SEA-LION v4.5 vs Gemma 4 31B vs Qwen3.8 27B for Southeast Asia

Short answer: SEA-LION for seven Southeast Asian languages and short text; Gemma 4 31B or Qwen3.8 27B for long documents, images and tools

LLMTR is a gateway that gives access to language models from different providers through one OpenAI-compatible API. Qwen-SEA-LION-v4.5-27B-IT (marscompute/qwen-sea-lion-v4.5-27b), Gemma 4 31B (gemma/gemma-4-31b-it) and Qwen3.8 27B (tensorx/qwen3.8-27b-tensorx) are called on the /v1/chat/completions endpoint at https://llmtr.com/v1 with the same API key from your llmtr.com account. For text jobs in one of the Southeast Asian languages the maker lists, and that fit in 8,192 tokens, use the marscompute/qwen-sea-lion-v4.5-27b id on LLMTR. When you need long documents, image input or tool calling, move to the gemma/gemma-4-31b-it or tensorx/qwen3.8-27b-tensorx id; for a JSON schema, move to the gemma/gemma-4-31b-it id.

This article does not rank quality and claims no quality advantage for any model. The comparison rests on the differences in price, context window and request shape in the LLMTR catalog, and on what the makers write about languages on their own model cards. Find out which model works in your language and for your job by trying them on your own examples.

SEA-LION v4.5, Gemma 4 31B or Qwen3.8 27B for Southeast Asian languages?

Four questions decide it: what language is the text in, do prompt and reply fit in 8,192 tokens, do you need images or tool calling, and what is the budget? The table shows the three rows with the values in the LLMTR catalog and documentation as of 30 October 2026; prices are per 1M tokens.

The language row comes from the makers' cards and is not a quality claim: AI Singapore lists the languages one by one, Google gives numbers but no list, and Qwen's card gives no language list at all. The account consent row shows the setting you have to make in your account for requests to work.

What does the same request cost on each model?

Take a 2,000-token prompt and a 500-token reply, for example a draft reply to a complaint written in Thai. On Qwen-SEA-LION-v4.5-27B-IT this request costs $0.003, on Gemma 4 31B $0.00048 and on Qwen3.8 27B $0.002. Over 1,000 requests the totals are $3.00, $0.48 and $2.00 respectively. Qwen3.8 27B thinks by default and the thinking tokens are billed as output: if 1,000 thinking tokens come on top of the same reply, the request costs $0.0044. Among these three rows Gemma 4 31B has the lower price; language, window and request shape weigh in the choice as much as price does.

The Python example below sends the same Thai message to all three models and works out the cost from the usage field of each response; the price table holds the catalog prices of 30 October 2026. tensorx/qwen3.8-27b-tensorx returns 403 provider_consent_required until the TensorX data-transfer consent is given on your account; in that case the example prints the error and carries on with the other models. Token counts differ from model to model; compare the replies on your own examples.

OpenAI Python SDK example that sends the same Thai message to Qwen-SEA-LION-v4.5-27B-IT, Gemma 4 31B and Qwen3.8 27B through https://llmtr.com/v1/chat/completions and works out the cost from the usage field.

import os

from openai import OpenAI, PermissionDeniedError

client = OpenAI(base_url="https://llmtr.com/v1", api_key=os.environ["LLMTR_API_KEY"])

# Per 1M tokens on 30 October 2026: (input, cache read, output); None = no cache price.
PRICES = {
    "marscompute/qwen-sea-lion-v4.5-27b": (0.60, None, 3.60),
    "gemma/gemma-4-31b-it": (0.14, None, 0.40),
    "tensorx/qwen3.8-27b-tensorx": (0.40, 0.10, 2.40),
}
messages = [
    {"role": "system", "content": "You are a support agent. Reply in Thai, politely, in at most three sentences."},
    {"role": "user", "content": "สั่งรองเท้าไปสองคู่ แต่ได้รับมาแค่คู่เดียวค่ะ"},
]

for model, (inp, cache, out) in PRICES.items():
    try:
        # 2,048 gives Qwen3.8 27B's thinking more room (raise it if finish_reason is length) and, with this short prompt,
        # stays within the 8,192 tokens SEA-LION shares between prompt and reply.
        response = client.chat.completions.create(model=model, messages=messages, max_tokens=2048)
    except PermissionDeniedError as error:
        # tensorx/qwen3.8-27b-tensorx answers 403 provider_consent_required until the TensorX consent is given.
        print(model, error.status_code, error.body)
        continue
    usage = response.usage
    details = usage.prompt_tokens_details
    cached = (details.cached_tokens or 0) if details and cache is not None else 0
    cost = ((usage.prompt_tokens - cached) * inp + cached * (cache or 0) + usage.completion_tokens * out) / 1_000_000
    print(model, response.choices[0].message.content, usage.prompt_tokens, usage.completion_tokens, f"${cost:.6f}")

Language support: what do the makers write?

AI Singapore writes that the post-training of Qwen-SEA-LION-v4.5-27B-IT was carried out in English and Southeast Asian languages, and lists the languages one by one: Burmese, Indonesian, Filipino, Malay, Tamil, Thai and Vietnamese. The model's name says the same: SEA-LION stands for Southeast Asian Languages In One Network. Turkish is not on this list.

