OpenAI OSS-20B (Fully Tested) + Free APIs: This Model is very weird! Can it be your next AI Coder?

Key Concepts

• GPTOSS 20B/21B Model: A large language model with 21 billion parameters and a Mixture of Experts (MoE) architecture.
• GLM 4.5 Air: A superior alternative to GPTOSS 120B, according to the speaker.
• Mixture of Experts (MoE): An architecture where multiple "expert" models exist, and only a subset is activated for each input token.
• Tool Calls: The ability of a language model to use external tools or APIs to perform tasks.
• Reasoning Effort: The amount of computational resources a model dedicates to reasoning about a problem.
• Quen 3 Coder Flash/Devstral: Smaller models specifically trained for coding and tool calling.
• Rue/Kilo: Platforms or tools for interacting with language models.
• AMA (Likely referring to Ollama): A tool for running language models locally.
• Dart: A project management tool with AI features.

GPTOSS 20B/21B Model Analysis

• Model Architecture: The GPTOSS 20B model is actually a 21B model. It utilizes a Mixture of Experts (MoE) architecture with 3.6 billion parameters. It has 32 experts, but only four are activated per token.
• Benchmark Claims vs. Reality: The model supposedly scores similarly to GPT-3.5 and GPT-4 Mini in benchmarks. However, the speaker's testing reveals significant shortcomings, particularly in coding.
• Coding Performance: The model performs poorly in coding tasks. Even with high reasoning settings, the generated code is often non-functional.
• Tool Calling: The model is good at tool calls, but this functionality can become unreliable over time.
• Censorship and Training Data: The model is heavily censored, leading to instances where it refuses to answer harmless questions, even if it simply lacks the knowledge. Example: Refusal to explain "how to walk a fish" due to perceived offensiveness.
• Compromises in Smaller Models: Smaller models require compromises. General-purpose small models like Quen 3 small are not as good at most tasks. Specialized models like Quen 3 Coder Flash and Devstral excel in specific areas like coding and tool calling.

Comparison with Other Models

• GLM 4.5 Air Superiority: The speaker reiterates that GLM 4.5 Air is a better option than GPTOSS 120B. The 20B model doesn't come close to GLM 4.5 Air, even without reasoning.
• Quen 3 Coder Flash/Devstral for Coding: These models are preferred for coding tasks due to their specialized training. They work well with tools like Rue and Kilo. General-purpose models like Mistral small or Quen 3 small are not as effective for coding.

Practical Usage and Tools

• Rue and Kilo Integration: The speaker demonstrates how to use the GPTOSS 20B model with Rue and Kilo.
  • Local Model Setup: Download and install the model from AMA (Ollama).
  • Rue Configuration: Create a new profile, select the appropriate provider (e.g., Open Router, Ollama), and choose the model.
  • Kilo Configuration: Select the provider and model.
  • Kilo Code: Offers $15 of free credits for testing.
• Limitations of Rue and Kilo: Both platforms lack the option to adjust reasoning effort.
• Reasoning Limitations of GPTOSS 20B: Even with high reasoning settings, the model's reasoning is limited to 2K-4K tokens.

Sponsor: Dart

• Overview: Dart is a project management tool that integrates traditional features with AI capabilities.
• AI Features: Dart's AI can brainstorm project ideas, generate task lists, and complete assignments.
• Custom Agents: Users can create custom AI agents that trigger from built-in integrations, N8N workflows, or custom webhooks. Examples: coding agent for GitHub pull requests, marketing agent for campaigns, mailing agent for outreach.
• Integration: Dart integrates with existing workflows through its MCP server, connecting to Claude, Chat GPT, and other AI tools.
• Pricing: Most features are free, with premium options starting at $8 per month.

Future Expectations

• GPT-5 Anticipation: The speaker expresses anticipation for the launch of GPT-5 but remains skeptical about its potential impact. The hope is that it will at least surpass Sonnet.

Conclusion

The GPTOSS 20B model, while interesting, is ultimately disappointing due to its poor coding performance, excessive censorship, and limited reasoning capabilities. It is not a viable alternative to GLM 4.5 Air or specialized coding models like Quen 3 Coder Flash and Devstral. The speaker highlights the importance of specialized models for specific tasks and expresses hope for improvements in future releases like GPT-5. The video also showcases Dart as a project management tool that leverages AI to enhance productivity.