Beam reads தமிழ் well. Its tokenizer makes Tamil pay 6× for it.
80 million people speak Tamil, and most of them type it two ways: in Tamil script and in “Tanglish” (Tamil in Latin letters). I asked Beam-1.0-501B-A23B the same 50 questions in English, Tamil and Tanglish, plus 22 questions that only exist in Tamil, and measured accuracy, prompt tokens, reasoning tokens and dollars.
accuracy on the parallel set, in all three languages, at default and minimal reasoning effort
more input tokens for the same question in Tamil script (tokenizer: tokens per Tamil word)
more reasoning tokens when the question is in Tanglish; Beam quietly translates before it thinks
cost per query vs English (Tamil / Tanglish) at Baseten’s $0.40 in / $1.20 out per 1M
New in v2:900 Belebele reading-comprehension questions in four input modes, a tokenizer league against 7 other open models, a 56-item hard Tamil-script split, a Tamil writing test, and the Token Saver.
accuracy drop on Belebele when the same passage and question are in Tamil instead of English ()
Tamil tokens per English token on 488 FLORES passages. Beam shares Qwen3's tokenizer; Gemma 3 needs only
fewer input tokens when the Tamil passage is auto-converted to Tanglish. The catch: accuracy falls to and Beam thinks 2× longer, so the bill goes up 1.5×
cost of a Tamil Belebele query vs English, because Beam also thinks longer in Tamil
Belebele: same 900 questions, four ways to ask
Accuracy by input mode
English: all English. Tamil: all Tamil. Mixed: Tamil passage, English question and options (how most Indian apps actually prompt). Auto-Tanglish: the Tamil prompt machine-transliterated to Latin letters by the Token Saver below.
Cost per 1,000 questions and what it buys
Tokenizer league: what Tamil costs on 8 open models
Same 488 FLORES-200 passages (the Belebele source), tokenized with each model's own tokenizer.json. Bars are Tamil tokens ÷ English tokens: 1.0× would be no tax. Beam's counts are identical to Qwen3's on every passage, so Beam inherits Qwen's weak Tamil vocabulary (about 0.9 Tamil characters per token, so a Tamil letter often costs more than one token).
Hard Tamil: script-level tasks
on hard items × 3 samples. Answers are computed by code, not by hand.
Beam knows Tamil grammar, ancient numerals, Thirukkural and the calendar cold. It breaks when it has to manipulate letters, like reversing a word or counting letters across a sentence. That's what you'd expect when one Tamil letter is split across byte-level tokens.
Does Beam write good Tamil?
fluency / 5 on 15 open-ended prompts (rent advance, UPI fraud, dengue, a complaint letter…), judged by Gemma-3-27B and DeepSeek-V3.1
correctness / 5, same judges
Tanglish questions answered in Tanglish. The others switched to Tamil script or a Tamil-English mix (checked by script share, not by a judge)
more output tokens for a Tamil-script answer than the same answer in Tanglish. Output is billed at 3× input
v1: the parallel set (English / Tamil / Tanglish)
Where the tokens go (avg per question, default effort)
Tamil script is expensive to read; Tanglish is cheap to read but expensive to think about. They land at almost the same price.
Cost per 1,000 questions: default vs reasoning_effort="minimal"
Minimal effort kept 100% accuracy in every language here and cut the bill by .
Reasoning tokens by task type
Logic puzzles in Tanglish make Beam think ~4× longer than the same puzzle in English, the biggest gap in the set.
Token Saver: paste Tamil, see what Beam will charge
Runs Beam’s real tokenizer in your browser, converts your Tamil to Tanglish with the same transliterator used in the Belebele run, and estimates what other models would charge for the same text.
Loading tokenizer…
Auto-Tanglish version
Same text on other models (estimated)
Native Tamil: questions with no English twin
accuracy on 22 native items × 3 samples: Tamil number words, ancient numerals (௲ ௱ ௰), letter counting by Tamil grammar, Sangam literature, grammar, the Tamil calendar
avg reasoning tokens on native items, about 2.6× the parallel Tamil set: counting எழுத்துகள் through byte-level tokens takes real work
The one miss
Asked for the last word of Kural 1 (அகர முதல… உலகு), one of three samples answered “என்று” (“that”), echoing the prompt’s phrasing instead of the couplet.
What this means if you build for Tamil users
v2: Beam is measurably weaker reading Tamil than English on Belebele, and the Tamil prompt costs about 4× more. Keep instructions and options in English when your content is Tamil: the mixed mode gets most of the accuracy back.
v2: keep the Tamil passage, ask in English. The Mixed mode scored 95.6% (vs 88.3% all-Tamil) at two-thirds of the all-Tamil cost. It is the best setup for Tamil RAG on Beam.
v2: don't auto-transliterate to save money. Auto-Tanglish cut input tokens 56%, but accuracy dropped to 80.3%, 32 of 900 answers ran out of tokens, and total cost rose 1.5× over Tamil script. Cheap to read, expensive to think.
v2: pick the model by tokenizer for Tamil-heavy RAG. On the same text, Gemma 3 uses about a quarter of Beam's tokens.
Budget ~1.7× English per Tamil or Tanglish query on Beam. Script choice barely changes the price; it moves cost between input and reasoning.
Turn reasoning down.reasoning_effort="minimal" lost nothing on everyday Tamil tasks and saved ~16–21%.
Long Tamil documents hurt most. At ~0.9 characters per token, Tamil RAG context costs ~3.7× English: summarise or translate retrieved chunks before stuffing a 1M context.
Tanglish is the cheaper input for chat UIs (2.5 chars/token) if you can tolerate the extra thinking; for short-answer apps Tamil script is slightly cheaper overall.
Beam is ready for Tamil. 100% on 300 parallel runs and 98.5% on native tasks means accuracy isn't the problem. Cost is.
Every answer, with Beam's reasoning
ID
Lang
Question
Gold
Beam
In
Think
Tap a row to read Beam's reasoning trace (first 700 characters).
Method
50 parallel items (20 math word problems with Indian context such as auto fares, rent splits and cents of land, 8 logic, 15 Tamil culture and knowledge, 7 everyday), each written in English, Tamil script and colloquial Tanglish, plus 22 native-Tamil items. Every prompt ends with “answer only” in its own language. Model: reflection/Beam-1.0-501B-A23B via Hugging Face Inference Providers (Baseten), max_tokens 6000. Parallel set: 2 samples per language at default effort and 1 at minimal effort; native set: 3 samples. Scoring: last number for numeric answers, alias match for words (Tamil or Latin script, Tamil digits normalised). Token counts come from the API's usage field; tokenizer stats come from Beam's own tokenizer.json. Cost uses Baseten list price ($0.40/1M input, $1.20/1M output incl. reasoning) as of Oct 2026. Total spend for the whole study: .
v2 method: Belebele (facebook/belebele, tam_Taml + eng_Latn, 900 aligned questions) × 4 input modes, default effort, one sample, scored on the option number. Tokenizer league: 488 unique FLORES passages through each model's tokenizer.json (Beam, Qwen3-8B, Gemma-3-27B, Llama-3.3-70B, DeepSeek-V3, gpt-oss-120b, Sarvam-M, Mistral-Small-3.1). Hard split: 56 items with gold computed by code (letter counts by Tamil grammar, grapheme reversal, uyirmei split, traditional numerals, Thirukkural endings, calendar), 3 samples. Writing: 15 prompts, judged 1–5 by two non-Beam models; register measured by script share.
Limits: one sample per Belebele question; judges are LLMs, not native-speaker raters; the transliterator is a simple rule-based scheme, not how every Tamil speaker types Tanglish.