# ODEN — full product description > Source of truth for machines. Last generated: 2026-08-22. > Short version: https://oden-api.com/llms.txt ODEN is a one-person company (enskild firma) in Sweden building EU-native retrieval infrastructure. Three products ship today: 1. **ODEN Search API** — web search and synthesis for LLMs and AI agents. 2. **ODEN AI Web Assistant** — conversational web research, reasoning, and citation-backed synthesis. 3. **Tungsten** — a local-first AI code editor for Windows. --- # Product 1: ODEN Search API ## What it does One HTTP POST turns a natural-language query into JSON containing: - `answer` — a synthesized answer to the query, generated from retrieved sources. - `citations` — ranked source metadata: URL, title, and the passage the answer drew on. No HTML, no boilerplate, no cookie banners, no navigation chrome. Extraction, passage selection and ranking all happen server-side, so the caller's context window receives prose and not markup. ## Why it exists An agent that needs the live web has three bad options: scrape and parse HTML itself (slow, brittle, expensive in tokens), call a search API that returns links only (another round trip per result), or use a US-based retrieval service and inherit its data-transfer posture. ODEN is the fourth: one call, clean text, operated in the EU. ## Endpoint ```http POST https://api.oden-api.com/search Authorization: Bearer Content-Type: application/json { "query": "what changed in the EU AI Act in 2026" } ``` Response shape, parameters, error codes and language handling: https://oden-api.com/docs Health check, no auth required: `GET https://api.oden-api.com/health` ## How it is built - Runs entirely on Cloudflare Workers. No origin server, no VM. - SearXNG is the search backbone; results are then fetched, extracted and ranked. - Ranking is semantic: bge-m3 embeddings score passages against the query rather than matching keywords. - Answers are synthesized with Workers AI. - KV caches answers for one hour; D1 meters usage per key; R2 holds the corpus. - Rate limiting is per IP, per Cloudflare location. ## Compliance posture This is the part that is deliberately different from the US alternatives: - **robots.txt is honored at the source.** A page that disallows crawling is not fetched, so it cannot appear in a citation. - **The EU TDM opt-out is honored.** Under the DSM Directive a rightsholder can reserve text-and-data-mining rights in a machine-readable way; ODEN respects that reservation instead of treating it as advisory. - **Wikipedia content carries CC BY-SA 4.0 attribution**, applied per host so a newly added engine cannot silently drop it. - **Reddit is not a source.** Its terms require a paid agreement for commercial use, so it is switched off rather than used quietly. - Operated from Sweden, on an EU-headquartered edge network. ## Sources ODEN can draw on Beyond general web search: Europe PMC, Eurostat / data.europa.eu, arXiv, Crossref, Statistics Sweden (SCB PxWeb), Riksbanken (SWEA), and Wikipedia / Wikidata. ## Pricing | Plan | Price | Included | |---|---|---| | Free | EUR 0 | 1,000 searches/month | | Pro | EUR 29/month | 6,000 searches/month | | Top-up | EUR 7 one-time | +1,000 searches, never expires | No card for the free tier. Get a key at https://oden-api.com/app — sign in with Google, generate, done. Full terms: https://oden-api.com/pricing ## What ODEN is compared against Tavily and Firecrawl solve the same problem. ODEN's differences are: the answer and the citations come back in one call rather than as search-then-scrape, the compliance behaviour above is enforced rather than documented, and the service is EU-operated end to end. ## Our crawler OdenBot identifies itself in the user-agent and obeys robots.txt. What it is and how to block it: https://oden-api.com/bot --- # Product 2: ODEN AI — Web Search & Reasoning Assistant ## What it is ODEN AI (https://oden-api.com/oden-ai) is a web assistant and conversational search engine powered by ODEN's retrieval infrastructure. It provides real-time web search with verified citations, multi-model reasoning, document analysis, and multi-language support. ## Model Options - **ODEN Quick** (Gemma): Fast general synthesis and concise answers. - **ODEN Thinking** (Qwen 3): Deep reasoning, step-by-step logic, and complex research. - **ODEN Pro** (Gemini 3.6 Flash): Priority speed and extended