Why does the free model feel dumber?
Model size, free vs paid tiers · Beginner · 5 min · llm, training · after Training: loss and gradient descent
Free tiers often run a smaller or older model, with tighter limits. In one model family, the smaller sizes score far lower on knowledge and hard reasoning, while simple jobs like following a format barely change.
1 Guess first
2 Look
Same family, seven sizes
1 / 4Knowledge (MMLU-Redux)
0.6B44.61.7B64.44B77.38B79.514B82.032B85.7235B89.2model size in parameters · score out of 100
Qwen3 family, thinking off · Qwen3 Technical Report (2025), Tables 12–20 · 235B = 235B-A22B
Knowledge: the smallest model scores 44.6, the largest 89.2. Size holds the facts.
In one family, size barely matters for following a format but matters a lot for knowledge and hard reasoning; and the vendors keep their biggest models off the free tier.
A model’s parameters are the numbers it learned in training. More parameters hold more facts and longer chains of reasoning, but cost more to run for every token. That is why a free tier, which someone pays to serve, tends to use a smaller model or an older one, and caps how much you can send.
4 You are the model
You are the router that picks a model for each job. Choose the smallest size that reaches the target score: bigger costs more, too small fails the job.
job 1 of 4 · pick the smallest model that scores at least 80 on Following instructions (IFEval)
Reformat notes into a fixed template
Qwen3 family, thinking off · Qwen3 Technical Report (2025), Tables 12–20 · 235B = 235B-A22B
You are the router: bigger costs more, too small fails the job.5 Check
6 Say it back
7 Mission · 5 min
In Google AI Studio (free), ask Gemini 3.5 Flash-Lite and then Gemini 3.8 Flash the same hard question, for example: “A train leaves at 9:40 and arrives at 13:05 after two 12-minute stops. How long was it moving?” Compare the answers, then try an easy reformatting job on both.
Three quick ones
Drawn at random from this lesson's pool; different next time.
Done?
Go deeper
Sources
- https://arxiv.org/abs/2505.09388 (Qwen3 Technical Report, 2025-05-14; checked 2026-10-21): non-thinking Tables 12, 14, 16, 18, 20 — MMLU-Redux 44.6/64.4/77.3/79.5/82.0/85.7/89.2, AIME’24 3.4/13.4/25.0/29.1/31.7/31.0/40.1, GPQA-Diamond 22.9/28.6/41.7/39.3/54.8/54.6/62.9, IFEval strict 54.5/68.2/81.2/83.0/84.8/83.2/83.2 for 0.6B/1.7B/4B/8B/14B/32B/235B-A22B
- https://ai.google.dev/gemini-api/docs/pricing (updated 2026-10-19, checked 2026-10-21): Gemini 4 Argon and 3.1 Pro Preview “Not available” on the free tier; 3.8 Flash, 3.5 Flash-Lite, 2.5 Pro “Free of charge”; free tier “Used to improve our products”: Yes
- https://chatgpt.com/pricing (checked 2026-10-21 in Chrome): Free “Limited memory and context”; Plus “Expanded memory and context”, “Advanced reasoning models”
- https://arxiv.org/abs/2001.08361
Published Oct 21, 2026 · updated Oct 21, 2026