Assistant: Handles casual conversation, technical Q&A, text summarization, and structured JSON / Markdown output across multi-turn sessions.
CPU & Mobile Deployment: Runs at 55-95 tok/s on Apple Silicon, 20-40 tok/s on x86_64 AVX2 CPUs, and 12-28 tok/s offline on iOS/Android (PocketPal AI, Termux) via Ueuee82683/Puma-1.0.

Local API Usage (Python)

from openai import OpenAI

client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")
resp = client.chat.completions.create(
    model="Merkium/puma-1.0",
    messages=[{"role": "user", "content": "Write a Python function is_prime(n) with type hints."}],
    temperature=0.3,
)
print(resp.choices[0].message.content)

Links

Hugging Face Weights (FP16 + GGUF): https://huggingface.co/Ueuee82683/Puma-1.0

# Puma 1.0

Puma 1.0 is a 1.42-billion-parameter (`1,418,911,232` parameters) open-weight language and code generation model created by Merkium and released under the Apache 2.0 license. Built on a 28-layer `LlamaForCausalLM` decoder (`hidden_size = 1536`, `intermediate_size = 9216`, 6:1 Grouped-Query Attention, 32K RoPE) paired with a custom `49,152`-token byte-level BPE vocabulary, it runs locally on desktop CPUs, laptops, and smartphones in `~1.44 GB` (`Q8_0` GGUF).

## Quickstart

```bash
ollama run Merkium/puma-1.0
```

## Architecture

| Specification | Value |
| :--- | :--- |
| **Architecture** | 28-layer pre-norm Transformer decoder (`LlamaForCausalLM`, `arch: llama`) |
| **Parameters** | `1,418,911,232` (`1.42B`, tied embeddings) |
| **Dimensions** | `hidden_size = 1536`, `intermediate_size = 9216` (`6.0x` SwiGLU), `12` Q / `2` KV heads |
| **Context Window** | `32,768` tokens (`num_ctx = 8192` default) |
| **Vocabulary** | `49,152` custom byte-level BPE (`<|bos|>`, `<|eos|>`, `<|system|>`, `<|user|>`, `<|assistant|>`) |
| **Quantization** | `Q8_0` (`1.44 GB`) |
| **License** | Apache-2.0 |

## Capabilities

- **Software Engineering & Scripting:** Generates clean Python, SQL, Bash, regexes, data structures, and unit tests (`90%` Python AST validity, `70%` strict unit-test pass rate on held-out algorithmic tasks).
- **Multi-Turn Local Assistant:** Handles casual conversation, technical Q&A, text summarization, and structured `JSON` / Markdown output across multi-turn sessions.
- **CPU & Mobile Deployment:** Runs at `55-95 tok/s` on Apple Silicon, `20-40 tok/s` on x86_64 AVX2 CPUs, and `12-28 tok/s` offline on iOS/Android (`PocketPal AI`, `Termux`) via [`Ueuee82683/Puma-1.0`](https://huggingface.co/Ueuee82683/Puma-1.0).

## Local API Usage (Python)

```python
from openai import OpenAI

client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")
resp = client.chat.completions.create(
    model="Merkium/puma-1.0",
    messages=[{"role": "user", "content": "Write a Python function is_prime(n) with type hints."}],
    temperature=0.3,
)
print(resp.choices[0].message.content)
```

## Links

- **Hugging Face Weights (`FP16` + `GGUF`):** https://huggingface.co/Ueuee82683/Puma-1.0

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)