What Speakleash does

Speakleash (also known as Spichlerz) is a Polish non-profit open-science foundation founded in 2022 and based in Warsaw. Its mission is to build the largest curated Polish-language text corpus and to train and release Polish Large Language Models so that Polish-language AI development does not depend on foreign models. The foundation operates a community of more than 160 contributors drawn from Polish enterprises (Deviniti, Bank Pekao, BTC), research institutions (AGH Cyfronet, CLARIN-PL, NASK PIB), and AI students, with collaboration coordinated through Discord, GitHub, and HuggingFace.

The core product surface comprises the Spichlerz dataset (exceeding 1 TB of Polish text and AI-Act compliant), the Bielik family of generative pre-trained transformer models (Bielik-7B, Bielik-2 11B, and the Bielik-PL v3 series with a dedicated Polish-optimized tokenizer), evaluation platforms (Open PL LLM Leaderboard, MT Bench PL, EQ-Bench PL, Chat Arena PL), and the Sówjka guardrails model for content safety. The Obywatel Bielik citizen-science project extends the effort into multimodal data, targeting 1 million cultural images. Compute is provided by Cyfronet AGH's 35-petaflop supercomputers (Athena and Helios with NVIDIA GH200 accelerators), and compression techniques for Bielik-PL-Minitron-7B were co-developed with NVIDIA engineers.

Speakleash operates on a freemium, open-source model with no disclosed pricing tiers or revenue streams. Distribution is entirely self-serve via HuggingFace and GitHub; sponsorships (Deviniti) and in-kind contributions from Intel, NVIDIA, and AGH Cyfronet substitute for cash funding. Named institutional counterparties include the Polish Ministry of Defence (which adopted Bielik for the national AI implementation center) and Cyfronet AGH as the primary compute partner.

Name

Speakleash

Legal name

Speakleash.org

Website

https://speakleash.org

Company type

Private

Founded year

2022

Operating status

Operating

Headcount range

11–50 employees

Short description

Speakleash is a Polish open-science foundation building the largest curated Polish text dataset (Spichlerz) and the Bielik family of Polish-language foundation models, serving researchers, developers, and the Polish government with free open-source LLMs and benchmarks.

Ownership category

akta.pro rank

Product category

Open-Source Language Models

NAICS

Web Search Portals and All Other Information Services (519290)

SIC

Services-Prepackaged Software (7372)

akta.pro primary industry

Question Answering & Retrieval-Augmented Generation (RAG) (HDAAADAC)

Keywords

• Polish language models
• Open-source LLM
• Text dataset curation
• AI guardrails
• Generative AI community

Headquarters

HQ city

Warsaw

HQ country

Poland

HQ region

Europe

Offices1 record

Markets served

GTM type

B2B

Offering type

Software

Cost components

Technology or R&D, Personnel, Infrastructure, Operations, Marketing or Sales

Revenue model

1. Free/Open-Source Services: All services provided free of charge. The project operates as an open-science initiative with volunteer contributors and no commercial revenue model.

Pricing tiers

Go-to-market motion2 records

Distribution channels2 records

Marketing channels7 records

Core offering

Speakleash (aka Spichlerz) is a non-profit open-source initiative that curates a 1TB+ Polish text dataset and develops the Bielik family of large language models optimized for Polish, along with supporting tools such as the Sówjka guardrails model and the Chat Arena PL comparison platform. All artifacts are released free of charge under an open-science model.

Product overview

SpeakLeash (also known as Spichlerz) is an open-science initiative that operates as a platform-plus-products ecosystem for Polish language AI. The core offering is the Spichlerz dataset—a 1TB+ Polish text corpus that is the largest and most documented Polish language dataset—which powers the Bielik family of large language models (7B, 11B, and Minitron variants). Supporting products include CHAT ARENA PL (LLM comparison platform), evaluation benchmarks (MT BENCH PL, EQ-Bench PL), the Sójka guardrails model for content safety, and the Obywatel Bielik citizen science project for multimodal data. The PLLuM initiative represents broader collaboration on Polish LLM development. All products are oriented around advancing Polish language AI capabilities in an open-science manner.

Differentiator

Problem solved

Functional benefit

Brands

• Bielik AI: Polish language AI model family developed by SpeakLeash, including Bielik-7B, Bielik-11B, and Bielik-PL variants with Polish-optimized tokenizer.
• CHAT ARENA PL
• Sójka

Products and services

• Bielik-PL-11B-v3.0-Instruct An 11 billion parameter instruction-tuned large language model for Polish, with a dedicated Polish-optimized tokenizer; achieves 65.93 on the Open LLM Leaderboard 5-shot average, outperforming larger models such as Meta-Llama-3.1-70B-Instruct. Intended for AI developers and researchers building Polish-language applications.
• Bielik-PL-Minitron-7B-v3.0-Instruct A compressed 7B-parameter Polish LLM derived from the 11B model via structured pruning and knowledge distillation, delivering 33% smaller model size and up to 50% faster inference while retaining approximately 90% of the full model's quality. Targets developers needing efficient Polish language inference.
• Bielik-7B-v0.1 The first Bielik large language model with 7 billion parameters, trained on Polish data from the Spichlerz dataset; available in base and instruct variants for Polish language NLP tasks.
• Sówjka Guardrails Model An AI guardrails model designed to detect and filter harmful content including hate speech, vulgarities, and violence, used for comment moderation, chat analysis, and filtering AI responses to maintain content safety in Polish-language applications.
• Chat Arena PL An interactive LLM comparison platform where users can compare Polish-language models (including Llama3, Mixtral, Bielik, and GPT-3.5) and models are ranked using an ELO rating system based on user evaluations.
• Open PL LLM Leaderboard A leaderboard for ranking Polish language LLMs, allowing comparison of model performance on Polish language tasks.
• MT Bench PL & EQ-Bench PL Polish language evaluation benchmarks for measuring LLM performance on various language understanding and generation tasks, including EQ-Bench PL at https://huggingface.co/spaces/speakleash/polish_eq-bench.
• Spichlerz Dataset The Polish language text dataset that has collected over 1 TB of Polish text; the largest and most documented Polish language dataset, prepared according to AI ACT requirements, with a live dashboard for ongoing data collection.

