# DeepSeek Projects at AI Tinkerers

> Canonical HTML: https://aitinkerers.org/technologies/deepseek
> Markdown URL: https://aitinkerers.org/technologies/deepseek.md
> Technology record last updated: 2026-02-26T12:16:41Z
> Generated: 2026-09-22T08:38:48Z

DeepSeek is a Chinese AI research firm: they develop high-performance, cost-efficient, open-source Large Language Models (LLMs) that disrupt industry cost benchmarks.

DeepSeek, founded in July 2023 by Liang Wenfeng (CEO and co-founder of High-Flyer), specializes in developing state-of-the-art, open-source LLMs from its Hangzhou, China headquarters. The company gained global recognition for its cost-effective approach: its flagship DeepSeek-R1 model, for example, was reportedly trained for a fraction of the cost of rivals like OpenAI's GPT-4. DeepSeek's product portfolio includes the reasoning-focused DeepSeek-R1, the multi-purpose DeepSeek-V3 (a 671B-parameter Mixture-of-Experts model), and the specialized DeepSeek-Coder series. This innovation in training efficiency has positioned DeepSeek as a significant, disruptive competitor in the global AI landscape.

- Official technology site: https://deepseek.com
- Public AI Tinkerers demos and talks: 12
- Result page: 1 of 1

## Recent Public Talks and Demos

### [Construindo Linr, um editor interativo de timelines com Deepseek rodando no Cursor](https://curitiba.aitinkerers.org/talks/rsvp_xqVhn6TV5MI)

O Linr é um editor online de timelines visuais - útil para checkpoints, milestones, diagramas, eventos historicos, registros de viagem, funil de vendas e afins. Configurei os modelos DeepSeek V4 dentro do Cursor usando o Cline como harness, porque o agente nativo do Cursor não aceita rotear para a API da DeepSeek. Com esse setup rodando, o Linr saiu do zero. Na demo ao vivo eu mostro, em três atos: O setup : por que o caminho obvio não funciona (o agente do Cursor recusa a chave externa) e como o Cline resolve isso dentro da mesma IDE. O coração: pego o Linr num estado inacabado e dou uma tarefa real ao agente rodando DeepSeek ao vivo - uma feature que falta ou um bug - e vocês veem o modelo construindo software de verdade, com o custo em dolar subindo na tela. O payoff: abro o app funcionando e mostro quanto custou, em tokens e em dólar, chegar ali.

- Event context: AI Tinkerers Curitiba: Encontro de Agosto (no EBANX) — 2026-08-26 — Curitiba
- Public talk page: https://curitiba.aitinkerers.org/talks/rsvp_xqVhn6TV5MI

### [inteligencia Artificial y creatividad: Dos fronteras cada vez mas juntas](https://pereira.aitinkerers.org/talks/rsvp_z2VSRgVfmfk)

Una herramienta para la escritura de libros, ebooks, biografias impulsada con inteligencia artificial.

- Event context: AI Tinkerers Pereira: Innovando con IA desde el SENA — 2026-04-30 — Pereira
- Public talk page: https://pereira.aitinkerers.org/talks/rsvp_z2VSRgVfmfk

### [AI safety in healthcare](https://vienna.aitinkerers.org/talks/rsvp_jm2eqpK8Vtk)

AI by itself is currently not good enough to be used without extensive supervision in the professional/production environment of a lot of industry, especially law enforcement or healthcare. Because of that, I would talk about how I design a more safe and deterministic AI system for industries that require a high level of complexity, accuracy, and certainty. I would walk through one of the project related to my award-wining research at AI in health conference of Ken Kennedy Institute of Rice University where I evaluate AI system for safety, before implementing solutions to mitigate those limitations. This include closely evaluating accuracy and saftey of AI outputs with feedback loop and "AI-as-a-judge", LLM jailbreaking, and careful prompt engineering

- Event context: AI Tinkerers Vienna: 2026 — 2026-02-19 — Vienna
- Public talk page: https://vienna.aitinkerers.org/talks/rsvp_jm2eqpK8Vtk

### [Assembly of Experts: Linear-time construction of ultra-large language model variants](https://munich.aitinkerers.org/talks/rsvp_5nHjZ4TDvtA)