Google's Gemma 4 model card says the family offers out-of-the-box support for more than 35 languages and was pre-trained on more than 140 languages; the page does not list which languages those numbers include. Qwen's Qwen3.8-27B model card gives no list of supported languages. So if you are going to choose Gemma 4 31B or Qwen3.8 27B for a job in Thai, Indonesian or Vietnamese, check the result in that language on your own examples; this article says no more about the two models' languages than their cards do.

• AI Singapore, Qwen-SEA-LION-v4.5-27B-IT model card · 2026-10-30
• Google, Gemma 4 model card · 2026-10-30
• Qwen, Qwen3.8-27B model card · 2026-10-30

Which one when?

The choices below follow from the differences in price, window, request shape and the makers' language statements. Compare output quality on your own examples.

• A support reply, translation or summary that fits in 8,192 tokens, in one of the seven Southeast Asian languages the maker lists: Qwen-SEA-LION-v4.5-27B-IT on LLMTR, marscompute/qwen-sea-lion-v4.5-27b, at $0.60 input and $3.60 output per 1M tokens.
• A long document or a long conversation that does not fit in 8,192 tokens: Gemma 4 31B or Qwen3.8 27B, both with a 262,144-token context window.
• A request with an image, such as a photographed receipt or a screenshot: Gemma 4 31B or Qwen3.8 27B; Qwen-SEA-LION-v4.5-27B-IT refuses a request carrying an image part with 400.
• An agent step that calls tools, or JSON that follows a schema: Gemma 4 31B, gemma/gemma-4-31b-it, with tool calling, json_object and json_schema. Qwen3.8 27B calls tools as well.
• High-volume short jobs where budget comes first: Gemma 4 31B; among these three rows, Gemma 4 31B has the lower input and output price.
• Jobs that repeat the same long system message: on Qwen3.8 27B, tokens reported as cached_tokens cost $0.10 per 1M tokens, and the other two rows have no cache discount. Output costs $2.40 there, so compare the total with Gemma 4 31B at your own prompt and reply sizes.

Moving between the models

Moving is usually a matter of changing the model field: all three models are called on the same endpoint with the same OpenAI-compatible request shape. When you move from Qwen-SEA-LION-v4.5-27B-IT to Gemma 4 31B or Qwen3.8 27B, you no longer have to fit max_tokens and the prompt into 8,192 together, and you can add tool definitions or image parts. On Qwen3.8 27B thinking is on by default and the thinking tokens are billed as output; if the reply comes back with finish_reason length, raise max_tokens. reasoning_effort returns 400 on all three models; on Gemma 4 31B, turn thinking on with reasoning: true or :think.

In the other direction, from Gemma 4 31B or Qwen3.8 27B to Qwen-SEA-LION-v4.5-27B-IT, remove tools, response_format and image parts from the request, check that the prompt text stays under about 32,000 characters, and set max_tokens so that it and the prompt stay within 8,192. For tensorx/qwen3.8-27b-tensorx, give the TensorX data-transfer consent on your account from the Data-transfer consents card under Dashboard > Settings; without it the request returns 403 provider_consent_required.

Comparing SEA-LION v4.5, Gemma 4 31B and Qwen3.8 27B on the same job

Send the same message to all three models, compare the reply, token usage and cost, and choose the model id for your own language and job.

1. Key and consent. Create an API key on llmtr.com, top up your balance and, for tensorx/qwen3.8-27b-tensorx, give the TensorX data-transfer consent under Dashboard > Settings.
2. Same message. Send your own texts with all three model ids to the https://llmtr.com/v1/chat/completions endpoint; on Qwen-SEA-LION-v4.5-27B-IT keep prompt and max_tokens together below 8,192.
3. Choice. Have the replies checked by a native speaker of the language, compare token counts and cost, then choose the model id.

Frequently asked questions

Does Qwen-SEA-LION-v4.5-27B-IT support Turkish?

No. Turkish is not among the languages AI Singapore lists on the model card. For Turkish text, pick another model and test it on your own examples; this article says nothing about Turkish for Gemma 4 31B and Qwen3.8 27B.

What does a request with a 2,000-token prompt and a 500-token reply cost?

$0.003 on Qwen-SEA-LION-v4.5-27B-IT, $0.00048 on Gemma 4 31B and $0.002 on Qwen3.8 27B. On Qwen3.8 27B the thinking tokens also count as output.

Why does tensorx/qwen3.8-27b-tensorx return 403?

The account has not given the TensorX data-transfer consent. Give it from the Data-transfer consents card under Dashboard > Settings; without it the request returns 403 provider_consent_required and is not billed.

Related posts

• Qwen-SEA-LION-v4.5-27B-IT API: first request with LLMTR
• A language-model API for Thai, Indonesian, Vietnamese and Malay
• SEA-LION's 8,192 tokens: long text, max_tokens and cost

• Qwen-SEA-LION-v4.5-27B-IT model
• Gemma 4 31B model
• Qwen3.8 27B (TensorX) model
• MarsCompute documentation
• Gemma 4 31B documentation
• TensorX documentation
• Tool calling documentation
• Multimodal images