Quantifiable outcome

• Bielik-11B-v3.0-Instruct achieves 5-shot average of 65.93 on Open LLM Leaderboard, outperforming Meta-Llama-3.1-70B-Instruct
• +2 more outcomes

Named customers2 records

Segments3 records

Ideal customer profiles3 records

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

AI capability3 records

Feature5 records

Partnerships

Nine partnerships are on record, tiered core and minor.

• IntelcoreTechnology or Integration · 19 February 2026Intel supported the development of Bielik AI model. Poland announced use of Dell equipment for AI implementation center for Polish-language AI model called Bielik, developed by SpeakLeash with Intel's support.
• DevinitiminorOthersPartner and sponsor. Listed in partners and sponsors section of the website. Provides organizational support.
• Cyfronet AGH (Academic Computer Centre)coreStrategic or Co-development PartnerStrategic partnership with Akademickie Centrum Komputerowe Cyfronet AGH. First commercial customer of latest NVIDIA GH200 accelerators (35 petaflops). Supercomputers Athena and Helios used for training Bielik models.
• NVIDIAcoreTechnology or IntegrationNVIDIA engineers collaborated on structured pruning and knowledge distillation techniques for Bielik-PL-Minitron-7B compression. Compression methodology first introduced at NVIDIA GTC March 2026.
• CLARIN-PLcoreStrategic or Co-development PartnerAcademic collaboration with CLARIN-PL (Common Language Resources and Technology Infrastructure). Participated in CLARIN-PL 10th anniversary conference.
• NASK PIBcoreStrategic or Co-development PartnerAcademic research collaboration with NASK PIB (Research and Academic Computer Network).
• PLLuM (Polish Large Language Universal Model)coreStrategic or Co-development PartnerCollaborative initiative with key Polish AI units on Polish Large Language Universal Model. SpeakLeash data contributes to PLLuM development.
• Hack To The RescueminorGTM or Marketing PartnerOfficial partner of hackathon. Mentors Maria Filipkowska and Adrian Gwoździej participated. World's largest Generative AI event.
• ML in PLminorGTM or Marketing PartnerEvent partnership with ML in PL conference.

Scale indicators5 records

Recent moves7 records

Expansion highlights6 records

Direct peers

• PLLuM (Polish Large Language Universal Model): Polish government-backed initiative to build a universal Polish LLM, with Speakleash contributing data. Directly comparable as a parallel Polish-language LLM effort sharing data and research alignment.
• EleutherAI: Volunteer-driven open-source AI research collective that released the GPT-Neo and Pythia model families. Closest peer in terms of community-led, open-science LLM development without a commercial product.
• Allen Institute for AI (AI2): Non-profit AI research institute producing open-source models (e.g., OLMo, Tulu). Directly comparable as a non-profit foundation releasing open LLMs and datasets with research-grade rigor.

Emerging players

• Stability AI: Open-source generative AI company (Stable Diffusion, Stable LM). Comparable as an open-source generative AI player with a community-driven distribution model, though focused on multimodal rather than Polish-specific language.
• Together AI: Open-source AI cloud platform supporting open model training and inference. Comparable as an infrastructure provider for open-source LLM development, though operating commercially rather than as a foundation.
• Hugging Face: Open-source ML platform hosting models, datasets, and evaluation leaderboards. Comparable as the distribution and community infrastructure that Speakleash uses to publish Bielik models and the Spichlerz dataset.

Broad incumbents

• Cohere: Enterprise-focused LLM provider with multilingual capabilities. Comparable as a foundation model developer, but operates commercially and serves enterprise customers rather than a non-profit open-science mission.
• Mistral AI: European-based LLM developer with open-weight models. Comparable on European language AI focus and an open-weight distribution model, but operates as a commercial entity rather than a foundation.
• OpenAI: Creator of GPT-family LLMs with multilingual support including Polish. Comparable as the dominant commercial LLM developer whose models serve as the benchmark and competitor for Bielik performance claims.
• Meta AI (Llama): Releases open-weight foundation models (Llama 3.x) that Serve as the architectural base for Bielik. Comparable as the upstream open-source LLM ecosystem that defines the competitive baseline for Polish-language models.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat5 records

Key risks6 records

Key highlights7 records

Customer concentration

• Xx.com/Speak_Leash
• GitHubgithub.com/speakleash
• Facebookfacebook.com/speakleash
• LinkedInlinkedin.com/company/speakleash

Revenue estimate

Valuation estimate

Number of profiles

Profiles14 records

Funding overview

Funding rounds

Investors

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Speakleash