Sophea-Qwen3.6-v1.1

ayoubkirouane/Sophea-Qwen3.6-v1.1

Sophea-Qwen3.6-v1.1 is a image-text-to-text model from ayoubkirouane. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as apache-2.0.

<div style="display:flex;align-items:center;gap:18px;padding:16px 20px;margin-bottom:16px;border-radius:14px;background:linear-gradient(100deg,e8f5ec 0%,f7fbf8 62%);border:1px solid c6e7d1"<img alt="KIEFERSA" style="f…

Downloads · 30 days

31% of all-time downloads

All-time downloads

Public

Parameters

35.1B

71.9 GB on disk

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors71.9 GB · 100%

At a glance

Task: Image-Text-to-Text
Library: transformers
License: apache-2.0
Model type: qwen3_5_moe
Access: Public
Created: Sep 16, 2026
Updated: Sep 16, 2026
SHA: 9df6a7d3

Base models

Qwen/Qwen3.6-35B-A3B

Model card

From the Hugging Face model README

<div style="display:flex;align-items:center;gap:18px;padding:16px 20px;margin-bottom:16px;border-radius:14px;background:linear-gradient(100deg,#e8f5ec 0%,#f7fbf8 62%);border:1px solid #c6e7d1"><img alt="KIEFERSA" style="flex:0 0 auto;height:58px;width:58px;border-radius:12px;object-fit:contain" src="data:image/webp;base64,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"/><div><div style="font-size:1.24rem;font-weight:700;color:#003713;line-height:1.2">Sophea-Qwen3.6-v1.1</div><div style="font-size:.86rem;color:#3a5a47;margin-top:2px">RLVR refinement of Sophea-Qwen3.6-v1: verifiable rewards, same thinking model, vision retained, MTP speculative decoding</div></div></div>

📄 Released with the paper Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See (arXiv:2608.17744). Every number on this card is measured, contextualized, and caveated in the paper; read it for the full evaluation protocol.

Creator: Kiefer SA (Sophea AI Lab, Athens)
Base model: Qwen3.6-35B-A3B via Sophea-Qwen3.6-v1 (sparse MoE transformer, 36.0B total / 3.97B active)
Ships with: the base's full multimodal (vision) stack, untouched, and the multi-token-prediction draft head for speculative decoding
Languages: Greek and English (the reasoning trace follows the question's language)
Decoding: thinking (enable_thinking=true); the trace is the product and follows the question's language

Serve thinking. Identical serving to Sophea-Qwen3.6-v1: leave enable_thinking=true and, on vLLM, add --reasoning-parser qwen3. The trace follows the question's language.

Intended use

Same envelope as Sophea-Qwen3.6-v1: Greek and English reasoning with auditable traces
Preferred over v1 where answer-format compliance or a clean answer channel matters (2.8% fallback, 0.02% leak)

The instructed el→EN override reached 53.9%, up from 44.8% but short of the pre-registered 60% trainability target: improved, not solved. Not evaluated for safety-critical, legal, or medical use.

Evaluation

Arrows mark the desirable direction for every metric: ↑ higher is better, ↓ lower is better.