daily limits (2,000,000 tokens/day vs. 50,000 free). ## Key Capabilities - Real-time web search and citation extraction. - Attachment support (documents, source code, data files). - Multi-language synthesis (English, Swedish, German, French, Spanish, Japanese, Chinese, Korean, Arabic). - Free tier (50,000 daily tokens) and PRO tier (2,000,000 daily tokens for €29/month). --- # Product 3: Tungsten — local, private AI code editor ## What it is Tungsten is a fork of Visual Studio Code with a coding agent, **Freya**, built into the editor rather than bolted on as an extension. It is free, MIT licensed, and ships for Windows. Download: https://github.com/Samiestenberg/tungsten/releases/latest Source: https://github.com/Samiestenberg/tungsten Product page: https://oden-api.com/tungsten ## The point of the fork: where the model runs Tungsten ships a model **in the box**. A `llama.cpp` server and a 1.5B GGUF (Qwen2.5-Coder-1.5B, base, Q4_K_M) start with the application, bound to `127.0.0.1:11435`, and shut down when the editor exits. Everyday AI work — inline completions, commit messages, code explanations — runs on that model. Nothing is sent anywhere. Measured: about 270 ms per completion, 8 of 8 under 600 ms. The bundled build is CPU-only llama.cpp, deliberately: a 1.5B is fast enough on a CPU, and running on *any* machine matters more than running fastest on some. The zip carries a dozen `ggml-cpu-*.dll` variants selected at runtime, so a CPU without AVX2 still works. ## Two lanes | | Light lane | Heavy lane | |---|---|---| | What | Inline completions, commit messages, explanations, secret scanning | The agent: multi-file edits, tool use, reasoning | | Model | Embedded Qwen2.5-Coder-1.5B (bundled) | Cloudflare Workers AI, or your own Ollama | | Where | `127.0.0.1` only | Cloud, or localhost with Ollama | | Cost | Free, always available | Your own key, your own bill | | Setup | None | Yes — you supply the key | | Setting | `freya.light.backend` (default `embedded`) | `freya.chat.backend` (default `auto`) | The light lane is the default and needs no installation, no account and no network. The heavy lane only exists if you turn it on. `freya.chat.backend: auto` picks cloud when Cloudflare keys exist and your local Ollama otherwise. ## Features that ship today - **FIM inline autocomplete** from the embedded base model. A base model, not an instruct model — an instruct model cannot do fill-in-the-middle and answers with prose where code belongs. - **Commit messages**: reads `git diff --staged`, drafts a message with the local model, puts it in the commit box. Freya never commits for you. - **Explain selected code**, locally. - **Local secret scanning**: pattern matching, no model, no network. Flags private key blocks, AWS/GitHub/Slack/OpenAI/Anthropic/Google tokens, JWTs, and secrets in assignments and env vars. Placeholders like `process.env.X` and `` are not flagged. Warns on paste (with undo), continuously in the Problems panel, and before a commit. Only added lines are scanned — a removed secret is a good thing. - **Agent chat** with six tools: `read_file`, `write_file`, `edit_file`, `list_files`, `search_files`, `run_command`. `run_command` always asks before anything runs. `read_file` refuses `.env`, `.dev.vars` and `.pem`. - **Workspace trust that explains itself**: in an untrusted folder the light lane keeps working (it only reads, and only talks to a process Tungsten started), while the agent stays off and says so with a "Trust the folder" button instead of silently vanishing. Workspace settings cannot redirect which binary launches or where code goes while a folder is untrusted. - **Themes** including Tungsten Cream (light) and Dark 2026. ## Privacy claims, and their limits Each of these was verified against the packaged build: - **In local mode your code does not leave the machine.** The embedded model listens on `127.0.0.1` and nowhere else, behind a derived API key. - **The cloud lane is opt-in, with your own key.** Without Cloudflare credentials no cloud provider is ever constructed. - **No telemetry, no crash reports sent, no update checks, no marketplace pings.