A local, safe coding agent for small and medium tasks, with your skills. Local, safe, with real documents and measurements. Source-available, yours.
On the native Windows install the agent doesn’t run commands (no sandbox there); on Linux it runs them isolated.
All numbers: one test machine, NVIDIA RTX 4000 Ada 20 GB + 125 GB RAM, Linux. Coding and speed tests: October 24, 2026. Your hardware will move them.
--cpu-moe), 18 tok/s (--n-cpu-moe 44). This is generation speed, not task time: a task also depends on reading the prompt and the number of steps.A research chat on the world coffee market, then “Write an ebook on this research”. Forge wrote the text, checked the facts against the sources, drew the cover and three chapter openers and laid out the PDF, all on our own computer.
Ebook in 4 min 28 s · research + ebook 14 min 17 s
| Sourced research (3 questions, reading ICO and USDA pages) | 9 min 49 s |
|---|---|
| Outline | 14 s |
| Writing the 3 chapters | 2 min 05 s |
| Fact check | 14 s |
| Cover and 3 illustrated openers | 1 min 51 s |
| PDF and Word layout | 4 s |
| Ebook, from request to PDF | 4 min 28 s |
| Research + ebook | 14 min 17 s |
Made with Velquaria Forge, local models: text by gemma4:26b, images by Qwen-Image 2512 (⚡ Fast mode), on an NVIDIA RTX 4000 Ada 20 GB. Text and images are AI-generated and declared as such (metadata in the file, EU AI Act art. 50); the figures come from the cited sources, and the claims in yellow are interpretations not found in the sources.
Each item states its limit too. Timings are measured on our test machine.
Type «write an ebook about…». Before it starts, Forge asks for language (Italian or English), length (short ~15, medium ~25, long ~40 pages) and cover, then up to three questions from the model on genre, audience and style. Every one can be skipped. Write in English and Forge detects it and asks its questions in English. The PDF comes out laid out like a book; the Word copy is there to edit it.
| Model | Where | Time |
|---|---|---|
| gemma4 26B | Local GPU | 177 s |
| qwen3.6 35B | Local GPU | 255 s |
| gpt-oss-120b | Free cloud | ~330 s |
| Layout, 25 pages | The PC’s Chrome or Edge | ~4 s |
With gemma4 26B: 25 pages, 5,408 words, about 70 tok/s. RTX 4000 Ada 20 GB + 125 GB RAM, Linux; also tested on Windows with Edge (there the agent doesn’t run commands). No extra dependency for the layout.
Essay, manual, professional, novel, short stories. A novel is written in scenes and dialogue, keeps continuity between chapters and has no tables. Each genre gets its typeface: EB Garamond for fiction, Newsreader for essays, Archivo for manuals. Six color and cover styles, picked by the model or by you. Real test: the novel «Luce d’Ombra», 17 pages, 5 chapters.
Qwen-Image, through ComfyUI, paints the art behind or inside the cover.
Limit: only where ComfyUI and a GPU are available. Otherwise the cover is graphic.
Highlighted figures, percentages in tables, and bar or ring charts stay only if they come from sourced material in a 🔎 Research. No sources, no statistics, and the colophon says so. Also: long words and URLs don’t shrink the book, no repeated numbers or charts, and short chapters are set tighter to avoid near-empty pages.
A second model rates numbers, dates and names 🟢 🟡 🔴 and gives its reason. There is a button under every answer; in non-fiction ebooks it runs on its own, with a Word report. Measured: 5 claims in 62 s with a free cloud model. In a test ebook it caught a real error.
Limit: it’s a second opinion from another model, not proof. Sources still need checking.
Real .docx and .xlsx files: the sheets hold real numbers, filters and formulas, not pasted text. Generated by the app with no external libraries.
Attach SDI XML, P7M or ZIP files and get the invoices, VAT by rate, due dates and anomalies (duplicates, totals that don’t add up). No AI model involved: nothing leaves the PC.
Limit: XML, P7M and ZIP only. No PDFs or scans. With the English interface the summary is in English; the Excel sheets stay in Italian.
Describe tasks in plain words on a folder, or ask for an ebook, and pick a time. They run on their own; in the morning you find one chat per task.
Limit: the PC has to stay on, not asleep. No commands and no deletions.
Interface in Italian and English. A native Windows installer, and a single zip for Linux and Windows.
Limit: On the native Windows install the agent doesn’t run commands (no sandbox there); on Linux it runs them isolated.
The agent reads the repository’s status, diffs and history. It commits only when you ask and never pushes. Before every answer Forge takes a hidden snapshot of the folder: «Undo» also restores files that commands created, changed or deleted, without touching history, branches or the index.
Limit: no push, branches or merges: those stay with you.
