Role of AI

Choose the AI arrangement that fits your course.

Privacy, cost, model capability, hardware, and institutional policy do not have one universal answer. LLMGrader lets the instructor choose how AI is provided while retaining responsibility for every grade, feedback decision, and instructional conclusion.

Return to LLMGrader

Instructor-led, AI-assisted

The AI proposes. The instructor decides.

Why AI is used

AI can draft grades and feedback, compare many responses, and prepare a first analysis of class-wide patterns that would otherwise require substantial instructor time.

Human review remains essential

The instructor reviews and can edit every suggested grade, feedback response, common-theme analysis, and interpretation of progress.

You choose what is sent

Student identifiers are not sent to the model. When a cloud model is selected, answer text and the grading materials needed for the task are sent to that provider.

A practical private option

For many LLMGrader tasks, a local model is all you need.

A gaming PC or a recent Apple-silicon Mac can run useful models locally through Ollama. No student response data needs to be sent to a model provider, and there is normally no per-request cloud charge.

Local models are especially practical when the questions, rubrics, and student answers fit comfortably within the model’s context window and the instructor provides clear grading guidance. A smaller or medium model can often handle routine grading and analysis effectively; larger or more difficult assignments may benefit from a more capable local model or a frontier cloud model.

The desktop version connects directly to Ollama running on the same computer. An institution can instead host a compatible model service behind its firewall and make it available through LLMGrader’s Custom Endpoint option.

Local means local. With Ollama running on the instructor’s desktop, prompts, rubrics, solutions, and student answer content remain on that computer rather than being transmitted to a cloud model provider.

Question and rubric development

Refine questions and solutions, develop rubrics, and check whether grading criteria are clear.

Grading and individualized feedback

Evaluate short and medium-length responses using the instructor’s solution, rubric, base prompt, and grading guidance.

Common Theme Analysis

Identify recurring strengths, misconceptions, incomplete explanations, and areas for instructional follow-up across a class.

Course Progress analysis

Compare completed assessments to examine conceptual development by a student and across the class.

Realistic starting points

What might run on your computer?

These examples are starting points rather than guarantees. Requirements depend on model architecture, quantization, context length, operating-system use, and whether memory is shared or dedicated.

Mac with 16 GB unified memory

Smaller models in roughly the 3B–4B range are realistic starting points for many routine tasks.

Mac with 24–32 GB unified memory

Quantized models in roughly the 7B–14B range may be practical, depending on context length and other applications.

Mac with 64 GB or more

Larger quantized models in the 20B–30B range become feasible, although speed varies by model and context.

Gaming PC with 8 GB of VRAM

Quantized models in roughly the 3B–8B range are generally the most practical choices.

Gaming PC with 16–24 GB of VRAM

Quantized models in roughly the 12B–27B range may fit, depending on model architecture and context length.

Gaming PC with 32 GB of VRAM

A capable local model in roughly the 27B–30B range can be practical for grading and analysis work.

Model file size is only part of the requirement. The computer also needs memory for the operating system, LLMGrader, the active context, and temporary working space. Quantized models reduce memory requirements, sometimes with a tradeoff in quality or speed.

When local is not the best fit

Local, institutional, and cloud models are complementary choices.

Local desktop model

Maximum local privacy and predictable cost. Best when the available hardware can run a model capable of the intended task.

Private and practical

Institutional endpoint

An organization can operate an approved compatible model service behind its firewall, with its own access, retention, and auditing policies.

Organizational control

Cloud provider

Cloud services provide broad access to free, inexpensive, and frontier models without requiring local model hardware.

Maximum model choice

LLMGrader supports OpenRouter, OpenAI, Anthropic, Custom Endpoints, and—in the desktop version—Local Ollama and Ollama Cloud. The application’s Settings screen provides setup guidance and official API-key links for each option.

Responsible use

Evaluate the model in the context where it will be used.

Test a representative set of answers before grading a full class. Consider accuracy, consistency, preservation of nuance, response time, context capacity, and whether the feedback is appropriate for the course and assignment.

Follow institutional policy. Before using any AI service with student work, review your institution’s requirements for artificial intelligence, student data, privacy, approved vendors, and disclosure. LLMGrader provides choices; it does not determine whether a particular model or use is permitted.