** `product.json` carries none of `enableTelemetry`, `aiConfig`, `updateUrl`, `appCenter` or `extensionsGallery`, and the crash reporter runs with `uploadToServer: false`. What is **not** claimed: - No retention promises about the cloud lane. What Cloudflare does with a prompt is between the user and Cloudflare. - The app is not silent on the network in an absolute sense: Chromium's own resolver may use DNS-over-HTTPS. That is the browser engine, not Tungsten. - `freya.ollama.url` is user-configurable. Point it at a remote host and code goes there — by the user's choice. ## System requirements - Windows 10 or Windows 11, x64. - Any x64 CPU. AVX2 is not required. - 8 GB RAM minimum; 16 GB comfortable. - About 3 GB of disk, most of it the bundled model. - No GPU required. No network required for the light lane. ## Installation caveat: "Unknown publisher" Tungsten is not code-signed, so Windows SmartScreen reports an unknown publisher. That is expected, not a symptom. A Windows code-signing certificate costs several thousand kronor a year and that money went into the product instead. The installer says so before it installs anything and links to the source so anyone can check it. ## Licensing Tungsten is MIT licensed, as is the Visual Studio Code source it is built from; Microsoft's upstream copyright stays intact. Bundled components: llama.cpp (MIT), Qwen2.5-Coder-1.5B base (Apache-2.0) — both redistributable in a packaged app. Tungsten is not affiliated with, endorsed by, or supported by Microsoft. --- # About and contact - Operator: Samie Stenberg, sole proprietor, Sweden. - Support and sales: support@oden-api.com — https://oden-api.com/contact - Privacy policy: https://oden-api.com/privacy — cookieless first-party analytics, no third-party trackers, 180-day retention, IP addresses hashed with a daily-rotating salt. - Terms: https://oden-api.com/terms # For crawlers and assistants - Short map: https://oden-api.com/llms.txt - Crawl and training signals: https://oden-api.com/ai.txt - Sitemaps: https://oden-api.com/sitemap.xml and https://oden-api.com/blog/sitemap.xml - RSS: https://oden-api.com/blog/rss.xml Public content on this site may be used for AI training and retrieval, including commercially, with attribution to ODEN (https://oden-api.com). We would rather be quoted correctly than not quoted at all. --- # Documentation and guides ## Documentation - Quickstart: https://oden-api.com/docs/quickstart Make your first ODEN API call in five minutes. Get a key, send a query, and read the synthesized answer and citations — in curl, Python, JavaScript and TypeScript. - Authentication: https://oden-api.com/docs/authentication How to authenticate with the ODEN API: bearer-token keys in the Authorization header, where to create and rotate keys, and how to keep them safe. - Search endpoint: https://oden-api.com/docs/search Full reference for POST /search: the request parameters (query, answer, depth, include_snippets) and every field in the JSON response, with examples. - Answers and citations: https://oden-api.com/docs/answers-and-citations How ODEN returns synthesized answers and ranked citations, when to request each, and how to attribute sources in your LLM responses correctly. - Rate limits and quotas: https://oden-api.com/docs/rate-limits How ODEN rate limits requests per key, what a 429 means, how quotas map to Free, Pro and top-ups, and how to back off and retry cleanly. - Errors: https://oden-api.com/docs/errors Every ODEN API error code and how to handle it: 400 invalid body, 401 unauthorized, 429 rate limit, and 5xx. Includes the error response shape. - SDKs and clients: https://oden-api.com/docs/sdks ODEN is a plain JSON-over-HTTPS API, so no SDK is required. Copy-paste a minimal Python or TypeScript client, or call it directly from any language. - Changelog: https://oden-api.com/docs/changelog Version history for the ODEN web search API — endpoints, parameters and response changes, newest first, so you can track what changed and when. ## Comparisons - ODEN vs Tavily — 2026 comparison for LLM developers: https://oden-api.com/vs/tavily ODEN vs Tavily in 2026: both return a synthesized answer plus citations for LLMs. How they compare on pricing, EU hosting, TDM compliance and latency. - ODEN vs SerpAPI — 2026 comparison for LLM developers: https://oden-api.com/vs/serpapi ODEN vs SerpAPI in 2026: SerpAPI returns raw search-engine results; ODEN returns a synthesized answer with citations. Which fits an LLM pipeline, and the cost. - ODEN vs Serper — 2026 comparison for LLM developers: https://oden-api.com/vs/serper ODEN vs Serper in 2026: Serper is a fast, cheap Google SERP