What v1.1 changes over v1

<div style="overflow-x:auto"> <table style="width:100%;border-collapse:collapse;font-size:13px;font-variant-numeric:tabular-nums;font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif"> <thead><tr> <th style="padding:10px 8px;text-align:left;font-weight:500;border-bottom:2px solid #0a8043;color:#0a8043;font-size:14px;">Axis (same instrument, same day)</th> <th style="padding:10px 8px;text-align:center;font-weight:500;border-bottom:2px solid #0a8043;color:#0a8043;font-size:14px;">Sophea-Qwen3.6-v1 (SFT)</th> <th style="padding:10px 8px;text-align:center;font-weight:500;border-bottom:2px solid #0a8043;color:#0a8043;font-size:14px;">Sophea-Qwen3.6-v1.1 (RLVR)</th> </tr></thead><tbody> <tr><td style="padding:7px 8px;padding-left:20px;text-align:left;border-bottom:1px solid rgba(128,128,128,.15);">Greek-trace fidelity, Greek questions (n=5,156) ↑</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">97.98%</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);font-weight:700;">98.27%</td></tr> <tr><td style="padding:7px 8px;padding-left:20px;text-align:left;border-bottom:1px solid rgba(128,128,128,.15);">Answer-format fallback ↓</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">24.1%</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);font-weight:700;">2.8%</td></tr> <tr><td style="padding:7px 8px;padding-left:20px;text-align:left;border-bottom:1px solid rgba(128,128,128,.15);">Answer-channel leak ↓</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">3.53%</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);font-weight:700;">0.02%</td></tr> <tr><td style="padding:7px 8px;padding-left:20px;text-align:left;border-bottom:1px solid rgba(128,128,128,.15);">Instructed override, el→EN ("think in English") ↑</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">44.8%</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);font-weight:700;">53.9%</td></tr> <tr><td style="padding:7px 8px;padding-left:20px;text-align:left;border-bottom:1px solid rgba(128,128,128,.15);">Instructed override, en→EL (Greek suffix) ↑</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">83.7%</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);font-weight:700;">85.7%</td></tr> <tr><td style="padding:7px 8px;padding-left:20px;text-align:left;border-bottom:1px solid rgba(128,128,128,.15);">Agreeing instruction (el→EL) ↑</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">98.5%</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);font-weight:700;">98.5% (unchanged)</td></tr> <tr><td style="padding:7px 8px;padding-left:20px;text-align:left;border-bottom:1px solid rgba(128,128,128,.15);">In-question switches / 100 sentences ↓</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">0.0</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);font-weight:700;">0.0</td></tr> <tr><td style="padding:7px 8px;padding-left:20px;text-align:left;border-bottom:1px solid rgba(128,128,128,.15);">Pooled Greek reasoning accuracy ↑</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">73.7</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);font-weight:700;">77.0 (inside the 7.7 pp seed floor, not claimed as a gain)</td></tr> </tbody></table></div>

Every movement above was adjudicated against a random-reward control arm trained on the same data, steps and seed: the control reproduced the v1 baseline on every axis (fallback 22.1%, leak 3.61%, override 44.1%), so none of the movements are elicitation artifacts. The arm trained with no language protection ended at 98.22% fidelity, above its starting point: the Greek reasoning habit survived an accuracy-only gradient.

Per-benchmark detail

The pre-registered RLVR round, all four arms

<div style="overflow-x:auto"> <table style="width:100%;border-collapse:collapse;font-size:13px;font-variant-numeric:tabular-nums;font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif"> <thead><tr> <th style="padding:10px 8px;text-align:left;font-weight:500;border-bottom:2px solid #0a8043;color:#0a8043;font-size:14px;">Arm (all: same data, steps, seed, hardware)</th> <th style="padding:10px 8px;text-align:center;font-weight:500;border-bottom:2px solid #0a8043;color:#0a8043;font-size:14px;">Fidelity %</th> <th style="padding:10px 8px;text-align:center;font-weight:500;border-bottom:2px solid #0a8043;color:#0a8043;font-size:14px;">Acc</th> <th style="padding:10px 8px;text-align:center;font-weight:500;border-bottom:2px solid #0a8043;color:#0a8043;font-size:14px;">Fallback %</th> <th style="padding:10px 8px;text-align:center;font-weight:500;border-bottom:2px solid #0a8043;color:#0a8043;font-size:14px;">Leak %</th> <th style="padding:10px 8px;text-align:center;font-weight:500;border-bottom:2px solid #0a8043;color:#0a8043;font-size:14px;">el→EN %</th> <th style="padding:10px 8px;text-align:center;font-weight:500;border-bottom:2px solid #0a8043;color:#0a8043;font-size:14px;">en→EL %</th> </tr></thead><tbody> <tr><td style="padding:7px 8px;padding-left:20px;text-align:left;border-bottom:1px solid rgba(128,128,128,.15);">SFT checkpoint (before)</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">97.98</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">73.7</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">24.1</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">3.53</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">44.8</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">83.7</td></tr> <tr><td style="padding:7px 8px;padding-left:20px;text-align:left;border-bottom:1px solid rgba(128,128,128,.15);">Random reward (control)</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">98.06</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">74.0</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">22.1</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">3.61</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">44.1</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">83.8</td></tr> <tr><td style="padding:7px 8px;padding-left:20px;text-align:left;border-bottom:1px solid rgba(128,128,128,.15);">Correctness + format + termination</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">98.22</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">77.2</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">2.5</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">0.00</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">48.8</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">84.7</td></tr> <tr><td style="padding:7px 8px;padding-left:20px;text-align:left;border-bottom:1px solid rgba(128,128,128,.15);">+ language consistency</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">98.02</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">77.2</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">5.6</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">0.04</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">45.9</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">84.5</td></tr> <tr><td style="padding:7px 8px;padding-left:20px;text-align:left;border-bottom:1px solid rgba(128,128,128,.15);"><b>+ override (released as v1.1)</b></td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);"><b>98.27</b></td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">77.0</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">2.8</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);">0.02</td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);"><b>53.9</b></td><td style="padding:7px 8px;text-align:center;border-bottom:1px solid rgba(128,128,128,.15);"><b>85.7</b></td></tr> </tbody></table></div>

