AI Potluck
Contribute
Back to Gap Map Organization

SpeakLeash Foundation

foundation · Poland

Scores

1 product on the map — 1 open-ish.

Bielik

Openness

3 high confidence
3.0
weights
open(safetensors on HF speakleash/Bielik-11B-v3.0-Instruct, behind an automatic click-through that asks for contact consent)
post-training-data
described(SpeakLeash's own instruction corpus (20M instructions and dialogues) and a 114,000-example preference set, described in the paper, not released)
code
closed(no training code released
license
Apache-2.0(OSI
every Bielik v3 checkpoint on the Hub)+Bielik-Terms-of-Use-1.1(non-OSI conduct terms on top of Apache-2.0
users must be 18 or older, and unlawful or harmful use and processing personal data are prohibited

Bielik 11B v3 ships under Apache-2.0, an open-source license, after an automatic click-through on Hugging Face. SpeakLeash's terms of use add conduct rules, an age minimum of 18 and a ban on processing personal data among them, without limiting commercial use. The instruction and preference data are SpeakLeash's own, described in the paper but not released, and no training code is published.

  • https://arxiv.org/pdf/2601.11579 recorded 2026-10-13

    Bielik 11B v3 paper, section 4.1: "we developed a proprietary dataset that undergoes continuous expansion and refinement by human annotators"; SFT ran "for 3 epochs on 20 million instructions and dialogues"; the preference set was expanded "to over 114, 000 examples"; RL "was performed using the Volcano Engine Reinforcement Learning (VERL) framework". The paper links no dataset release and no training code.

  • https://bielik.ai/en/terms/ recorded 2026-10-13

    Terms of Use, "version 1.1 of 31 July 2025", supplementary to the license: "The Bielik 3.0 Model is made available under the Apache 2.0 licence". "Use of the Model is permitted only to persons who are at least 18 years of age", and use for "unlawful activities", harmful content and "processing any personal data" is prohibited. It sets no commercial, user or revenue limit.

  • https://huggingface.co/api/models?author=speakleash&limit=200 recorded 2026-10-13

    SpeakLeash's model listing: every Bielik v3 repository, from Bielik-1.5B-v3 through Bielik-11B-v3.0-Instruct and Bielik-Minitron-7B-v3.0-Instruct and their quantized copies, carries the tag license:apache-2.0.

  • https://huggingface.co/api/models/speakleash/Bielik-11B-v3.0-Instruct recorded 2026-10-13

    Hub record: "private": false, "gated": "auto", cardData license "apache-2.0", base_model speakleash/Bielik-11B-v3-Base-20250730, extra_gated_fields asking only "I agree to be contacted for feedback about Bielik models"; files model-00001-of-00005.safetensors to model-00005-of-00005.safetensors. No dataset tag. The raw model card returns 401 without accepting the gate.

Adoption

1 high confidence
1.0

Adoption is measured by Hugging Face downloads of the full-precision instruct model. SpeakLeash also publishes quantized AWQ and GGUF copies, which are not declared and so are not counted here.

Capability

3 medium confidence
3.0

Bielik is not listed on Artificial Analysis, so this reading comes from the paper's own English benchmark table. There it clearly outperforms Mistral 7B Instruct v0.2, a tune of the same base it was scaled up from, most of all in math, and Llama 3 8B Instruct. In Polish it scores close to open models several times its size.

  • https://artificialanalysis.ai/leaderboards/models recorded 2026-10-13

    The leaderboard's embedded model records give intelligenceIndex 14.19 for Mistral Medium 3.5, 12.83 for Nemotron 3 Super 120B A12B (Reasoning), 11.27 for Mistral Small 4 (Reasoning), 8.99 for Mistral Small 4 (Non-reasoning) and 8.44 for GPT-4o (Nov '24). No SEA-LION or Bielik model appears on the page. The digest drifts between fetches because the page carries live pricing and throughput figures.

  • https://arxiv.org/abs/2601.11579 recorded 2026-10-13

    "Bielik 11B v3: Multilingual Large Language Model for European Languages", submitted 30 Dec 2025. The model "extends the Mistral 7B v0.2 architecture, scaled to 11B parameters via depth up-scaling", trained in four stages: "continuous pre-training, supervised fine-tuning (SFT), Direct Preference Optimization (DPO), and reinforcement learning".

  • https://arxiv.org/pdf/2601.11579 recorded 2026-10-13

    Table 11, "Open LLM Leaderboard results for selected instruction-tuned models": Bielik-11B-v3.0-Instruct AVG 72.45, MMLU 71.11, GSM8K 85.60; Meta-Llama-3-8B-Instruct 66.87; Mistral-7B-Instruct-v0.2 65.71, MMLU 60.78, GSM8K 40.03. Table 5, Open PL LLM Leaderboard for instruction-tuned models: Bielik-11B-v3.0-Instruct 65.93, below Llama-3.3-70B-Instruct at 66.40 and above Qwen2-72B-Instruct at 65.87.