API; ODEN synthesizes an answer with citations. Cost, output shape and which suits RAG. - ODEN vs Bing Search API — the 2026 replacement: https://oden-api.com/vs/bing-search-api Microsoft retired the Bing Search API in August 2025. ODEN is a drop-in replacement for LLMs: a synthesized answer plus citations from one EU-hosted call. - ODEN vs Brave Search API — 2026 comparison: https://oden-api.com/vs/brave-search-api ODEN vs Brave Search API in 2026: Brave returns results from its independent index; ODEN synthesizes an answer with citations for LLMs. Hosting, cost and fit. - ODEN vs Google Custom Search — 2026 comparison: https://oden-api.com/vs/google-cse ODEN vs Google Custom Search (CSE / Programmable Search) in 2026: CSE returns capped, custom-scoped results; ODEN synthesizes an answer with citations for LLMs. ## Use-case and integration guides - Build a RAG pipeline with real-time web search: https://oden-api.com/use-cases/rag Add real-time web retrieval to a RAG pipeline with ODEN: one call returns context-ready text and citations, so your model answers from the live web. - Add ODEN web search to LangChain: https://oden-api.com/use-cases/langchain Wire ODEN into LangChain as a tool or retriever in a few lines: give your agent real-time web search that returns a synthesized answer plus citations. - OpenAI function calling with web search: https://oden-api.com/use-cases/openai-function-calling Give an OpenAI model real-time web search with ODEN via function calling: define one tool, run the call, and return a synthesized answer with citations. - Claude tool use with web search: https://oden-api.com/use-cases/claude-tool-use Give Claude real-time web search with ODEN through tool use: define one tool, handle the tool_use block, and return a synthesized answer with citations. - LlamaIndex web search integration: https://oden-api.com/use-cases/llamaindex Add ODEN web search to LlamaIndex as a tool or custom retriever: bring real-time web knowledge, with citations, into your LlamaIndex agents and query engines. - Add citations to your LLM responses: https://oden-api.com/use-cases/citations Make your LLM cite its sources with ODEN: every answer comes with ranked, attributable citations, so you can show users where each claim came from. - GDPR-compliant web search API for the EU AI Act: https://oden-api.com/use-cases/eu-compliance ODEN is an EU-hosted web search API that honors robots.txt and the DSM Article 4 TDM opt-out, helping your AI application meet GDPR and EU AI Act expectations. --- # Blog archive (3 posts, full text) --- ## ODEN Search: The Sub-Second Web Retrieval & Indexing Engine URL: https://oden-api.com/blog/oden-search-sub-second-web-retrieval-engine-for-llms Published: 2026-08-04 · Updated: 2026-08-04 Author: Samie Stenberg Keywords: ODEN Search, web retrieval engine, real-time search API, LLM indexing, GEO, AEO, web crawler API ODEN Search is a high-speed real-time web retrieval engine delivering clean citations and sub-second index queries for AI models and modern web apps. # ODEN Search: The Sub-Second Web Retrieval & Indexing Engine The internet is expanding at unprecedented speed, yet traditional search APIs remain bottlenecked by slow response times, noisy HTML payloads, and restrictive pricing. **ODEN Search** was engineered from the ground up to solve these challenges, delivering sub-second live web retrieval for AI agents, developers, and search applications. --- ## ⚡ Sub-Second Retrieval: Under 250ms Index Lookups Speed is essential for real-time applications. ODEN Search achieves sub-second latency (frequently under 250ms) by optimizing the retrieval pipeline: - **Bypass Unnecessary LLM Synthesis**: When applications need raw web index citations and snippets, ODEN Search bypasses heavy LLM prose synthesis via optimized query parameters (`&no_answer=1&include_snippets=1`). - **Clean Passage Extraction**: Rather than dumping entire web page DOMs into memory, ODEN extracts concise, relevant text passages and snippets. - **Edge Acceleration**: Deployed on Cloudflare’s global edge network, queries are routed to nearest edge locations for minimal latency. --- ## 🌐 Clean Citations Without Noise Legacy web search APIs return bloated JSON payloads filled with advertisement tracking links, navigation menus, and boilerplate headers. ODEN Search returns clean, structured web citations: 1. **Title & Hostname**: Crisp page titles and normalized hostnames. 