# Welvet Β· OpenFluke AI engine > Welvet is OpenFluke's pure Go AI engine for Windows, Linux, macOS, Android, and iOS. Copy .entity models between devices; 34 numeric types, quantization, CPU/SIMD/WebGPU, zero CGO. Canonical: https://openfluke.com/welvet --- The engine Welvet Welvet is the brain of the whole project : an AI engine, built from scratch in Go, that can train and run neural networks on the hardware you already own. No cloud, no special chip, no black box. Windows Linux macOS Android iOS Read the feature book β†— Source on GitHub Run models with Octo In plain English What is an AI engine? An AI "model" is the trained thing that answers questions or makes pictures. An engine is the program that actually loads that model and runs it. Welvet is that program: a single, self-contained library you can drop into an app so it can think locally, the same way SQLite lets an app store data without a database server. Train once, then run the exact same model on Windows, Linux, macOS, Android, or iOS. πŸ“¦ Same model, every OS Runs on Windows, Linux, macOS, Android, and iOS. Copy a .entity model between them and it loads the same way, no re-export and no cloud. 🎯 Predictable results Same input, same output, every time, across phones, laptops, and desktops. That reliability is the whole point. 🧩 Runs real models Load popular open models (via Octo) and squeeze them down to fit on modest hardware, offline. The story so far From Loom to Welvet Welvet is the next chapter of our earlier engine, Loom . Most of the ideas carry straight over. This is mainly a cleaner structure: one feature per folder, honest error messages, and a strict rule that tests and demo apps live outside the engine. The headline change for users is that the model app, Octo, now runs several kinds of AI (chat, images, and speech), not just one. Old /loom links now point here. The engine's public home is github.com/openfluke/welvet . For developers Under the hood Pure Go, one feature per folder. Dense is the shared MatVec microkernel; storage-truth numeric types and quants; three backends: CPU tiled, Plan 9 SIMD, and WebGPU. Tests live in w2a , never in the engine. Layers: Dense, MHA, SwiGLU, RMSNorm/LayerNorm, CNN/RNN/LSTM, transformers. Numeric types & quants: 34 dtypes and 20 quant formats, packed on disk, not FP32-only. Backends: CPU tiled, Plan 9 SIMD (AVX2/NEON), and WebGPU in the browser. Model IO: .entity checkpoints, Hugging Face import, tokenizer, sampling. Systems: DNA, evolution, tween, TANHI telemetry, volumetric grid runtime. Honesty: no QAT, no silent fallbacks; half-done paths are stamped as such by w2a. Install go get github.com/openfluke/welvet@latest # then explore the runnable examples in the feature book v0.76 Β· scorecard β†’ v1.0 (76/100) Apache-friendly open source no silent fallbacks Deep dive every package with runnable go run . examples: the Welvet feature book β†— Architecture Dense is the shared MatVec microkernel Most transformer and CNN FLOPs are weights Γ— activations. Welvet keeps one Dense stack for that (FormatNone Γ— 34 dtypes + 20 quants Γ— CPU / SIMD / WebGPU). Layers whose expensive bit is GEMV reuse it. Kind Examples What runs where Native layer math Dense, RMSNorm, LayerNorm, Softmax, Embedding Own fwd/bwd; norms and softmax have real WebGPU shaders Composite β†’ Dense MHA, SwiGLU, RNN, LSTM, Residual, Sequential Projections are Dense children; attention / SiLU / recurrence stay local im2col β†’ Dense CNN1 / CNN2 / CNN3 Host im2col, then Dense GEMV (tiled conv shaders still open) One MatVec surface means one place for quant bugs, dtype wires, and backend parity. Separate native kernels pay off when the calc is not GEMV. Honest gaps today: k/IQ/Affine SIMD often inflate-once then DotTile (not true fused asm); MHA attention and some Softmax/SiLU paths still use host ALU; WebGPU device ALU is typically f32 at the boundary. w2a stamps these; nothing silent counts as β€œdone.” Layout Engine, proof, and apps Welvet is the engine only. Tests and the model shell live in sibling repos, never inside engine packages. 