Spectre (Mistral Large 3) : This UNLIMITED FREE AI Model on KiloCode is ACTUALLY INSANE!

Key Concepts

• Spectre Model: A new, stealthy language model available on Kilo Code.
• Kilo Code: A platform where the Spectre model can be accessed and used.
• Context Window: The amount of text a model can consider at once (256k tokens for Spectre).
• Tokens per Step: The maximum number of tokens processed in a single operation (8,000 for Spectre).
• Multimodal Support: The ability of a model to process different types of data, such as images (currently lacking in Spectre).
• Tool Calling: The capability of a model to invoke external tools or functions.
• Reasoning Model: A model that explicitly shows its thought process or reasoning steps.
• Non-Reasoning Model: A model that performs tasks without displaying explicit reasoning traces.
• Mistral Large: A potential identity of the Spectre model, known for its large parameter count.
• Codestral: Another potential identity of the Spectre model, likely focused on code generation.
• Transformers Repo: A repository for natural language processing models, which has added support for Mistral models.
• Parameter Count: A measure of a model's size and complexity (675 billion for Mistral Large).
• Open Sourcing: Making model weights and code publicly available.
• Reinforcement Learning (RL): A type of machine learning where an agent learns by trial and error.
• Apache Licenses: A type of permissive open-source license.
• Agentic Benchmarks: Performance tests specifically designed for AI agents.
• Chain of Thought (CoT): A prompting technique that encourages models to show their reasoning steps.
• Latency: The delay between input and output in a model.
• Multimodal Model: A model capable of processing multiple data types (e.g., text and images).

Spectre Model: Introduction and Access

The video introduces a new, stealthy language model named Spectre, currently accessible via Kilo Code. Key technical specifications include a substantial 256k token context window and a maximum of 8,000 tokens per step. While it currently lacks image or multimodal support, this is anticipated in the near future.

Accessing Spectre:

1. Navigate to the VS Code marketplace.
2. Search for and install the Spectre model.
3. Once installed, it will appear in the VS Code sidebar.
4. To enable Spectre, go to Settings, create a new profile, and name it.
5. Select the Kilo gateway provider.
6. Choose the Spectre model from the available options.

Performance and Capabilities

Spectre is described as a non-reasoning model, meaning it does not display explicit thinking traces or reasoning steps, nor does it offer options to adjust reasoning effort. Currently, it operates with unlimited usage and no rate limits, which is highlighted as a significant advantage.

Observed Performance:

• Seamless Tool Calling: The model excels at accurately invoking the correct tools.
• Conciseness: It is not verbose, typically only outputting the tool call itself without additional commentary.
• Problem Solving:
  • Successfully solved the first part of an Advent of Code 2025 day one question.
  • Initially struggled with the second part but autonomously corrected itself, demonstrating impressive self-correction.
• Comparison with Other Models:
  • GLM4.6 failed to solve the second part of the Advent of Code problem.
  • Miniax also failed on the second part.
  • Kimmy successfully solved the second part.

Potential Identity and Technical Speculation

The presenter has strong suspicions that Spectre is either Mistral Large or Codestral. This is supported by the recent addition of Mistral models to the Transformers repo.

Speculative Details:

• Mistral Large: Believed to be a 675 billion parameter model, making it exceptionally large.
• Size Comparison: The model is estimated to be larger than GLM but smaller than Kimmy, based on observed performance.
• Mistral's Comeback: This release could signify a strong resurgence for Mistral, especially in the Western open-source model landscape, which is currently perceived as having fewer strong contenders.
• Open Sourcing Shift: Mistral had previously moved away from open-sourcing, losing momentum. This new model appears to be open, potentially adopting strategies from models like Deepseek.
• RL and Tool Calling: Unlike many previous Mistral models that did not use reinforcement learning and were less adept at tool calling, this new model might represent a generational leap.
• Licensing: Mistral's past licensing issues are mentioned, but there's an indication they might return to Apache licenses. This could benefit projects like Cursor building their Composer model.
• Confirmation: The presenter is "pretty confident" it's a Mistral model due to observed behaviors and confirmation from other sources.

Performance on Agentic Benchmarks

While not yet tested on general benchmarks due to a preference for API access, Spectre has been evaluated on Agentic benchmarks.

Agentic Benchmark Findings:

• Comparable to GLM: Its performance is on par with GLM on these benchmarks.
• Tool Call Accuracy: Exhibits almost no tool call errors.
• Diff Edit Failures: No failures observed in diff editing tasks.
• Limitations:
  • Does not perform "God do godo" well (specific task unclear from transcript).
  • Not particularly strong at front-end development.
• Strengths:
  • Awesome at back-end development.
  • Tends to write simple, un-overcomplicated code.

Spectre vs. Other Models: Reasoning and Latency

A key distinction highlighted is Spectre's non-reasoning nature.

Reasoning Models and Trade-offs:

• Spectre: A raw model performance without explicit reasoning steps.
• Miniax, GLM, Kimmy, DeepSeek: These models now incorporate reasoning, which improves performance but increases speed and latency.
• Spectre's Advantage: As a non-reasoning model, it likely offers faster inference times.
• Future Possibility: The presenter expresses a desire for a reasoning variant of Spectre.

Future Expectations and Multimodality

Spectre is expected to evolve into a multimodal model, which would set it apart from many current open models.

Multimodality Landscape:

• Spectre: Expected to become multimodal.
• Miniax, GLM, Kimmy: Currently not multimodal, though GLM has a vision model coming soon.

Upcoming Releases:

• A Ministral model is also expected soon.
• Both Spectre and Ministral are anticipated to be released around December 25th (Christmas).

Conclusion and Call to Action

The presenter expresses enthusiasm for Spectre, hoping it signifies Mistral's return to strong open releases. The model is considered "pretty cool" and a potential "godsend" for the Western open-source community. The presenter encourages viewers to share their thoughts, subscribe, and consider supporting the channel.