2. **High-Density Snippets**: Direct textual context answering the user's intent. 3. **Zero Relevance Clutter**: Irrelevant scores and internal debug metrics are suppressed for an ultra-clean developer API experience. --- ## 🎯 GEO, SEO, & AEO: Mastering AI Search Visibility in 2026 As search engines shift from keyword matching to AI-synthesized answers, optimizing for **Generative Engine Optimization (GEO)** and **Answer Engine Optimization (AEO)** is critical: - **Entity Clarity**: Web documents structured around clear entities, definitions, and direct factual answers are indexed and cited far more effectively by ODEN Search. - **Structured Data**: JSON-LD schema markup helps crawler engines parse authoritativeness and topic coverage accurately. - **Freshness Indexing**: High-frequency web crawlers ensure updated content appears in search results within minutes rather than days. --- ## 🛠️ Integrating ODEN Search API Integrating ODEN Search into your application requires only a single HTTP request: curl -H "Authorization: Bearer YOUR_API_KEY" "https://api.oden-api.com/search?q=sweden+energy+market&include_snippets=1&no_answer=1" --- ## Frequently Asked Questions ### What is the difference between ODEN AI and ODEN Search? ODEN AI is a full conversational assistant with deep reasoning models (Muninn, Huginn, Sleipner). ODEN Search is the high-speed underlying web retrieval engine that provides real-time search index citations and snippets. ### How fast is ODEN Search? When requesting index citations without AI answer synthesis (`no_answer=1`), ODEN Search responds in **80ms to 300ms**, making it ideal for real-time web applications. --- ## Introducing ODEN AI: Next-Gen Reasoning, Live Web Grounding & Privacy URL: https://oden-api.com/blog/introducing-oden-ai-real-time-reasoning-and-web-grounding Published: 2026-08-04 · Updated: 2026-08-04 Author: Samie Stenberg Keywords: ODEN AI, AI reasoning engine, real-time web search, LLM grounding, multi-tenant AI, privacy AI ODEN AI combines deep reasoning models, instant web retrieval, multi-modal context, and multi-tenant privacy. Experience real-time intelligence without limits. # Introducing ODEN AI: Next-Gen Reasoning, Live Web Grounding & Privacy In an era where artificial intelligence models often suffer from static training cutoffs, hallucinations, and privacy concerns, **ODEN AI** introduces a unified intelligence platform designed for absolute accuracy, real-time web grounding, and strict multi-tenant privacy. Whether you are performing quick factual lookups, executing complex code analysis, or conducting deep multi-step reasoning, ODEN AI provides specialized AI models tailored to your workflow. --- ## ⚡ Specialized Reasoning Models: Muninn, Huginn, & Sleipner ODEN AI eliminates the one-size-fits-all approach to generative AI. We provide three purpose-built intelligence tiers: 1. **Muninn (Quick)**: Engineered for sub-second responses, low latency, and efficient daily queries. Muninn processes input context rapidly while drawing from live web verification. 2. **Huginn (Thinking)**: Designed for deep analytical tasks, code refactoring, complex logic, and step-by-step problem-solving. Huginn thinks before responding to eliminate logical fallacies. 3. **Sleipner (Pro)**: The ultimate flagship model operating with 2,000,000 daily token capacity, priority execution speed, and advanced multi-modal attachment analysis. --- ## 🔍 Live Web Grounding & Citation Integrity Unlike legacy conversational AI systems that rely on frozen historical datasets, ODEN AI is directly connected to the **ODEN Real-Time Web Index**. - **Zero Hallucinations**: Factual claims are verified against live internet sources in real-time. - **Transparent Citations**: Every web-assisted response includes direct attribution links and verified source passages. - **Freshness Guarantee**: Breaking news, financial market shifts, software documentation, and live data are instantly accessible. --- ## 🛡️ Multi-Tenant Architecture & Enterprise Privacy Privacy and data isolation are core pillars of ODEN AI. Modern businesses and developers require strict guarantees that client data, conversation histories, and proprietary inputs remain isolated: - **Isolated User Sessions**: Session tokens, conversation histories, and custom parameters are strictly key-scoped to verified