Introducing ZorQelis AI and the SoreQen S1 models

ZorQelis AI is an AI research company based in New Delhi, India. It was founded on 24 August 2026 by Ved Saini, who leads it as Founder and CEO. We build agentic models with deep reasoning: models that hold a problem across many steps, decide what to do next, and notice when an earlier step was wrong.

SoreQen is where that work ships. It is a chat assistant at soreqen.com, a developer API at platform.soreqen.com, and a desktop app for Windows and Linux. All three run the same family of models, the SoreQen S1 series, which we trained in August 2026 and published with open weights. This post is the introduction we did not write at launch.

The SoreQen S1 models

There are three sizes, all trained with one recipe. They are deliberately small: cheap enough to offer generously and fast enough to feel immediate, rather than large enough to top a leaderboard.

Each one is a low-rank fine-tune of Qwen3.5, the open model family from Alibaba Cloud released under Apache 2.0. The fine-tune taught Hinglish, tool use, structured output and professional writing, and it was built to keep what the base already did well: its thinking behaviour, vision, context length and output ceiling. On the API, all three models take vision, tool calling, structured output and every reasoning setting the model supports, with a 256k-token context window. There is no tier that unlocks a feature. The models in detail.

Why Hinglish

Most Indians who type to an assistant do not write formal English or Devanagari Hindi. They write Hinglish: Hindi and English mixed in one sentence, in Roman script, like "yaar ye samajh nahi aa raha, explain karo". Most assistants treat that as broken English, answer in English, and lose the register the question was asked in.

SoreQen reads Hinglish as Hinglish and answers in it. We measured that rather than asserting it. On 24 Hinglish prompts, SoreQen S1 Mega gave a genuinely code-mixed answer 20 times, against 15 for the Qwen3.5 4B base it was trained from. It also drifted wholly into English once, where the base did not.

24 prompts is a small sample, and we say so. An evaluation that only reports its wins is marketing, so the one result that went against us is published beside the ones that did not.

Open weights, and the data behind them

Weights alone let you run a model; they do not let you check it. So we published both. The three models are on Hugging Face in safetensors and in GGUF (Q4_K_M and Q8_0) for running locally, and the 600,229 training rows they learned from are published as three datasets:

• SoreQen Hinglish: 36,326 training rows. 13,657 rows rejected for too little Hindi.
• SoreQen Reasoning: 540,175 training rows. 6,293,545 near-duplicate rows removed.
• SoreQen Writing: 23,728 training rows. 17,162 truncated rows discarded.

The rejection counts matter as much as the row counts. How big a corpus is tells you little; how much was thrown away, and why, tells you whether anyone looked at it. As far as we know, the Hinglish set is the only curated public instruction corpus of its kind. Weights, data and licences.

What went wrong on the way

The S1 models came out of the second of two training campaigns. The first, from scratch, was a mixture-of-experts model trained from nothing, and it did not ship. The second, fine-tuning Qwen3.5, produced the models you can download. Across both, we recorded forty-five failures.

The ones worth writing down were silent. A run that crashes teaches you something immediately; a run that completes while doing the wrong thing teaches you nothing until you check. One checkpoint was quietly training from random initialisation in a run whose loss curve looked ordinary. A deduplication pass, correct for a maths corpus, deleted almost all of the small hand-written set that teaches a model what it is. Those rows are supposed to resemble each other.

Every filter in the pipeline now prints what it rejected and why, and every checkpoint is diffed against its base. The full postmortem is published, because a company that only reports its results is asking to be taken on faith. What went wrong, and why.

Where to use SoreQen today

• In the browser: soreqen.com. Free to start, and you can try it without an account: a signed-out visitor gets a working chat immediately.
• Through the API: the SoreQen API is OpenAI-compatible, so existing clients and SDKs work by changing the base URL and the key. Credit is prepaid per token, with no subscription.
• On the desktop: the SoreQen app for Windows and Linux keeps itself up to date, with every update signed and checked before it installs.
• On your own hardware: the weights and GGUF builds are at huggingface.co/zorqelis-ai.

What this blog is for

This is where we will write about the models, the research behind them and the company that makes them. We will hold it to the standard the rest of this site keeps: numbers that come from the artefact they describe, and results that include the ones that do not flatter us. There is an RSS feed if you would rather be told when something new is up.