All models
    MIST QG family · Question generation

    MIST-QG-1.1

    The questions a passage answers, in 324 languages, as many as you ask for.

    MIST-QG-1.1 reads a passage and writes the questions it answers, in the passage’s own language or across languages, and you can say how many you want. It is a 4B model trained on 324 languages, built to make the (question, passage) pairs that embedders and rerankers learn from, in languages where those pairs barely exist.

    Also in MIST QG:MIST-QG-1.0
    Previous generation:mist-qg-1.5b
    MIST-QG-1.1 banner
    324
    Languages
    77%
    Cross-lingual questions usable, best tested
    0.701
    Questions that find their own passage (R@1)
    4B
    Parameters
    / Why it exists

    The problem it solves

    Search and RAG systems learn from pairs of a question and the passage that answers it. In English there are millions of them. In most of the world’s languages there are almost none, so retrieval quietly works worse for most of the world.

    MIST-QG-1.1 turns raw text in any of 324 languages into those pairs. It is small enough to run over millions of passages on one GPU, and its questions are written to find their own passage again, which is what a retrieval model needs to learn from.

    / Benchmarks

    The numbers

    Same language, all 238 test languages

    Same language, all 238 test languages

    46% of its questions are usable: on par with Qwen3.5-27B and ahead of Llama 3.3 70B. Only Gemma 4 31B, nearly eight times its size, scores higher.

    Cross-lingual

    Cross-lingual

    English or French questions on passages in 153 other languages: 77% usable, the best of every model tested.

    By how much training data a language had

    By how much training data a language had

    From 90 major languages down to 83 rare ones. In the rarest languages it scores 28%, against 15–25% for Qwen, gpt-oss and Llama up to 70B; Gemma 4 31B leads there at 32%.

    / What you get

    Highlights

    Bigger-model quality at 4B

    Averaged over 238 test languages, 46% of its questions pass every check, on par with Qwen3.5-27B (47%) and ahead of Llama 3.3 70B, Qwen3.5-9B and gpt-oss-20b. Gemma 4 31B, nearly eight times its size, scores higher at 51%.

    Best across languages

    English or French questions on passages in 153 other languages: 77% usable, more than every model tested, Gemma 4 31B (67%) included. About 1.75 questions per passage, against one for MIST-QG-1.0, so more usable pairs from every passage.

    You choose how many

    Ask for “3 questions in Swahili” and get three. Cross-lingual, it returned exactly the number asked every time from 1 to 5; in the same language it treats the number as a maximum instead of padding.

    More directions

    Any of 324 languages into English or French, plus two directions 1.0 can’t do: English or French into 30 widely spoken languages (59% usable, 99.3% in the right language) and between any two of those 30.

    / Quick start

    Run it in a few lines

    $pip install vllm
    questions.py
    from vllm import LLM, SamplingParams
    
    def prompt(passage, lang, qlang, n=None):
        ask = f"{n} {'question' if n == 1 else 'questions'}" if n else "Questions"
        return (f"Passage ({lang}):\n{passage}\n\n"
                f"{ask} in {qlang} that this passage answers, in your own words:\n")
    
    llm = LLM(model="olaverse/MIST-QG-1.1", max_model_len=4096)
    params = SamplingParams(max_tokens=320, temperature=0.3, top_p=0.9, stop=["\n\n"])
    
    passage = "The Nile is the longest river in Africa. It flows north for about 6,650 km..."
    prompts = [prompt(passage, "English", "English"),        # the model decides how many
               prompt(passage, "English", "Swahili", n=3)]   # three questions in Swahili
    for out in llm.generate(prompts, params):
        print([l[2:].strip() for l in out.outputs[0].text.split("\n") if l.startswith("- ")])
    / MIST QG family

    Which MIST QG model?

    On the Olaverse API as mist-qg-1.1, with n: "auto" to let it decide how many. Would rather self-host a model that always lets the passage decide? Use MIST-QG-1.0.

    MIST-QG-1.1THIS PAGEMIST-QG-1.0 →
    Parameters4B4B
    Passage languages324324
    How many questionsYou choose 1 to 5: exact every time cross-linguallyThe passage decides
    Question languagesOwn language, English, French, and 30 moreOwn language, English or French
    English or French → 30 languages59% usable, 99.3% in the right languageNot supported
    Between two of those 30 languages47% usableNot supported
    Cross-lingual questions per passageAbout 1.75About 1
    Questions that find their own passage (R@1)0.701, the best tested0.696
    Best forA set number of questions, in more languages and directionsLetting the passage decide how many
    On the Olaverse APIYesNo, self-host only
    / Where it shines

    Best for

    • Training pairs for embedders and rerankers in languages that have none.
    • Cross-lingual retrieval data: English or French questions over passages in any language.
    • Test sets that check your search finds the right passage, in every language you serve.
    • Generating questions over millions of passages on one GPU, with your text kept in-house.
    / Built with
    Qwen3.5-4B-Base vLLM llama.cpp (GGUF) Apache-2.0