It reads files in chunks, with numbered lines and search with context, and several files in the same step. For questions about the code there is «explore», a read-only helper that answers with facts, paths and line numbers.
Limit: one helper per question, read-only.
velquaria "do X" works on the current folder and exits with 0, 1 or 2, so it fits scripts and cron.
Limit: the app has to be open. On Windows you run it as python -m studio.cli.
A new model stays «to be tested». One click runs the standard test on your PC, new against old, and gives you the verdict with the numbers.
Now for images and video too. When a new or rival model comes out, Forge downloads it without removing the current one, runs both on the same prompts and shows you time, GPU memory and the results side by side: you choose which to keep.
Qwen-Image 2512
Previous versionForge’s agent in Write mode, 2 attempts, at most 15 steps, tests run by the program. Not Aider’s official score: 87 of 225 exercises, 3 of 6 languages.
| Model | Python /34 | Rust /30 | C++ /23 | 1st attempt | 2nd attempt |
|---|---|---|---|---|---|
| Gemma 4 26B | 19 | 13 | 18 | 35/87 (40.2%) | 50/87 (57.5%) |
| Qwen3.6 35B-A3B | 18 | 11 | 14 | 34/87 (39.1%) | 43/87 (49.4%) |
By language: solved on the second attempt. Aider protocol with 2 attempts, tests run in a sandbox with no network; Forge’s agent uses its own tools, not Aider’s edit formats. Go, JavaScript and Java not tested; 3 C++ exercises excluded because they fail even with the reference solution on g++ 13. During the Qwen3.6 run other processes used the GPU for a few minutes. The test script ships with Forge: tests/bench_polyglot.py.
Licenses of the models named here, checked on Hugging Face: the Gemma 4 family, Qwen3.6-35B-A3B, Qwen-Image 2512 and gpt-oss-120b are Apache-2.0. Strata is MIT. The book fonts (EB Garamond, Newsreader, Archivo, JetBrains Mono) are OFL.
Cline, Aider, Continue and Open WebUI already talk to Ollama. The hard part is making local models good enough to rely on.
Every recommended model is tested on our standard set of small tasks, on the same machine. The numbers on this page say how we measured them, and the requirements table says which model fits your PC.
Ebooks in PDF and Word, Word and Excel under every answer, e-invoices to Excel, claim check: see What’s new. Plus web research with citations checked sentence by sentence, notebooks on your own sources, Artifacts (report, infographic, pitch deck, table, mind map), images and video with ComfyUI, voice, reverse engineering with REA and Ghidra, multiple users, and use from your phone.
Forge runs Strata, an open-source engine (MIT) that keeps the experts of a mixture-of-experts model in RAM and the rest on the GPU: Qwen3.8-Flash-Next 125B generates 60–80 tok/s on a 20 GB card. That’s generation speed, not task time.
Uses your Claude Code skills (167 on our test machine), adapted automatically, and ships with 21 of its own. Each one loads only when the task needs it. Local MCP servers work too.
Honest note: Forge is built for small and medium tasks that have to stay on your machine. For large codebases and long tasks, cloud agents are still the reference.
Based on our measurements. Linux and Windows today, macOS coming. On the native Windows install the agent doesn’t run commands (no sandbox there); on Linux it runs them isolated.
Based on our measurements on one machine; your speeds will differ.
| Use | Hardware | Models |
|---|---|---|
| Use the app | Any Linux or Windows PC (on Windows the agent doesn’t run commands), 8 GB RAM, no GPU needed | Small models on CPU, or free cloud tiers with your own keys |
| Local coding agent | GPU with 8 GB VRAM, 16 GB RAM | qwen3.5 9B, tested on our standard set of small tasks |
| Recommended | GPU with 16–24 GB, 32 GB RAM | gemma4 26B, qwen3.6 35B-A3B |
| The fast 125B | 20 GB GPU + ~128 GB RAM | Qwen3.8-Flash-Next 125B with Strata (measured on an RTX 4000 Ada 20 GB + 125 GB RAM) |
Everything on this page (screenshots, videos, speeds) was produced locally on an NVIDIA RTX 4000 Ada 20 GB + 125 GB RAM, with the models named next to each sample. With less memory you use the models in the rows above: slower or simpler answers.
Forge ships under the FSL-1.1-ALv2 license: you can read, audit, change and use the code, at work too; you can’t turn it into a competing product. Each version becomes open source (Apache-2.0) two years after release. Every feature stays in the app. Paid plans will cover the service around it — early updates, signed installers, tested model and skill packs, support — never the features themselves.
Free. The list runs on Substack (velquaria.substack.com); unsubscribe in one click. Sender: Velquaria, Italy — company details coming soon. Privacy
