The SoreQen S1 series
Three open-weight reasoning models at 0.8B, 2B and 4B parameters. S1 is where the line starts, not what it is — later generations will be more capable, and the direction does not change.
The three
| Model | Parameters | Context | Max reply |
|---|---|---|---|
| SoreQen S1 Mini | 0.8B | 256k | 32k |
| SoreQen S1 | 2B | 256k | 32k |
| SoreQen S1 Mega | 4B | 256k | 32k |
SoreQen S1 Mini
0.8Bsoreqen-s1-miniFastest. Good for quick questions and everyday chat.
Reasoning is a switch on this model rather than a dial. Its useful range is “don’t think”, “think a little” and “think harder”; asked for more, a 0.8B model spends its whole output allowance reasoning and returns no answer. That is measured, not assumed.
SoreQen S1 Mega
4Bsoreqen-s1-megaDeepest reasoning. Best for hard problems and long documents.
What every one of them can do
Capability is not a tier here. The chat product gates some of this by plan because a subscription is an allowance; the API does not, because a call is a purchase.
- Vision — images in the message content
- Tool calling, with results coming back as turns
- Structured output against a strict JSON Schema
- Streaming, with a usage frame at the end
- Reasoning traces you can read
- 256k of context on all three
Evaluation
One evaluation has been run and published: 24 Hinglish prompts, SoreQen S1 Mega against the Qwen3.5 4B base it was fine-tuned from, scored on measurable properties of the answer rather than on preference.
| Measure | S1 Mega | Base |
|---|---|---|
| Stayed in Roman scriptNeither model slipped into Devanagari. The product promises Roman-script Hinglish. | 24 | 24 |
| Genuinely code-mixedAt least three Hindi markers and two English content words in the same answer. | 20 | 15 |
| Carried sequence or causeWords like pehle, phir, kyunki — the difference between an answer and a greeting. | 23 | 24 |
| Drifted into EnglishAnswering a Hinglish question wholly in English. One case, against none. | 1 | 0 |
| Produced the artefact asked forOf the three prompts that asked for one. Too few to read anything into. | 3 | 3 |
| Hindi markers per answerMean across all 24. Higher means more actual code-mixing. | 22.5 | 21.1 |
What this does not tell you
Twenty-four prompts is a small sample, and one evaluation of one model is not a benchmark suite. It says the fine-tune measurably increased code-mixing — twenty answers against fifteen — and left informativeness roughly where it found it. It says nothing about maths, coding, long-context recall or tool use.
We have not run MMLU, GSM8K, HumanEval or any of the public leaderboards, so there are no scores for them on this page. A table of numbers nobody produced is the thing every model page has and the reason nobody believes any of them. When those runs happen, the results go here whichever way they come out.
Running them yourself
Every model is published in safetensors and in GGUF. Nothing needs to be signed into and nothing phones home. Two quantisations are built for each: Q4_K_M (the default. best size-to-quality trade; runs on a modest laptop) and Q8_0 (near-lossless. take it if you have the ram).
All three derive from Qwen3.5 (Alibaba Cloud, Apache 2.0). The full statement of what was changed, and under what terms each artefact is released, is on the open-source page rather than repeated here.