Writer/palmyra-mini-thinking-a-MLX-BF16

Palmyra Mini Thinking A - MLX BF16

Model Description

This is a bfloat16 precision version of the palmyra-mini-thinking-a model, optimized for Apple Silicon using the MLX framework. This model is based on the Qwen2 architecture and is specifically designed for reasoning tasks with explicit thinking capabilities through special <think> and </think> tokens.

Quick Start

Installation

pip install mlx-lm

Usage

from mlx_lm import load, generate

# Load the model
model, tokenizer = load("/Users/thomas/Documents/Model Weights/SPW2 Mini Launch/palmyra-mini-thinking-a/MLX")

# Generate text with thinking
prompt = "Solve this step by step: What is 15% of 240?"
response = generate(model, tokenizer, prompt=prompt, verbose=True, max_tokens=512)
print(response)

Technical Specifications

Model Architecture

• Model Type: qwen2 (Qwen2 Architecture)
• Architecture: Qwen2ForCausalLM
• Parameters: ~1.7 billion parameters
• Precision: bfloat16
• Specialization: Reasoning and thinking tasks

Core Parameters

Parameter	Value
Hidden Size	1,536
Intermediate Size	8,960
Number of Layers	28
Attention Heads	12
Key-Value Heads	2
Head Dimension	128
Vocabulary Size	151,665

Attention Mechanism

• Attention Type: Full attention across all 28 layers
• Max Position Embeddings: 131,072 tokens
• Attention Dropout: 0.0
• Sliding Window: Not used
• Max Window Layers: 21

RoPE (Rotary Position Embedding) Configuration

• RoPE Theta: 10,000
• RoPE Scaling: None

Thinking Capabilities

• Thinking Tokens: <think> (151648) and </think> (151649)
• Reasoning Mode: Explicit step-by-step reasoning
• Chat Template: Automatically adds <think> tag for generation prompts

File Structure

palmyra-mini-thinking-a/MLX/
├── config.json                    # Model configuration
├── model.safetensors              # Model weights (3.3GB)
├── model.safetensors.index.json   # Model sharding index
├── tokenizer.json                 # Tokenizer configuration
├── tokenizer_config.json          # Tokenizer settings
├── special_tokens_map.json        # Special tokens mapping
├── chat_template.jinja            # Chat template with thinking
└── README.md                      # Model documentation

Performance Characteristics

Hardware Requirements

• Platform: Apple Silicon (M1, M2, M3, M4 series)
• Memory: ~3.3GB for model weights
• Recommended RAM: 12GB+ for optimal performance
• Precision: Full bfloat16 precision

Layer Configuration

All 28 layers use full attention mechanism as specified in the layer_types configuration, providing consistent attention patterns across the entire model depth.

Training Details

Tokenizer

• Type: LlamaTokenizerFast with 151,665 vocabulary size
• Special Tokens:
  • BOS Token ID: 151646 ( )
  • EOS Token ID: 151643 ( )
  • Pad Token ID: 151643 ( )
  • Think Start: 151648 (<think>)
  • Think End: 151649 (</think>)

Model Configuration

• Hidden Activation: SiLU (Swish)
• Normalization: RMSNorm (ε = 1e-06)
• Initializer Range: 0.02
• Attention Dropout: 0.0
• Word Embeddings: Not tied
• Use Cache: False (optimized for thinking tasks)

Chat Template

The model uses a specialized chat template that automatically initiates thinking mode:

• User messages:
• Assistant messages: <｜Assistant｜><think>\n (automatically adds thinking prompt)
• Tool calling support with <tool_call> and </tool_call> tokens
• Vision and multimodal tokens included

Usage Examples

Reasoning Task

prompt = """
A train travels 120 miles in 2 hours. If it maintains the same speed, how far will it travel in 5 hours?
<｜Assistant｜><think>
"""

response = generate(model, tokenizer, prompt=prompt, max_tokens=300)

Problem Solving

prompt = """
Explain why the sky appears blue during the day.
<｜Assistant｜><think>
"""

response = generate(model, tokenizer, prompt=prompt, max_tokens=400)

Known Limitations

1. Platform Dependency: Optimized specifically for Apple Silicon; may not run on other platforms
2. Memory Requirements: Requires significant memory due to full precision weights
3. Thinking Overhead: Explicit thinking may increase response length and generation time
4. Cache Disabled: Model has use_cache: false which may impact inference speed

