We believe: the deeper we reach into the roots of mathematics and physics, the more openly we can share the results with the world.
Full-stack AI lab that turns deep research into CLI + open-source products covering LLM training · new architecture research · FPGA NPU synthesis · deterministic agent orchestration.
Public 3 + 1 coming: EulerForge (training) · EulerStack (architecture research) · EulerNPU (FPGA NPU · v0.1.0) · EulerAgent (deterministic CLI agent · coming soon).
No expensive cloud GPUs — turn the model you already have into an MoE.
EulerForge provides a standardized workflow for converting a dense model into an MoE-style trainable one, so experiments can be expressed in configuration rather than in glue code.
mixture_lora / moe_expert_lora| Injection | Dense LoRA · Mixture LoRA · MoE Expert LoRA · Native MoE Expert LoRA |
|---|---|
| Training | SFT · DPO · ORPO · RM · PPO |
| Backbones | Qwen2/3 · Llama 2/3 · Gemma 3 · Gemma 4 (dense+MoE) · Mixtral |
| Quantized training | nf4 / int4 / int8 (bitsandbytes) |
| Extras | HF Export · Optuna grid/bayesian search · integrated benchmark · 5-language CLI |
24 KO/EN tutorials + full CLI reference included
YAML-declarative LLM architecture language. Beyond an assembler, it's a research framework where cortical column · neural differentiation · tissue blocks can be ablated without code changes — letting you investigate hypotheses such as early emergence of reasoning · faster abstraction · independent development of world-knowledge vs reasoning/abstraction at the architecture level, and shifting frontier-LLM research from capital scale to design.
Organized by the v1 "industrial ordering principle", from validated industrial standards → recent hybrid/MoE → v1 experimental primitives: 24 llm_ (5 sizes × 4–5 variants, MLA included) + 33 arch_ (beginner 2 · intermediate 3 · advanced 5 · expert 23, of which 9 are *_mini). You can experiment with Phase B primitives — MLA, MoD, Titans, Dual-Stream — at arch-scale.
| Mixers | Attention · Mamba · RetNet · Hyena |
|---|---|
| FFN | MLP · Gated MLP (SwiGLU) · MoE (top-k routing) |
| Skill-level walkthrough | beginner (GPT-2/Llama) · intermediate (Mistral/Gemma2/Qwen) · advanced (Jamba/Samba/RetNet) · expert (MoE × mixer × depth/receptive-field 3D design space) |
| Compile target | HuggingFace model directory or JSON runtime config |
Three-stage validation (schema → cross-field compatibility → realism heuristics) catches design errors before compilation. All CLI messages are translated into 5 languages (ko/en/zh/ja/es). v0.1.5 adds μP scaling, differentiation auxiliary objectives, and the tissue organ declaration as backward-compatible spec extensions (existing YAML keeps working unchanged).
Reasoning ability emerges from less data. We directly test the hypothesis that structure becomes a function of data efficiency.
Abstraction and reasoning don't compete in the same layer. Inspired by cortical hierarchy, the two capabilities grow independently.
If world knowledge and reasoning/abstraction can be grown separately, the center of LLM research shifts from capital to design.
FPGA first. A single YAML spec runs inference on real FPGA silicon, and the same flow extends to ASIC as a long-term target.
DS-CNN + GRU INT8 full-graph NPU IP. 11/11 accuracy · 8.07× speedup vs CPU (3.17 ms/inference).
Hybrid configuration that offloads only the FFN block to the NPU. CPU↔NPU text 5/5 bit-identical — partial LLM inference demonstrated on a Zynq-7000-class FPGA.
spec → compile → bitstream → real FPGA — three NPUs validated end-to-end in one flow.
A predictable, local-first agent built on an 8-state machine plus HITL (human approval) gates. Runs open LLMs locally.
Learn moreAn 8-stage lifecycle plus HITL approval nodes. Execution flow is reproducible from logs, and risky stages require human approval before proceeding.
Runs open LLMs like Qwen and Llama locally. RAG, long-term memory (SQLite), and MCP integration ship as standard.
code.dev_loop.v2 — a code → test → self-fix loop running on top of the Pattern/Graph orchestrators and deterministic shell gates.
Preparing for release. Tutorials and CLI reference will be added at launch time.
A data engine and a robot behavioral-learning product family that run on top of the four products available today (Forge · Stack · NPU · Agent) are joining soon.
A data processing system that bridges the gap between raw datasets and production training. Local-LLM-based profiling, auto-generated domain filters, PPL/MinHash metrics, 1-GPU LoRA/MoE scale-down validation, GPU and cloud cost estimation, and MLOps integration.
An RL→IL pipeline that combines imitation learning (IL) with FastTD3 reinforcement learning (RL). First two public domains = car (EulerDrive, CARLA-verified) and humanoid (EulerWalk, RL→IL loco-manip). An 8-domain unified schema and Domain Plugin extensibility.
A full-stack AI lab researching and disseminating core AI technologies rooted in mathematics, physics, and humanities.
The name Eulerwa joins the mathematician Leonhard Euler with the Korean word "와 (wa)", meaning "with". The two halves of the name map directly onto the two halves of our motto.
Deep Roots — like Euler. Not chasing trends, but going to the foundations of mathematics and physics. Every Eulerwa product stands on that base.
Open Horizons — like 'wa (with)'. As Euler freely shared knowledge throughout his life, we keep our results open-source so everyone starts from equal opportunity. And the horizon those technologies head toward keeps widening.
We are building inference silicon with EulerNPU, intelligence structure with EulerStack and EulerForge, physical-behavior learning with EulerAtlas, and autonomous orchestration with EulerAgent. Each is an independent product, yet all point in a single direction — an intelligence that can understand and care for the physical world on its own.
The first destination is the ocean. Paired with power-supply barges, an autonomous composite that covers the surface, the underwater, and the air at the same time — collecting marine debris, mapping the spatiotemporal distribution of fish and generating fishing plans, monitoring navigation safety, and accumulating marine-ecology data. Surface vessels, submersibles, and aerial vehicles operating as one intelligent system, observing and caring for the ocean in three dimensions — this is the first physical application Eulerwa is preparing.
On the same technical base, our view extends further. Monitoring and protecting the near-Earth orbital environment is essentially the same problem as today's marine ecology monitoring. Voyaging deeper into space will demand a different level of autonomy and propulsion, but not giving up that direction is Eulerwa's long-term plan. The deeper the roots, the further the horizon.
Access specialized technical publications and the open-source ecosystem.
In-depth research books on quantum computing and AI architectures.
Community tools for data processing and model orchestration.
For the public good, so more people can share what we build.
AI should evolve so its benefits reach society as a whole, not just a handful of large technology companies. Eulerwa's research and products aim to contribute to public problems like industrial automation, safety, education, energy, and healthcare.
AI infrastructure and tools should not be concentrated in a few hands. From training and architecture research to the FPGA NPU inference stack, Eulerwa releases the full stack as open source so mid-size companies, startups, labs, and individual developers can reach the frontier in a measurable, reproducible form.
Outcomes are shaped by how technology is used. Eulerwa's tools can be applied across a wide range of industrial uses; for high-risk, safety-critical applications we make it an explicit principle that human judgment, transparent audit, and compliance with applicable law and regulation must accompany them.
Bold: currently public · Faded: in preparation
Deep in the roots of mathematics and physics. Open in sharing the results with the world.
Contact Us