🧠 openfluke/welvet Engine Layers, quant, Plan 9 SIMD, WebGPU, ENTITY, dispatch. One feature per folder. GitHub β†— βœ… openfluke/w2a Proof Tests, CABI, docs, menus. The honesty harness that stamps every claim. w2a page β†’ πŸ™ openfluke/octo Model shell HF download, convert β†’ ENTITY, quantize, run. The Lucy successor. Octo page β†’ Folder Contains (top) core , weights , quant , simd , webgpu , tiling , architecture , layers/ runtime/ forward, backward, training, step systems/ dna, evolution, tween, tanhi, telemetry model/ transformer, entity, tokenizer, sampling, hf apps/ octo, flux2, mosstts stub/ seed, serialization, hardware, memory, fountain, donate, accel loom/poly is legacy reference only. Welvet is the rewrite. Non-negotiable Rules the engine lives by No testing code in the engine β€” all checks live in w2a. No fallbacks β€” missing path β†’ hard error (no SIMDβ†’Go, no fake GPU). Nothing hardcoded to float32 β€” APIs are Tensor[T] / generics. No QAT β€” DType + QuantFormat are storage truth. One feature β†’ one folder. v1.0 = scorecard 100/100 β€” every board row βœ…. Coverage axes What β€œdone” means For each layer or op, every cell must work. No cell may silently substitute another cell. 3 backends 34 dtypes 20 quant formats 2 fwd + bwd Backend Status Requirement CPU tiled βœ… SC+MC; stream native + packed MatVec Plan 9 SIMD βœ… amd64 AVX2+FMA / arm64 NEON; unsupported arch β†’ hard error WebGPU βœ… Real device; FormatNone + quant GEMV; no host fake-GPU Version scorecard v0.76 toward v1.0 Formula: version = 0.{round(earned)} until 100 β†’ v1.0 . Today: 76 / 100 . Only this board sets the version number. βœ… Implemented Β· 100% of row 🚧 Partial Β· 50% of row ⬜ Not started Β· 0% # Section Wt Earned 1 Foundation β€” layout, core, weights, quant, simd, webgpu, tiling 15 15.0 2 Dense MatVec microkernel β€” FormatNoneΓ—34 + quants Γ— backends 15 13.5 3 Transformer stack β€” MHA, SwiGLU, norms, Embedding, Sequential… 14 14.0 4 CNN / RNN / LSTM β€” full timed 34Γ—20Γ—3 matrices 6 6.0 5 Extended layers β€” GDN, ConvT, Mamba, KMeans, Parallel… 7 3.5 6 Runtime + architecture β€” volumetric grid, fwd/bwd/train/step 8 8.0 7 Systems β€” dna, evolution, tween, tanhi, telemetry 5 5.0 8 Model / IO β€” tokenizer, entity, transformer, sampling, hf 8 8.0 9 Apps β€” Octo model shell 3 1.5 10 Stubs β€” seed, serialization, hardware, memory… 3 1.5 11 Accel β€” NPU / Metal / QNN plugins 2 0.0 12 Peak fused / no host ALU 14 0.0 Total β†’ v1.0 100 76.0 Biggest remaining chunks: peak fused (14), extended layers (~3.5), then apps / stubs / accel. Full dtype and quant boards live in the README on GitHub β†— . Honest snapshot What ships today Area Status Engine layout (one feature β†’ one folder) βœ… Rules: no engine tests / no fallbacks / no hardcoded f32 / no QAT βœ… core β€” 34 dtypes, Tensor[T], activations, backends βœ… weights β€” FormatNone Γ— 34 stream pack / MatVec βœ… quant β€” Pack / Unpack / MatVec all 20 formats (CPU) βœ… simd β€” Plan 9 kernels (amd64 / arm64) βœ… webgpu β€” Dense GEMV family + norm / softmax / SwiGLU fuse βœ… Dense FormatNone Γ— 34 Γ— 3 backends fwd+bwd βœ… Dense k/IQ/Affine SIMD (inflate + DotTile, not fused) 🚧 architecture β€” volumetric grid βœ… Model IO / transformer / entity / tokenizer / hf βœ… apps/octo β€” interactive model shell 🚧 stub/accel β€” NPU / Metal / QNN ⬜ Full v1 matrix (every cell peak-fused) ⬜ Feature book β†— See the proof run Full README on GitHub