Skip to main content

​
Overview

ollama-plan reads your installed Ollama models and hardware profile, then computes safe values for OLLAMA_NUM_CTX, OLLAMA_NUM_PARALLEL, and OLLAMA_MAX_LOADED_MODELS. It prevents out-of-memory crashes by planning memory usage before you start Ollama. 
llm-checker ollama-plan

​
Example Output

OLLAMA CAPACITY PLAN
Hardware: Metal (metal)
Memory budget: 20GB usable (reserve 2GB)

Selected models:
  - qwen2.5-coder:14b (14B, ~9.1GB base)
  - llama3.2:3b (3B, ~2.0GB base)

Recommended envelope:
  Context: 8192 (requested 8192)
  Parallel: 2 (requested 2)
  Loaded models: 2 (requested 2)
  Estimated memory: 15.2GB / 20GB (76%)
  Risk: LOW (18/100)

Notes:
  - Running both models simultaneously fits within your memory budget

Recommended env vars:
  export OLLAMA_NUM_CTX=8192
  export OLLAMA_NUM_PARALLEL=2
  export OLLAMA_MAX_LOADED_MODELS=2

Fallback profile:
  OLLAMA_NUM_CTX=4096 OLLAMA_NUM_PARALLEL=1 OLLAMA_MAX_LOADED_MODELS=1

​
Flags

​
--models
string[]
Model tags or family names to include in the plan. Matches against installed Ollama models by exact name, prefix, family, or substring. If omitted, all local models are included.
​
--ctx
number
Target context window in tokens.Default: 8192
​
--concurrency
number
Target number of parallel requests to support.Default: 2
​
--objective
string
Optimization objective. Accepted values: latency, balanced, throughput.Default: balanced
​
--reserve-gb
number
Memory to reserve for the OS and background processes (GB).Default: 2
​
--json
flag
Output the full capacity plan as JSON.

​
Usage Examples

# Plan for all installed models
llm-checker ollama-plan

# Plan for specific models only
llm-checker ollama-plan --models qwen2.5-coder:14b llama3.2:3b

# Optimize for low latency with a larger context window
llm-checker ollama-plan --objective latency --ctx 16384

# Throughput-optimized plan with 4 concurrent requests
llm-checker ollama-plan --objective throughput --concurrency 4

# Reserve 4GB for other workloads
llm-checker ollama-plan --reserve-gb 4

# Machine-readable output
llm-checker ollama-plan --json

​
Plan Output Fields

FieldDescription
Memory budgetUsable RAM after subtracting the OS reserve
ContextRecommended OLLAMA_NUM_CTX value
ParallelRecommended OLLAMA_NUM_PARALLEL value
Loaded modelsRecommended OLLAMA_MAX_LOADED_MODELS value
Estimated memoryProjected memory usage and utilization percentage
RiskRisk level (LOW, MEDIUM, HIGH) and score (0–100)
Fallback profileConservative fallback values for constrained environments

​
Applying the Plan

Copy the export lines into your shell profile or set them before starting Ollama: 
export OLLAMA_NUM_CTX=8192
export OLLAMA_NUM_PARALLEL=2
export OLLAMA_MAX_LOADED_MODELS=2
ollama serve
Or apply them inline for a single session: 
OLLAMA_NUM_CTX=8192 OLLAMA_NUM_PARALLEL=2 OLLAMA_MAX_LOADED_MODELS=2 ollama serve
Ollama must be running and you must have at least one model installed before running ollama-plan. Install a model with ollama pull llama3.2:3b if needed.

Build docs developers (and LLMs) love

Get started for freeTalk to us