Compatibility

• MLX-LM: Requires recent version with Qwen2 support
• Apple Silicon: M1, M2, M3, M4 series processors
• macOS: Compatible with recent macOS versions supporting MLX
• Transformers: Version 4.52.4+

License

Apache 2.0

Original model card below:

---

Palmyra-mini-thinking-a

Model Description

• Language(s) (NLP): English
• License: Apache-2.0
• Finetuned from model: Qwen/Qwen2.5-1.5B
• Context window: 131,072 tokens
• Parameters: 1.7 billion

Model Details

The palmyra-mini-thinking-a model demonstrates exceptional performance in advanced mathematical reasoning and competitive programming. Its capabilities are highlighted by an outstanding score of 0.886 on the 'MATH500' benchmark, showcasing a robust ability to solve complex mathematical problems. The strength of the model in quantitative challenges is further confirmed by its score of 0.8287 on 'gsm8k (strict-match)', which demonstrates proficiency in multi-step arithmetic reasoning. Additionally, the model proves its aptitude for high-level problem-solving with a score of 0.8 on 'AMC23'. The model also shows strong potential in the coding domain, achieving a score of 0.5631 on 'Codeforces (pass_rate)' and 0.5481 on 'Olympiadbench (extractive_match)', indicating competence in generating correct solutions for programming challenges.

Benchmark Performance

This section provides a detailed breakdown of the palmyra-mini-thinking-a model's performance across a standardized set of industry benchmarks. The data is presented in its original order from the source evaluation.

Benchmark	Score
gsm8k (strict-match)	0.8287
minerva_math(exact_match)	0.3842
mmlu_pro(exact_match)	0.2748
hendrycks_math	0.0054
ifeval (inst_level_loose_acc)	0.3657
mathqa (acc)	0.4171
humaneval (pass@1)	0.2378
BBH (get-answer)(exact_match)	0.462
mbpp	0.304
leadboard_musr (acc_norm)	0.3413
gpqa lighteval gpqa diamond_pass@1:8_samples	0.3826
AIME24(pass@1)(avg-of-1)	0.4333
AIME25(pass@1)(avg-of-1)	0.3667
Livecodebench-codegen (livecodebench/code_generation_lite v4_v5)	0.1784
AMC23	0.8
MATH500	0.886
Minerva	0.3493
Olympiadbench (extractive_match)	0.5481
Codecontests (pass_rate)	0.1778
Codeforces (pass_rate)	0.5631
Taco (pass_rate)	0.3083
APPS (all_levels)	0.0447
HMMT23 (extractive_match)	0.1
Average	0.380839

Use with transformers

You can run conversational inference using the Transformers Auto classes with the generate() function. Here's an example:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "Writer/palmyra-mini-thinking-a"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
    attn_implementation="flash_attention_2",
)

messages = [
      {
        "role": "user",
        "content": "You have a 3-liter jug and a 5-liter jug. How can you measure exactly 4 liters of water?"
      }
    ],

input_ids = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
)

gen_conf = {
    "max_new_tokens": 256,
    "eos_token_id": tokenizer.eos_token_id,
    "temperature": 0.3,
    "top_p": 0.9,
}

with torch.inference_mode():
    output_id = model.generate(input_ids, **gen_conf)

output_text = tokenizer.decode(output_id[0][input_ids.shape[1] :])

print(output_text)

Running with vLLM

vllm serve Writer/palmyra-mini-thinking-a

curl -X POST http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Writer/palmyra-mini-thinking-a",
    "messages": [
      {
        "role": "user",
        "content": "You have a 3-liter jug and a 5-liter jug. How can you measure exactly 4 liters of water?"
      }
    ],
    "max_tokens": 8000,
    "temperature": 0.2
  }'

Ethical Considerations

As with any language model, there is a potential for generating biased or inaccurate information. Users should be aware of these limitations and use the model responsibly.

Citation and Related Information

To cite this model:

@misc{Palmyra-mini-thinking-a,
  author = {Writer Engineering team},
  title = {{Palmyra-mini: A powerful LLM designed for math and coding}},
  howpublished = {\url{https://dev.writer.com}},
  year = 2025,
  month = Sep 
}

Contact [email protected]