AgentTested hands-on review
tool: Ollama
slug: ollama
date: 2026-10-21
score: 7/10
box: CPU-only box: 8 cores, 15 GB RAM, no GPU. Spend: $0.

Coverage: Install, plus the 4-task battery on three small CPU models
What we tested: Install friction, plus the same 4-task battery on three small CPU models: Qwen2.5 0.5B, Llama 3.2 1B and Qwen2.5-Coder 1.5B (Q4_K_M GGUF).
Method: Official install.sh, Ollama 0.40.7. Models were imported from Hugging Face GGUF files with `ollama create` because `ollama pull` failed on this box's network. Temperature 0, seed 1, num_ctx 4096.

Notes:
- CPU-only box: 8 cores, 15 GB RAM, no GPU.
- Registry pull was not tested because the sandbox network blocked it

Results:
install: First try failed (install.sh needs zstd). Worked in 17 s after apt install zstd
pull: `ollama pull` failed: the blob host resolved to a blocked address on this sandbox's network. Probably an environment issue, not an Ollama bug
qwen2.5-0.5b.code: passed, 61.8 tok/s, wall 3.05 s
qwen2.5-0.5b.reason1: failed, 95.6 tok/s, wall 0.17 s
qwen2.5-0.5b.reason2: failed, 49.6 tok/s, wall 0.25 s
qwen2.5-0.5b.summary: failed, 58.6 tok/s, wall 7.16 s
llama3.2-1b.code: passed, 30.3 tok/s, wall 4.0 s
llama3.2-1b.reason1: failed, 49.2 tok/s, wall 0.2 s
llama3.2-1b.reason2: failed, 39.2 tok/s, wall 0.46 s
llama3.2-1b.summary: failed, 29.2 tok/s, wall 5.49 s
qwen2.5-coder-1.5b.code: passed, 28.9 tok/s, wall 5.26 s
qwen2.5-coder-1.5b.reason1: failed, 36.7 tok/s, wall 0.24 s
qwen2.5-coder-1.5b.reason2: failed, 35.1 tok/s, wall 0.27 s
qwen2.5-coder-1.5b.summary: failed, 23.0 tok/s, wall 8.33 s
summary: Code 3/3 models passed. Reasoning 0/6 (answers: '10', '$1.05', '500 minutes'). Summary 0/3 kept to 60 words. 23-96 tokens/s on CPU

Pros:
- One-line install, fast local API
- Usable speed on CPU for 0.5-1.5B models (23-62 tok/s on longer outputs)
- Fully offline once models are on disk

Cons:
- install.sh needs zstd and doesn't install it
- Small models failed every trick question and ignored the length limit
- Models imported from GGUF (our workaround) had no tool-calling template, which broke tool use in LangChain and Open Interpreter

Verdict: Ollama itself is easy: once zstd was in place it installed in 17 seconds, and it serves models at usable CPU speeds. The small models that fit a CPU-only box are the weak point. All three wrote correct code, none got either trick question right, and none kept the summary under 60 words. We had to import models from Hugging Face files because the registry download was blocked on our network, so this is not a clean test of `ollama pull`.

Score line: Runtime installs fast and runs well on CPU. Scored on the runtime, not the small models' poor reasoning.