Eulerwa Inc. opens every layer of AI development — from model training to hardware acceleration to autonomous agents — through CLI and open source.
EulerForge for training, EulerStack for model design, EulerWeave for data, EulerAgent for execution, EulerNPU for hardware, and EulerAtlas for robot learning.
A research-oriented LLM fine-tuning framework that injects LoRA into HuggingFace models and lets you train a dense model as a Mixture-of-LoRAs or MoE Expert LoRA structure. On top of a familiar dense-SFT workflow, EulerForge adds Dense → MoE conversion and phase scheduling, so routing, expert specialization, and MoE stability can be studied reproducibly on a modest GPU budget. A single YAML preset carries you through SFT → DPO/ORPO → RM → PPO.
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
EulerStack is an Architecture Description Language for LLMs. It separates architecture out of PyTorch files where structure, training, and serving usually live tangled together — the same abstraction step the semiconductor industry took when Verilog and VHDL replaced schematics-plus-C. A single YAML spec flows through a 5-layer pipeline (DSL → Schema → IR → Compiler → CLI) and lands as a HuggingFace PreTrainedModel (config.json + model.safetensors) that EulerForge can pick up directly for training.
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).
A comprehensive data processing system that bridges the gap between raw datasets and production-ready model training.
Learn MoreDesigned for researchers and developers to refine and analyze data in local environments.
Business-critical features for scaling from a single GPU to cluster-level operations.
A local-first agentic framework with an 8-state machine, Pattern/Graph orchestrators, RAG, Long-Term Memory (SQLite), MCP integration, 30+ CLI commands, and a plugin system. Every action is logged, audited, and approved by humans. On top of deterministic gate nodes (kind: shell) that run verification as code (not the LLM), it provides a self-repairing coding agent (code.dev_loop.v2) that writes, tests, and fixes its own code.
Philosophy: Every action by an agent that changes the world (file writes, shell execution, external calls) must be approved by a human. The Pattern Orchestrator now mixes LLM, judge, and deterministic shell gate nodes (12 canonical patterns); bounded auto-approval (node.auto_approve) and a symbol gate (symbol_check) let it run fix→test loops autonomously.
30+ commands across core, approval, RAG, memory, workflow, pattern, and MCP groups. Built for developers who want autonomous capabilities without sacrificing control and security.
An NPU inference synthesis stack that composes an operator graph (spec.yaml) from 138 operators across 17 groups, then validates and compiles it (.npuart) to run on a host CPU reference, a functional NPU simulator, and Zynq-7000 (XC7Z020) · Zynq UltraScale+ (ZU3EG/ZU9EG) FPGAs. It provides INT4/INT8 quantization, 10 dtypes, 4 backends, and 15 CLI subcommands. See the product page for a demo of the humanoid-brain SoC (Project 3) being simulated.
Learn MoreSupports 138 operators (17 groups, A–Q) and 10 dtypes, compiling spec.yaml into .npuart artifacts.
| Operators | 138 (Core Math, Activation, Normalization, Conv/Vision, Sequence/Attention, Efficient Attention (Flash/Sliding/GQA), MoE/Sparse, Recurrent, Graph, Multimodal, Vision Encoder, Diffusion, Speculative Decoding, Quantization, Mamba/SSM, Cache Compress, Autonomy) |
|---|---|
| DType | fp32, int32, fp16, bf16, int8, uint8, int16, int4, fp8_e4m3, fp8_e5m2 — INT4 weight quantization supported |
| Target | Zynq-7000 (XC7Z020) · Zynq UltraScale+ (ZU3EG, ZU9EG) FPGA |
15 subcommands including validate, compile, run, sim, profile, explain, quantize, migrate-spec, generate, board smoke, calibrate, benchmark, replay, and compress-cache. Backends: cpu_ref · npu_sim · zynq_ps · zynq_pl_stub.
A plugin-based CLI robot behavioral learning framework running on an RL→IL pipeline that combines imitation learning (IL) with FastTD3 reinforcement learning (RL). The first two flagship domains we are releasing are car (EulerDrive — BEV autonomous driving, CARLA-verified) and humanoid (EulerWalk — RL→IL loco-manipulation), and both ship pretrained models so you can run them with no training at all. The unified schema is designed to cover 8 domains — car, drone, humanoid, robot dog, mobile manipulation, logistics, agriculture, shipyard — and domains keep expanding as plugins.
Learn MoreAll modules communicate consistently through fixed obs/action dimension schemas per domain.
| Classic (01-04) | Car 4/2 · Drone 6/3 · Humanoid 33/14 (loco-manip) · Robot Dog 12/8 |
|---|---|
| Industrial (05-08) | Mobile Manipulation 18/8 · Warehouse AGV 14/4 · Agri-Robot 16/6 · Shipyard 20/9 |
| Policy Models | BC-MLP, BC-RNN, BC-CNN, ACT, Diffusion Policy + SOTA (BEVFuser, TemporalTransformer, WholeBodyACT) + RL Actor |
| Levels | L0 Toy · L1 Intermediate · L2 Advanced (real backends: CARLA/MuJoCo/HumanoidBench) |
Adds train-rl, data collect/generate-demos, and sim rollout --render/--save-video. Includes the Domain Plugin architecture and edge deployment.
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.
Technology exists to serve humanity
We oppose the military automation of artificial intelligence. Technology must be used to protect life, not to take it — no technological achievement can take precedence over human safety.
Artificial intelligence must serve as a tool that elevates human dignity and strengthens democratic values. We oppose any use of AI that suppresses individual freedom or undermines democratic principles.
All software, models, and services released by Eulerwa may not be used for purposes that violate these principles. We take our responsibility for how our technology is used seriously.