Best Free AI APIs for Spanish, Portuguese and Indonesian in 2026: Published Scores and Free Tiers

For Indonesian there is a current, public ranking of today's models, and the free API that scores best is Gemini 3.8 Flash on the Gemini API free tier: 80.86 on AI Singapore's SEA-HELM leaderboard, level with paid models such as Claude Opus 4.8 (81.23). Gemini 3.6 Flash, Qwen 3.5 397B and Gemma 4 31B are within about one point of it, and Gemma 4 31B is the easiest to use for free. For Spanish and Portuguese we found no equivalent leaderboard of 2026 models that we could read, so the evidence is thinner: OpenAI's own MMMLU table for GPT OSS and an older MMLU-ProX table. We say below what each can and cannot show.

Every score comes from a published leaderboard, model card or paper, linked with its date; we read them on 2026-11-01 and did not run any test ourselves. Rows that link an offer page are free tiers in our verified catalog. Why models lose points outside English at all is explained in why AI is worse in low-resource languages.

Indonesian: current scores and where to use each model free #

SEA-HELM scores each model on Indonesian tasks (knowledge, instruction following, multi-turn chat, generation, reasoning, understanding, culture and safety) and averages them into one score out of 100. The table lists the models with a free API in our catalog:

Source: SEA-HELM Indonesian leaderboard (results dated 2026-09-18; 87 models: 71 open-weight of all sizes and 16 closed), read 2026-11-01; our offer pages for the free limits (catalog checks dated 2026-10-12 to 2026-10-31). The leaderboard's default view shows only open models up to 200B, so the link above sets Model size to All and turns on closed models. The 95% confidence interval is about ±1.2 points for the top models, so models within two points of each other are effectively tied.

For scale, the paid models at the top of the same table are GPT 6 Astra (81.40), Claude Opus 4.8 (81.23) and Gemini 3.1 Pro Preview (81.21, not on the free tier). The Gemini API free tier is listed as open in Indonesia, Spain, Portugal, Brazil, Mexico, Argentina, Colombia and Chile, among other countries, and it needs no card.

Spanish and Portuguese: what we could verify #

SEA-HELM covers Southeast Asian languages, so it has no Spanish or Portuguese. Two published benchmarks cover Spanish, Portuguese and Indonesian side by side, which lets you compare the three languages for the same model.

MMMLU for GPT OSS. OpenAI's GPT OSS model card reports MMMLU, a professionally human-translated version of MMLU in 14 languages, for the two open models that Groq, Cloudflare and OVHcloud serve on free tiers. Each model is scored at three reasoning-effort settings:

Source: gpt-oss-120b and gpt-oss-20b model card, table 2 (submitted 2025-08-08), read 2026-11-01. The card does not say which setting a hosted free API uses, so treat the low column as the floor and the high column as the ceiling.

The same model scores within two points in Spanish and Portuguese at every setting, and up to four points lower in Indonesian than in Spanish, with the gap widest for the 20B model at low effort (75.0 against 71.2). Going from low to high effort adds five to six points to the 120B model and six to eight to the 20B.

MMLU-ProX. MMLU-ProX is a harder multilingual version of MMLU-Pro, covering 29 languages including all three. Its project page lists eight models, each scored 5-shot with chain of thought:

Source: MMLU-ProX project page, read 2026-11-01; the page gives no date for the table, and the paper (arXiv 2503.10497) was first submitted 2025-03-13. Its leaderboard is marked "coming soon". These are older models, and none of these exact versions has a free API in our catalog, so use the table for the pattern, not for picking a model: scores in the three languages are close, and which one is lowest changes from model to model.

What this does not show: how Gemini, Gemma or Qwen 3.5 score in Spanish or Portuguese. Do not assume the Indonesian ranking carries over. Test it on your own prompts.

Gemini: the strongest free option in Indonesian #

SEA-HELM puts Gemini 3.8 Flash (80.86) and Gemini 3.6 Flash (80.85) inside the confidence interval of the best paid models, and Gemini 3.1 Flash-Lite (79.39) is not far behind. All three are on the Gemini API free tier, which also serves Gemma 4. The catch is that Google shows the free per-model limits only inside AI Studio, not on a public page, so check your quota there before you build on it. Steps: how to get a free Gemini API key.

SEA-LION: a free key for Southeast Asian languages #

AI Singapore's SEA-LION models are trained further on Southeast Asian languages, including Indonesian. The SEA-LION API gives one free trial key per Google account, limited to 10 requests a minute, with models including Qwen-SEA-LION v4.5 27B (76.51 on SEA-HELM Indonesian, function calling supported) and Llama-SEA-LION v3.5 70B for reasoning. AI Singapore publishes no price or end date for the trial key. It is not yet in our catalog, and it does not help with Spanish or Portuguese. On the table above the general Gemma and Qwen models score higher in Indonesian than this fine-tune, so use SEA-LION for its other languages or its function calling rather than for Indonesian alone.

Steps: sign in to the SEA-LION Playground with Google, open the API Key Manager, click Create New Trial API Key and save the key (it is shown once), then call https://api.sea-lion.ai/v1 with any OpenAI-compatible client.

Source: SEA-LION API guide, read 2026-11-01 (limit dated 2026-06-04).

How to choose #

• Best free quality in Indonesian: Gemini 3.8 Flash on the Gemini API free tier.
• Best open model: Gemma 4 31B on the Gemini API free tier or OpenRouter. Keep a second provider configured, since OpenRouter's free models allow only 50 requests a day until you buy credits.
• Many requests a day: GPT OSS 120B on Groq (1,000 requests a day). It scored 72.28 on SEA-HELM Indonesian, about seven points below Gemma 4 31B, and 85.3 to 85.9 on MMMLU in Portuguese and Spanish at high effort.
• Spanish or Portuguese: start with GPT OSS 120B (the only free model with a published per-language score for them) and test Gemini and Gemma 4 on your own prompts.
• Your own small model: download open weights and run them locally (how to run AI locally for free).

Whatever you pick, test it on 20 or 30 of your own prompts first. A leaderboard average hides weak spots, and these benchmarks mostly test exam-style tasks. If your task is translation rather than chat, see free translation APIs; for a language with a Vietnamese focus, see the Vietnamese version of this guide.

Limits and the catch #

• Other languages cost more tokens. The same sentence usually takes more tokens in Spanish, Portuguese or Indonesian than in English, so per-token free limits go less far: why other languages use more tokens.
• Benchmarks are not your task. MMMLU, MMLU-ProX and SEA-HELM measure exams and generic tasks; a model that scores two points higher may still be worse at your domain, or in the Spanish of Mexico against Spain.
• Different benchmarks, different scales. An 85 on MMMLU and an 80 on SEA-HELM are not comparable numbers; compare rows only within one table.
• Free tiers may train on your data. Google may use Gemini API free-tier prompts to improve its products; keep personal data out of free tiers unless the provider says it does not train on them.
• Scores and free models change. Leaderboards add models and free lineups rotate. We re-check monthly; the date at the top of this page is our latest check.