Requiring 10^13-10^15 FLOPs to calculate one 8 bit weight in an LLM is inefficient and not beautiful. Therefore, we developed the new "Assembly-of-Experts" construction method. It allows to create meaningful child variants of existing Mixture-of-Experts parent models in linear time. For this, parts of the parent models are recombined, guided by structural and semantic features. Using AoE, we constructed the DeepSeek R1T "Chimera", a 671B open-weights hybrid model combining DeepSeek's V3-0324 and R1 model variants. The child is achieving about R1-level intelligence, while using about 40% fewer output tokens, close to V3. Constructed without any fine-tuning or distillation, the Chimera exhibits surprisingly compact, orderly reasoning compared to its parent models. Moreover, it showed behaviour outside of the distribution, i.e. it is not just an interpolation of the parent models. Community response has been encouraging. The model has 8,200 downloads on Hugging Face and, at peak at the end of May, reached 1.35 billion processed tokens per day on OpenRouter and almost 5 billion tokens per day on chutes.ai. Since then, we have developed the second generation, DeepSeek-TNG R1T2 Chimera. It is a Tri-Mind that also leverages the new R1-0528, and uses a more refined construction method. It appears to be 20% faster than R1, while being almost always more intelligent. It is not quite as intelligent as R1-0528, but typically more than twices as fast, and also with a beautiful chain of thought like R1T. Since R1T2s release on OpenRouter on July 7th, it reached 1 billion tokens per day already on its first full day of operation. It broke 2 B/d on July 17th and grew to become the 5th most popular LLM on Chutes on July 19th. It currently runs on 30 Chutes instances of 8xH200 each, totalling 240 H200 GPUs.

- Event context: AI Tinkerers Munich - July 25 — 2025-07-25 — Munich
- Public talk page: https://munich.aitinkerers.org/talks/rsvp_5nHjZ4TDvtA

### ["Sports Videos Understanding at Scale" OR "Video Runner Bib OCR detection"](https://montreal.aitinkerers.org/talks/rsvp_IccvgKkVEhA)

there are two things I worked on in Q1 and 2024Q4 that might be interesting: 1. Using VLM to create a Visual Q&A using RAG and also using it for "guided" highlight detection 2. I also tried using the large context window offered in Gemini 2.0 Flash Thinking to read marathon runners' bib

- Event context: AI Tinkerers Montreal - May 2025 Meetup at LightSpeed — 2025-05-07 — Montreal
- Public talk page: https://montreal.aitinkerers.org/talks/rsvp_IccvgKkVEhA

### [AI Architect](https://hong-kong.aitinkerers.org/talks/rsvp_xdYfH6khNPM)

In this session, we will together explore how artificial intelligence is reshaping the world of architecture design, making it more accessible, efficient, and personalized than ever before. We will learn how to leverage AI to streamline design process, enhance creativity, and deliver tailored solutions for clients. Whether you’re a professional architect, a tech enthusiast, or an AI advocate, this session will provide actionable insights and inspirations to elevate your design approach and embrace the future of architecture.

- Event context: AI Tinkerers - Hong Kong Meetup (March) - Art Month & International Women's Day! — 2025-03-20 — Hong Kong
- Public talk page: https://hong-kong.aitinkerers.org/talks/rsvp_xdYfH6khNPM

### [Creating Browser agent based on screen recording or video.](https://atlanta.aitinkerers.org/talks/rsvp_H8XjHnsScjY)

Demonstrate how businesses can create browser agents to automate manual & repetitive processes. Companies can describe their processes in natural language, screen recording or user manual and instantly get automations for the tasks. For example, automating the transfer of data from medical EHR systems or place purchase orders, or just simply updating the most used software’s in the world: Excel, Word and PowerPoint.

- Event context: AI Tinkerers Atlanta February Meetup — 2025-02-27 — Atlanta
- Public talk page: https://atlanta.aitinkerers.org/talks/rsvp_H8XjHnsScjY

### [Sofia: Your AI, Your Data, Your Rules](https://manizales.aitinkerers.org/talks/rsvp_jFw3i8kiQtA)

Sofia is a private, OSINT-powered AI assistant designed to run locally, giving users full control over their data. Built with a modular, containerized architecture using Docker, it ensures seamless deployment and portability. Developed in Python, Sofia leverages Ollama to run DeepSeek as a local AI model, eliminating cloud dependencies. Secure, adaptable, and customizable, it enhances daily workflows while prioritizing privacy and efficiency

- Event context: Primera reunion AITinkerers Manizales — 2025-02-26 — Manizales
- Public talk page: https://manizales.aitinkerers.org/talks/rsvp_jFw3i8kiQtA

### [How to evaluate LLMs?](https://amsterdam.aitinkerers.org/talks/rsvp_VC8O-Ww8Q8Y)

I would give a live demo on how to evaluate LLM output using evaluators like 'LLM as a judge', functions like Cosine similarity, and JSON schema validators.

- Event context: AI Tinkerers Amsterdam - February Edition — 2025-02-26 — Amsterdam
- Public talk page: https://amsterdam.aitinkerers.org/talks/rsvp_VC8O-Ww8Q8Y

### [Agent Laboratory: Agentic AI Researcher](https://prague.aitinkerers.org/talks/rsvp_2skAj-j5HIE)

Agent Laboratory is an end-to-end autonomous research workflow meant to assist you as the human researcher toward implementing your research ideas. Agent Laboratory consists of specialized agents driven by large language models to support you through the entire research workflow—from conducting literature reviews and formulating plans to executing experiments and writing comprehensive reports.

