CacheWarden
Keeps assistant prompt caches warm while you step away, so you don't pay to rebuild them.
Flagship · VS Code
Forge
An open-source coding agent for VS Code, built for people who run their own models.
Forge loads GGUF models through llama.cpp and manages them properly: start, keep loaded, share across editor windows. Cloud models and the Claude Code and Codex CLI agents are there when you choose to add them.
- Local runtime, managed. Model loading, GPU placement and context limits are handled for you.
- Tool calls that hold up. Tuned for the context and output limits that real local models have.
- Every change reviewable. Inline diffs, per-turn checkpoints and Keep/Undo for each file.
- Long sessions. Context compaction keeps work going past the window limit.
- Plays well with others. MCP servers, delegation to CLI agents, and remote control from your phone.
Flagship · Desktop
HalluScribe
A private, searchable memory of every AI coding session you have had.
HalluScribe reads the session files your AI tools already write, summarises each one with a local model, and keeps a readable Markdown archive on your disk. Nothing is uploaded and there is no API bill.
- Many sources. Claude Code, Codex and Forge sessions, plus exported ChatGPT, Claude.ai, Gemini and Grok chats.
- Local summaries. Gemma 4 by default, turning 200K-token logs into short, structured entries.
- Runs on its own. A nightly sweep archives new sessions automatically.
- Memory for agents. A read-only MCP server lets coding agents search what earlier sessions learned.
Flagship · Desktop · Offline
Gemma4kids
An offline learning companion for children aged 6–11, in their own language.
A child can ask anything, from fractions to what to do in an earthquake, and get a plain, age-appropriate answer. When they want to make something, Gemma writes a live animation they can watch, read and change. It runs entirely on the family computer with Google's Gemma 4.
- No internet, no account. Models run locally through Ollama or llama.cpp; nothing leaves the machine.
- Learning by making. Prompts become HTML animations in an editor children can open, edit and save.
- Their language. English, German and Greek, with voice input and read-aloud through a local Piper voice.
- Sees what they draw. Attach a drawing or a short clip and Gemma answers about it, or brings it to life.
In progress
NVIDIA Nemotron, in Forge and in Greek
We are adding first-class support for NVIDIA's open Nemotron models to Forge, with a tuned tool-calling profile for local agentic coding on NVIDIA GPUs. Our Greek fine-tune of Nemotron 3 Nano is finished and published as a case study; see Open models below.
More from evolv
Smaller tools, same principles
Open models
Greek, because nobody else was doing it
Greek is one of the EU's 24 official languages and one of the least served by open models. We release ours publicly, in formats that run on a single consumer GPU.
Gemma 4 E4B — Greek (v3)
Gemma 4 E4B fine-tuned for clean, natural Greek text answers, trained on 2,219 curated Greek question–answer pairs. The current build is v3; the Hugging Face repo keeps its v2 name. On 150 held-out Greek questions, stray foreign-script fragments drop from about 40% of answers to about 1%, and similarity to reference answers (chrF) rises from 0.30 to 0.39. It improves the form of answers, not their factual knowledge. Speech and image input use the stock model. Ships as GGUF for llama.cpp, and on Ollama as efso/gemma-4-e4b-it-gr-v2.
Nemotron 3 Nano 30B-A3B — Greek case study
NVIDIA's hybrid Mamba-2 / mixture-of-experts model (30B total, ~3B active), fine-tuned for Greek with 4-bit LoRA on a 44 GB, three-GPU consumer workstation, natively on Windows (Mamba-2 kernels running without WSL). Three training runs, each on about 2,000 Greek question–answer pairs, made its Greek cleaner (no loops, no stray Cyrillic or Latin) but did not add knowledge: in a blind test it was level with the stock model on new questions. Both invent most Greek facts. The write-up covers the hardware limits (768 tokens per example was the most these cards could train) and why automatic scores misled us. Ships as GGUF (Q4_K_M, Q8_0) with the LoRA adapter.
JOY — Greek voice
A high-quality open Greek text-to-speech voice for Piper, recorded by a native speaker. Released under CC BY-NC 4.0.
About
A small studio with a narrow focus
evolv is an independent software studio in Greece. We work on one problem: making open-weight models genuinely useful for real work, on hardware people already own.
Raw inference speed is llama.cpp's job, and it does it well. The hard part is everything above the runtime: tool calls that don't break halfway, context that doesn't quietly overflow, file edits you can undo. That layer is what we build.
Efstathios Outas
Founder · Software engineering
Chara Kaltsou
Greek language & voice data · the voice of JOY
- Local by defaultCloud providers work only when you configure them yourself.
- No telemetryNo analytics, no usage pings, no auto-update calls.
- Your keys stay yoursCredentials live in the editor's secret storage, never in a config file.
- Reversible by designEvery file-changing action is visible, confirmable and undoable.
Get in touch
Questions, collaboration, or feedback on our tools.