Goliath 120B Model Documentation

Overview

Model Name: Goliath 120B Creator: alpindale Inference Model ID: alpindale/goliath-120b Provider: Mancer (via LangMart) Release Date: November 10, 2023

Description

Goliath 120B is a merged model that combines "two fine-tuned Llama 70B models into one 120B model" by merging Xwin and Euryale variants. The model was created using the mergekit framework by @chargoddard, with merge ratio optimization by @Undi95.

This model represents an advanced approach to model merging, leveraging the strengths of both fine-tuned variants to create a more capable 120B parameter model.

Technical Specifications

Model Architecture

• Model Group: Llama2
• Base Model: Llama 70B (merged variant)
• Total Parameters: 120 Billion
• Context Window: 6,144 tokens
• Instruction Format: Airoboros

Input/Output Capabilities

• Input Modalities: Text only
• Output Modalities: Text only
• Max Completion Tokens: 1,024 per request
• Default Stop Sequences: USER:, </s>

Pricing

Metric	Cost
Context Window	6,144 tokens
Input Token Cost	$6/Million tokens
Output Token Cost	$8/Million tokens
Max Completion Tokens	1,024 per request

Provider: Mancer 2

Cost Calculation Example

• Request: 100 input tokens + 500 output tokens
• Input cost: 100 × ($6/1M) = $0.0006
• Output cost: 500 × ($8/1M) = $0.004
• Total: $0.0046

Capabilities

Feature	Supported
Tool Use	No
Reasoning	No
Vision	No
Function Calling	No

Supported Parameters

The model supports a comprehensive set of parameters for fine-grained control:

Parameter Category	Supported Options
Response Format	JSON mode, text
Token Limits	max_tokens, min_tokens
Sampling	temperature, top_p, top_k, top_a, min_p
Penalties	frequency_penalty, presence_penalty, repetition_penalty
Control	stop sequences, logit_bias
Advanced	seed (for reproducibility), logprobs

Parameter Details

• temperature: Controls randomness (0.0 = deterministic, 2.0 = high randomness)
• top_p: Nucleus sampling parameter (0.0-1.0)
• top_k: Restricts sampling to top K tokens
• top_a: Threshold for token amplitude
• min_p: Minimum probability threshold
• frequency_penalty: Reduces token repetition based on frequency
• presence_penalty: Reduces token repetition based on presence
• repetition_penalty: Alternative token repetition control
• logit_bias: Adjusts logit values for specific tokens
• seed: Ensures reproducible outputs
• logprobs: Returns log probabilities of generated tokens

Use Cases

Goliath 120B is suitable for:

• Long-form text generation
• Advanced reasoning tasks
• Complex dialogue and instruction following
• Content creation and analysis
• Fine-tuned inference applications
• Creative writing and storytelling
• Code generation and technical explanations
• Document summarization and analysis

Limitations

• Context Window: 6,144 tokens is limited compared to modern models (32K-200K+)
• Modalities: Text-only (no vision, audio, or multimodal capabilities)
• Advanced Features: No function calling, tool use, or extended reasoning support
• Training Data: Based on Llama 2, may have knowledge cutoff limitations

Related Models

• Xwin-LM: One of the base models used in the merge (Llama 70B variant)
• Euryale: The second base model used in the merge (Llama 70B variant)
• Llama 2 70B: Base architecture for both merged models
• Goliath (Other Variants): Alternative Goliath model configurations

Development Credits

• Merge Framework: @chargoddard (mergekit)
• Merge Optimization: @Undi95 (ratio optimization)
• Creator: alpindale

Provider Information

Mancer 2 Details

• Provider Name: Mancer 2
• Hosting Endpoint: neuro.mancer.tech/oai/v1
• Data Retention Policy: No data retention for training purposes
• Privacy: Terms of service available at mancer.tech
• Hosting: Dedicated hosting for optimal performance

Integration & Access

LangMart Integration The model is available through LangMart for:

• Chat completions interface
• Model comparison tools
• Batch processing

Model ID: alpindale/goliath-120b

Model Weights: Publicly available on Hugging Face for community use and local deployment

Usage

API Request Example

curl -X POST https://api.langmart.ai/v1/chat/completions \
  -H "Authorization: Bearer $LANGMART_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "alpindale/goliath-120b",
    "messages": [
      {
        "role": "user",
        "content": "Hello, how are you?"
      }
    ]
  }'

Python Example

from openai import OpenAI

client = OpenAI(
    base_url="https://api.langmart.ai/v1",
    api_key="YOUR_LANGMART_API_KEY"
)

response = client.chat.completions.create(
    model="alpindale/goliath-120b",
    messages=[
        {"role": "user", "content": "Hello, how are you?"}
    ]
)

print(response.choices[0].message.content)

Performance Characteristics

• Size: 120 Billion parameters
• Context: 6,144 token context window (suitable for standard conversations and documents)
• Inference Speed: Optimized by Mancer for fast inference
• Quality: Result of advanced merging of fine-tuned Llama models
• Instruction Following: Airoboros format specialized instruction handling

Model Comparison

Compared to other models available on LangMart:

• vs Claude 3.5 Sonnet: Less capable but significantly cheaper, text-only
• vs GPT-4: Stronger for creative tasks, weaker for complex reasoning
• vs Llama 2 70B: Combined strengths through model merging
• vs Llama 3: Older architecture but still effective

Integration Guide

Using with LangMart

const response = await client.post('/v1/chat/completions', {
  model: 'alpindale/goliath-120b',
  messages: [
    { role: 'user', content: 'Your prompt here' }
  ],
  temperature: 0.7,
  max_tokens: 1000
});

Environment Variables

LANGMART_API_KEY=your_api_key_here
LANGMART_MODEL_ID=alpindale/goliath-120b

References

• Creator: alpindale
• Mergekit Framework: Created by @chargoddard
• Merge Optimization: @Undi95
• Hosting Provider: Mancer 2
• Platform: LangMart.ai
• Repository: Hugging Face (model weights available)

---

Last Updated: December 23, 2025 Source: LangMart Model Registry Model Card: Available on LangMart and Hugging Face