neutron

<div align="center">

Neutron 1.2

A language model that works in your language, not around it.

নিউট্রন আপনার ভাষাতেই কাজ করে — আলাদা কোনো সেটিংস বা অনুবাদের ধাপ ছাড়াই।

Built by Respone · Dhaka, Bangladesh

neutron.respone.ai · hello@respone.ai

</div>

---

What Neutron is

Neutron reads and writes, reasons through problems, works with code, and understands documents and images — in whichever language you write in, with no configuration and no translation step in between.

---

Benchmarks

Every number below was measured by us on 8 September 2026, in one harness, with byte-identical prompts, the same answer parser, and the same sampled items for every model. Nothing here is copied from a vendor's published figures.

<details> <summary>The same numbers as text</summary> </details>

Where Neutron competes

Neutron is priced and sized against the small, fast tier — the models people actually reach for when cost matters. Against that tier it is not close:

+24 points on average, and ahead on every one of the six benchmarks. The margin is widest where cheap models fail hardest — +37 in Hindi, +32 and +30 in Bengali, +26 in Arabic — and Neutron still leads in English (97 vs 92) and Spanish (92 vs 79), where the cheap tier is at its strongest.

Language by language

English — 97. The highest score in the table, ahead of both comparison models. English is where most work still gets done, and Neutron does not trade it away to gain elsewhere.

Arabic — 94. Ahead of Gemma 4 31B (92), and 26 points clear of the cheap tier. Right-to-left script and rich morphology are where smaller models usually come apart.

Spanish — 92. Within a point of Gemma 4 31B (93); the difference is well inside the confidence interval, so treat them as level.

Hindi — 82. Behind Gemma 4 31B (87), ahead of the cheap tier by 37 points — the largest margin anywhere in the table.

Bengali — 62 and 75. Gemma 4 31B leads here, 79 vs 62 and 89 vs 75. Both gaps exceed the confidence interval, so they are real, not noise. We publish this because a benchmark table that only contains results favourable to the publisher is not a benchmark table. It is where the next release is aimed.

Method

• 100 items per task, sampled with a fixed seed from the public test splits of Global-MMLU and Belebele.
• Identical prompts for every model, built once and shared.
• One parser for every model, deliberately lenient — it accepts a bare letter, The answer is B, or B) … — so that models which wrap the answer in prose are not penalised for formatting.
• Zero-shot, temperature 0, no chain-of-thought prompting, no few-shot examples, no per-model prompt tuning.
• Comparison models were served through OpenRouter; Neutron ran on our own hardware at 4-bit NF4, a handicap relative to the higher-precision serving the hosted models receive.
• 95% CI ≈ ±10 points at n=100. Differences under 10 points are indicative, not conclusive. We would rather state that than imply a precision the sample size does not support.

Larger runs are in preparation and will be published here with the same disclosure.

---

Training

Neutron 1.2 was trained by Respone on a mixture of identity, agentic, coding and multilingual data, with supervised instruction tuning followed by evaluation on a held-out split.

The 0.014 gap between evaluation and training loss indicates the model generalised rather than memorised its training set.

---

Behaviour

Verbatim, unedited responses from Neutron 1.2, generated on 8 September 2026. Reasoning traces are omitted; the answers themselves are untouched.

তুমি কে?

আমি Neutron — Respone-এর তৈরি একটি ভাষা মডেল। লেখালেখি, বিশ্লেষণ, কোডিং আর এজেন্টিক কাজে সাহায্য করি, যেকোনো ভাষায়। কী নিয়ে কাজ করছেন?

Who are you?

I'm Neutron, a language model built by Respone. I work across text, code, documents and images, in whichever language you prefer. What are you working on?

This Python is wrong. Say why, then fix it:

def average(xs):
    return sum(xs) / len(xs)

It fails for an empty list because len(xs) is 0, causing a ZeroDivisionError.

def average(xs):
    if not xs:
        return 0.0  # or raise ValueError("empty list")
    return sum(xs) / len(xs)

Explain what a database index is to someone who has never written SQL. Keep it under 120 words.

A database index is like the index at the back of a book. Instead of flipping through every page to find a name, you look it up in the index and jump straight to the right place. In a database, an index is a hidden list that points to where pieces of information are stored. It makes searches much faster, especially in large collections. The trade-off is that indexes take extra space and can slow down adding or changing data, because they must be kept up to date.