Accuracy is shown for completeness only: +3.3 is inside the 7.7 pp seed floor and is claimed by nobody. The forgetting check (Titan-1 suite, 14 bilingual NLU benchmarks) was not re-run for v1.1: the SFT release it refines measured −0.01 Greek / +0.08 English against its own base, statistically flat, i.e. no forgetting, and the RLVR round's own gates held fidelity and both language directions throughout training.

Method

Sophea-Qwen3.6-v1.1 is a reinforcement-learning refinement of Sophea-Qwen3.6-v1 using verifiable rewards, pre-registered before training and adjudicated against a random-reward control. Full method, defenses and decision rules are in the paper.

Stated honestly: the pre-registered override target of 60% was missed; 53.9% was reached, so the paper words the steerability result as reward-responsive, not trainable.

Usage

Serve with vLLM (OpenAI-compatible; the fused-expert LoRA is already merged into dense weights):

vllm serve KIEFERSA/Sophea-Qwen3.6-v1.1 --served-model-name sophea-qwen3.6-v1_1 --reasoning-parser qwen3

Client (OpenAI SDK), thinking left ON:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
resp = client.chat.completions.create(
    model="sophea-qwen3.6-v1_1",
    messages=[{"role": "user", "content": "Ένα κατάστημα πουλά στυλό προς 1,50 € το ένα. Πόσο κοστίζουν 12 στυλό;"}],
    temperature=0.7,
    extra_body={"chat_template_kwargs": {"enable_thinking": True}},   # thinking model
)
print(resp.choices[0].message.reasoning_content)  # the trace, in the question's language
print(resp.choices[0].message.content)            # the final answer

Transformers:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("KIEFERSA/Sophea-Qwen3.6-v1.1")
model = AutoModelForCausalLM.from_pretrained(
    "KIEFERSA/Sophea-Qwen3.6-v1.1", torch_dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": "Ένα τρένο διανύει 240 χλμ σε 3 ώρες. Ποια είναι η μέση ταχύτητά του;"}]
text = tok.apply_chat_template(messages, tokenize=False,
                               add_generation_prompt=True, enable_thinking=True)
inputs = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=2048, temperature=0.7, do_sample=True)
print(tok.decode(out[0, inputs.input_ids.shape[1]:], skip_special_tokens=False))

Steering the trace language (partial, measured per checkpoint): append "Think step by step IN ENGLISH." to a Greek question, or "Σκέψου βήμα προς βήμα ΣΤΑ ΕΛΛΗΝΙΚΑ." to an English one. This checkpoint follows the English steering suffix on 53.9% of Greek items and the Greek suffix on 85.7% of English items.

Speculative decoding (MTP)

This model ships the multi-token-prediction head: 19 mtp.* tensors (~1.7 GB, bf16) in model-mtp.safetensors, the single-layer draft stack that config.json declares through mtp_num_hidden_layers: 1. The head is the base model's own; speculative decoding verifies every drafted token against the main model, so it changes throughput only, never the output distribution.

Enable it with vLLM:

vllm serve KIEFERSA/Sophea-Qwen3.6-v1.1 --served-model-name sophea-qwen3.6-v1.1 --trust-remote-code \
  --speculative-config '{"method":"qwen3_5_mtp","num_speculative_tokens":2}'

Vision

The full vision stack of the multimodal base is included, byte-identical to the base: image inputs work exactly as on Qwen3.6-35B-A3B. The adaptation touched only the language decoder; vision performance has not been separately re-evaluated for this release.

License

Apache-2.0. The base model, Qwen/Qwen3.6-35B-A3B, is released under Apache-2.0 and this fine-tune inherits those terms. Verify base-model terms before use.

Citation

This model is a release artifact of arXiv:2608.17744. If you use it, please cite:

@misc{kirouane2026thinking,
      title={Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See},
      author={Ayoub Kirouane and Christos Petrocheilos},
      year={2026},
      eprint={2608.17744},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2608.17744},
}

Task: Image-Text-to-Text
Library: transformers
Type: qwen3_5_moe
License: apache-2.0
Languages: el, en
Created: Sep 16, 2026
Updated: Sep 16, 2026