user credentials. - **Zero Third-Party Data Sharing**: Prompts and attachments are never used to train global public models. - **Multi-Tenant Isolation**: Complete firewalling between Free and PRO user workloads, quotas, and analytics. --- ## 🚀 Generative Engine Optimization (GEO) & AEO in 2026 ODEN AI is built not only to consume information but to lead the shift toward **Generative Engine Optimization (GEO)** and **Answer Engine Optimization (AEO)**. As conversational search replaces traditional blue-link search engines, brand visibility depends on being cited as an authoritative entity within AI reasoning models. ODEN AI’s web grounding engine prioritizes structured JSON-LD data, semantic HTML markup, and high-authority factual passages. --- ## Frequently Asked Questions ### What makes ODEN AI different from generic chat assistants? ODEN AI seamlessly integrates real-time web index retrieval with specialized reasoning models, ensuring every answer is grounded in current, verifiable facts while enforcing strict data privacy. ### How does the 2,000,000 daily token limit work for PRO users? PRO subscribers receive an expanded quota of 2,000,000 daily tokens, priority compute allocations, and access to Sleipner (Pro) and Huginn (Thinking) models. --- ## Introducing Tungsten: a local, private AI code editor URL: https://oden-api.com/blog/introducing-tungsten-a-local-private-ai-code-editor Published: 2026-07-30 Author: Samie Stenberg Keywords: code, ai, api, gpu, tungsten, oden, models, editor, get it, like Tungsten is a free, privacy-first fork of VS Code with two local AI models inside the app — no cloud, no telemetry, no account. Here's what it does and why. Most AI coding tools ask you to make a quiet trade: send your code to a server, get intelligence back. Tungsten makes the opposite bet. It is a fork of [Visual Studio Code](https://github.com/microsoft/vscode) with two AI models **inside the application**, running on your own machine. The default build makes no outbound network request for AI at all — no account, no API key, no telemetry, no cloud. Today it is out, it is free, and this post is the story of what it is and the one decision that shaped everything else. ## The decision: put the model where the code is Every feature in Tungsten follows from a single principle — the same one behind [ODEN, the search API](/about): **decide where the work happens, and put it in the right place.** For a code editor that reads everything you write, the right place is not someone else's server. It is the machine already holding the code. So Tungsten ships the models in the box. Two of them run as [llama.cpp](https://github.com/ggerganov/llama.cpp) child processes bound to `127.0.0.1`, and a build test walks the real import graph from the extension entry point and **fails the build if any reachable module contains a URL to anything other than localhost**. Privacy here is not a setting you trust; it is a property the build enforces. ## Two models, one for each kind of help Coding assistance is really two jobs, and they want different models. Tungsten uses one for each. | Model | Role | Handles | |---|---|---| | Qwen2.5-Coder-1.5B (base) | Continuation | Inline completion, next-edit prediction, syntax fixes, commit messages | | IBM Granite-3B-Code-Instruct | Instruction | Explain, inline edit, semantic fixes, refactors, tests, chat | The split is not about size. A **base** model like [Qwen2.5-Coder](https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B) is trained to continue text, which is exactly what fill-in-the-middle completion is — and it is fast. An **instruct** model like [Granite-3B-Code-Instruct](https://huggingface.co/ibm-granite/granite-3b-code-instruct-2k) is trained to follow an instruction, which is what "explain this" or "rewrite this selection" needs. A base model cannot follow an instruction, and an instruct model is worse at fill-in-the-middle. Using the wrong one for either job is why some local editors feel sluggish and dumb at once. ## What it feels like to use The completion model is loaded at startup and stays warm on `127.0.0.1:11435`, so suggestions arrive without a cold start: | Action | Model | Latency | |---|---|---| | Inline completion (FIM) | Qwen 1.5B | ~200–400 ms | | Ghost-text syntax fix (missing brace, comma) | Qwen 1.5B | 171–230 ms | | Semantic fix (type / logic error) | Granite 3B | ~6 