- Event context: AI Tinkerers Prague - February 2025 Meetup — 2025-02-25 — Prague
- Public talk page: https://prague.aitinkerers.org/talks/rsvp_2skAj-j5HIE

### [LLMs can teach themselves to better predict the future](https://seattle.aitinkerers.org/talks/rsvp_zkiKbBlvnuQ)

In our first paper from Lightning Rod Labs (https://lightningrod.ai), we explore if AI can improve its forecasts via self-play and real-world outcomes: - 12,100 questions and outcomes from Polymarket (politics, sports, crypto, science, etc) - Base model generates multiple distinct reasoning traces and predictions per question - Rank predictions by how close they were to the actual outcome - Fine-tune with DPO on the ranked traces & predictions Result: +7-10% accuracy over control, bringing two small (14B) models on par with GPT-4o (over 10x larger). Arxiv Link: https://arxiv.org/abs/2502.05253

- Event context: AI Tinkerers Seattle - February 2025 Meetup — 2025-02-22 — Seattle
- Public talk page: https://seattle.aitinkerers.org/talks/rsvp_zkiKbBlvnuQ

### [Drop-in OpenAI Assistants Replacement](https://boston.aitinkerers.org/talks/rsvp_Pjqe613g4Z0)

I was an early user of the OpenAI Assistants API and have experienced most of its rather significant shortcomings. I decided to write a drop-in replacement that not only fixes the shortcomings, but opens up a myriad of additional capabilities. I can continue using the OpenAI SDKs, but now use non-OpenAI LLMs and embedding models. So, for instance, I can create vector embeddings that have specialized multi-language support coupled with ultra-low-cost LLMs such as Amazon, DeepSeek, Llama, etc.

- Event context: AI Tinkerers Boston Meetup January 2025 — 2025-01-27 — Boston
- Public talk page: https://boston.aitinkerers.org/talks/rsvp_Pjqe613g4Z0

## Related Technologies

- [Claude](https://aitinkerers.org/technologies/claude) ([Markdown](https://aitinkerers.org/technologies/claude.md)) — 174 public demos
- [Gemini](https://aitinkerers.org/technologies/gemini) ([Markdown](https://aitinkerers.org/technologies/gemini.md)) — 188 public demos
- [OpenAI API](https://aitinkerers.org/technologies/openai-api) ([Markdown](https://aitinkerers.org/technologies/openai-api.md)) — 520 public demos
- [Python](https://aitinkerers.org/technologies/python) ([Markdown](https://aitinkerers.org/technologies/python.md)) — 662 public demos
- [GPT-4o](https://aitinkerers.org/technologies/gpt-4o) ([Markdown](https://aitinkerers.org/technologies/gpt-4o.md)) — 57 public demos
- [Amazon](https://aitinkerers.org/technologies/amazon) ([Markdown](https://aitinkerers.org/technologies/amazon.md)) — 2 public demos
- [AMD MI325X](https://aitinkerers.org/technologies/amd-mi325x) ([Markdown](https://aitinkerers.org/technologies/amd-mi325x.md)) — 1 public demo
- [Browser agent](https://aitinkerers.org/technologies/browser-agent) ([Markdown](https://aitinkerers.org/technologies/browser-agent.md)) — 1 public demo
- [Cline](https://aitinkerers.org/technologies/cline) ([Markdown](https://aitinkerers.org/technologies/cline.md)) — 1 public demo
- [Cursor](https://aitinkerers.org/technologies/cursor) ([Markdown](https://aitinkerers.org/technologies/cursor.md)) — 65 public demos
- [DeepSeek API](https://aitinkerers.org/technologies/deepseek-api) ([Markdown](https://aitinkerers.org/technologies/deepseek-api.md)) — 5 public demos
- [deepseek-chat](https://aitinkerers.org/technologies/deepseek-chat) ([Markdown](https://aitinkerers.org/technologies/deepseek-chat.md)) — 1 public demo
- [DeepSeek V4 Flash](https://aitinkerers.org/technologies/deepseek-v4-flash) ([Markdown](https://aitinkerers.org/technologies/deepseek-v4-flash.md)) — 1 public demo
- [DeepSeek V4 Pro](https://aitinkerers.org/technologies/deepseek-v4-pro) ([Markdown](https://aitinkerers.org/technologies/deepseek-v4-pro.md)) — 1 public demo
- [Docker](https://aitinkerers.org/technologies/docker) ([Markdown](https://aitinkerers.org/technologies/docker.md)) — 147 public demos
- [Flash](https://aitinkerers.org/technologies/flash) ([Markdown](https://aitinkerers.org/technologies/flash.md)) — 14 public demos
- [Gemini-2](https://aitinkerers.org/technologies/gemini-2) ([Markdown](https://aitinkerers.org/technologies/gemini-2.md)) — 15 public demos
- [IDE](https://aitinkerers.org/technologies/ide) ([Markdown](https://aitinkerers.org/technologies/ide.md)) — 4 public demos