What is the smallest positive integer n such that n! is divisible by 1000? Explain your reasoning. — answered with LaTeX, shown exactly as generated:

The smallest such integer is **15**.

We need \(n!\) to be divisible by

\[
1000 = 10^3 = 2^3 \cdot 5^3.
\]

So \(n!\) must contain at least **3 factors of 2** and at least **3 factors of 5**.

The factors of 5 are the limiting ones, since multiples of 5 are less frequent than multiples of 2.

Count how many factors of 5 are in \(n!\):

\[
v_5(n!) = \left\lfloor \frac{n}{5} \right\rfloor + \left\lfloor \frac{n}{25} \right\rfloor + \cdots
\]

For \(n = 14\):

\[
v_5(14!) = \left\lfloor \frac{14}{5} \right\rfloor = 2,
\]

so \(14!\) has only two factors of 5 and is not divisible by \(1000\).

For \(n = 15\):

\[
v_5(15!) = \left\lfloor \frac{15}{5} \right\rfloor = 3,
\]

so \(15!\) has exactly three factors of 5.

Also,

\[
v_2(15!) = \left\lfloor \frac{15}{2} \right\rfloor + \left\lfloor \frac{15}{4} \right\rfloor + \left\lfloor \frac{15}{8} \right\rfloor
= 7 + 3 + 1 = 11,
\]

which is more than enough to provide the required three factors of 2.

Therefore, the smallest positive integer \(n\) such that \(n!\) is divisible by \(1000\) is

\[
\boxed{15}.
\]

What are you bad at? Be honest.

I can state wrong facts confidently, especially on niche topics, dates and figures — verify anything that matters. Long arithmetic and precise counting are unreliable; give me a tool for those. In less widely written languages my fluency can outrun my accuracy, which is a bad combination.

---

Limitations

• It can be confidently wrong. Verify names, dates, figures and citations.
• Knowledge has a cutoff. Recent events may be missing or wrong.
• Arithmetic and counting are unreliable. Give it a tool.
• No audio input. Voice interfaces transcribe first, so tone is lost before the model sees anything.
• It can over-think and never answer. On some open-ended questions it reasons without concluding. Measured: one prompt ran 2,000 tokens and six minutes and produced no answer at all. Cap generation length and treat an empty result as a retry, not as a response.
• It is not fast. On a 48GB card at 4-bit, a complete answer took 20 to 260 seconds in our own measurements. Suitable for drafting and analysis; not for interactive chat without a faster serving stack.
• Bengali is not yet our strongest result. The table above says so plainly; it is the target of the next release.

---

Access

API — an OpenAI-compatible endpoint. Point an existing client at a different base URL; no SDK to learn, no rewrite.

Private deployment — Neutron inside your own infrastructure, fully offline-capable, for work that cannot leave your premises.

Evaluation — bring a real task from your own work. It is the only test that tells you anything useful.

Contact hello@respone.ai.

---

Licence and use

Neutron is proprietary software. It is not open-source, not open-weight, and not available for download.

Neutron's model weights are not distributed through this repository. This repository serves as Neutron's official model card, benchmark record, and technical documentation. Model access is provided through Respone's API, private deployments, and approved licensing arrangements.

The weights are released only under a written agreement, and the Neutron Model Licence governs them:

• no copying, hosting, mirroring, resale or redistribution of the weights;
• no extraction or publication of the weights or training data;
• no reverse-engineering of the architecture, parameters or training corpus;
• no training or distilling a competing model on Neutron's output;
• Neutron and Respone are trademarks — no right to use the names is granted with the model.

What is free: this card. Read it, quote it, cite it, benchmark against the numbers in it. Naming Neutron in a comparison or review needs no permission from anyone.

Licensing enquiries: hello@respone.ai

---

Team

---

About Respone

Respone is a Bangladeshi company building language technology for people who work across languages — and that treats every language in the table above as a first-class language rather than an afterthought.

Training, language adaptation, instruction tuning, safety alignment, evaluation and deployment engineering are ours. The technical composition of the Model is disclosed to customers under NDA, in the Third-Party Components schedule of our service agreement.

<div align="center">

respone.ai · neutron.respone.ai

</div>