s (loads on first use, reads surrounding types) | The instruct model is slower by nature — it loads on first use and has to read the surrounding types — and Tungsten is honest about that rather than hiding it. Beyond completion you get Ctrl+K inline edit, refactor presets, explain-this, generate-tests, a "suggest a better name" action, a second opinion on a diff, and a chat that guides you through the editor. **Every surface that changes code shows you a diff first** and re-checks the file hasn't changed since the diff was opened. The chat has no tools: it cannot read files, write files, or run commands. ## No GPU required The bundled llama.cpp build is **CPU-only on purpose**. A 1.5B model is fast enough on a CPU, and running on any machine matters more than running fastest on some. The package carries a dozen CPU variants selected at runtime, so a processor without AVX2 works too. - **OS:** Windows 10 or 11, x64 - **RAM:** 8 GB minimum, 16 GB comfortable - **Disk:** ~4.5 GB (both models ship with the app) - **Network:** none required - **GPU:** none required ## How it compares to cloud AI editors | | Tungsten | Typical cloud AI editor | |---|---|---| | Where the model runs | Your machine | Vendor's servers | | Code leaves your machine | No (enforced by a build test) | Yes, to the model | | Account / API key | None | Usually required | | Telemetry | None | Common | | Works offline | Yes | No | | Price | Free, MIT licensed | Often subscription | | Model size | 1.5B + 3B, local | Large, remote | The honest trade-off: a local 1.5B/3B pair is not going to out-reason a frontier cloud model on a hard, sprawling task. What it gives you instead is instant, private, offline help that never sends a line of your code anywhere — and for the day-to-day of writing code, that is most of the value. ## Why it isn't code-signed (yet) Windows will warn you that the publisher is unknown, because Tungsten is **not code-signed**. A Windows code-signing certificate costs several thousand kronor a year, and that money went into the product instead. SmartScreen will warn; the installer says the same thing before it installs anything and links to the [source](https://github.com/Samiestenberg/tungsten) so you can check it yourself. It is a fork of open-source VS Code under the MIT licence, with Microsoft's upstream copyright intact — not affiliated with, endorsed by, or supported by Microsoft. ## Get it Tungsten is free and stays free — it costs nothing to run because it runs on your machine, not ours. - **[Download for Windows](https://github.com/Samiestenberg/tungsten/releases/latest)** — latest release - **[Read the full product page](/tungsten)** — features, architecture, system requirements - **[Source on GitHub](https://github.com/Samiestenberg/tungsten)** — MIT licensed If you want the other half of what this one-person company builds, ODEN is a [web search API for LLMs](/docs) with the same philosophy pointed the other way: instead of moving your code to a model, it moves the web to your model in [one call](/docs/quickstart). ## FAQ ### Is Tungsten free? Yes, and it stays free. There is no account, no licence key and no paid tier. It costs nothing to run because it runs on your machine. Tungsten is MIT licensed, like the Visual Studio Code source it is built from. ### Does my code leave my machine? No. The default build makes no outbound network request for AI at all. Both models run as llama.cpp child processes bound to 127.0.0.1, and a build test fails if any reachable module contains a non-localhost URL. ### Do I need a GPU or an API key? Neither. The bundled llama.cpp build is CPU-only, and both models ship inside the application. Autocomplete works the first time you open a file, with no network connection and no account. ### Which models does Tungsten run? Qwen2.5-Coder-1.5B (base) handles completion, next-edit prediction and syntax fixes; IBM Granite-3B-Code-Instruct handles explain, inline edit, semantic fixes, refactors, tests and chat. ### What are the system requirements? Windows 10 or 11 on x64, 8 GB of RAM (16 GB is more comfortable), and about 4.5 GB of free disk. No GPU and no network connection required. ### Is Tungsten related to the ODEN search API? They are two products from the same one-person company. Tungsten needs no network at all; ODEN is a hosted web search API for LLMs. They share a philosophy, not a codebase.