Citation and evidence

Jan

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

Developer ToolsLarge Language ModelsOpen Source AI

Cite this article

Jan is an open-source desktop application for downloading and running large language models on a user's own computer, with optional connections to hosted model providers. Its GitHub repository describes it as "an open source alternative to ChatGPT that runs 100% offline on your computer", and its documentation calls it "an open-source replacement for Claude and ChatGPT".[1][2] Jan runs GGUF models through a built-in llama.cpp engine, can run MLX models on Apple silicon, serves loaded models over a local OpenAI-compatible API, and acts as a host for Model Context Protocol (MCP) servers.[3][4][5][6]

The application is developed by Menlo Research (legal entity Menlo Research Pte. Ltd.), which earlier operated as Homebrew Computer Company and, before that, simply as Jan.[7][8][9] The code has been licensed under the Apache License 2.0 since May 2025, after earlier periods under a "Sustainable Use License", the GNU GPL v3 and the GNU AGPL v3.[10][11][12][13] As of November 17, 2026 the latest stable release is version 0.8.8, published on October 28, 2026, and the repository janhq/jan has 45,755 stars and 3,178 forks.[14][1]

Developer and history

From Jan to Homebrew to Menlo Research

The janhq/jan repository was created on August 17, 2023, and the earliest release still listed on GitHub is version 0.2.0 of October 13, 2023.[1][15] On July 30, 2024 the team wrote that it had started work on Jan a year earlier, that the app had passed one million downloads, and that it was "becoming Homebrew, an AI research lab" with Jan, a self-hosted platform called Cortex, model research and hardware as its areas of work.[16] The project's About page from that period is titled "Homebrew Computer Company" and describes the company as "the creators and maintainers of Jan, Cortex and other tools"; it explains the name as a reference to the 1970s Homebrew Computer Club and calls the company a bootstrapped startup.[8] The README of January 2025 likewise said the company was bootstrapped and had no external investors.[17]

On March 25, 2025 the company announced that "Homebrew (Jan) is now Menlo Research", describing Menlo as "an open R&D lab building the brain for robots" and stating that Jan, Cortex and its models would stay open.[9] Jan's privacy policy names Menlo Research Pte Ltd as the operating company, and the Linux package on Flathub is published under the same name.[7][18] Runa Capital's ROSS Index, which ranked Jan 20th among open-source startups by growth in GitHub stars during 2024, lists the company as founded in 2023 and headquartered in Singapore.[19] Jan's own handbook says the team is small and "mostly based in Ho Chi Minh City, Vietnam", while Menlo's handbook says its staff work across Singapore, Vietnam and the United States.[20][21]

Menlo now describes itself as a robotics company. Its handbook says it launched Asimov, an open humanoid robot, in April 2026, and describes Jan as an open-source AI agent with 7 million downloads.[21] The GitHub organization followed the renamings. Release notes from August to October 2025 link to github.com/menloresearch/jan, and from version 0.7.3 (November 13, 2025) they link to github.com/janhq/jan again.[22][23]

Release history

Jan has stayed on 0.x version numbers throughout. The table lists selected releases; dates are GitHub publication dates in UTC, which in a few cases differ by a day or more from the date on the project's own changelog.[24][25]

VersionDateNotable changes
0.2.02023-10-13Earliest release listed on GitHub [15]
0.4.32023-12-21Remote OpenAI models through an API key [26]
0.4.52024-01-29Local API server [27]
0.4.132024-05-20Nitro engine replaced by Cortex [28]
0.5.72024-10-24Stable, beta and nightly channels [29]
0.5.122024-12-30Opt-in and opt-out settings for analytics [30]
0.5.162025-03-24Security fixes for flaws found in Cortex [31]
0.6.02025-06-19Tauri build replaces Electron; Apache 2.0 license [32]
0.6.32025-06-26MCP support in the stable build; Jan-Nano-128k [33]
0.6.62025-07-31Cortex models migrated to the llama.cpp extension; Hugging Face provider [34][35]
0.6.92025-08-28Image input; MCP leaves experimental status [36]
0.7.02025-10-02Projects for grouping chats [37]
0.7.62026-01-27Reworked chat interface; Anthropic Messages API format on the local server [38]
0.7.72026-02-11MLX backend on macOS [39]
0.7.82026-03-11Jan CLI; OpenClaw integration (removed in 0.7.9) [40][41]
0.8.02026-05-22Single llama.cpp router process; Multi-Token Prediction; inline MCP approval [42]
0.8.42026-07-23Built-in web search and fetch tools; credentials moved to the OS keyring [43]
0.8.52026-10-08llama.cpp compiled into the app; local models dropped on Intel Macs; CLI unbundled [44]
0.8.82026-10-28Security release; newer bundled engine; origin checks on the local API [14]

Expanded article table

License

The LICENSE file has changed four times, and each change is visible in the repository history.

Date of commitLicenseNotes
2023-09-06"Sustainable Use License"First LICENSE file in the repository [10]
2023-10-15GNU GPL v3[11]
2023-10-25GNU AGPL v3In force from 0.2.1 through the 0.5 series [12]
2025-05-20Apache License 2.0Pull request 5042; first shipped in 0.6.0 [13][32]
2025-08-26Apache License 2.0File shortened to a notice, "Copyright 2025 Menlo Research" [45]

Expanded article table

The current file is a short Apache 2.0 notice in Menlo Research's name, with one added sentence: "Attribution is requested in user-facing documentation and materials, where appropriate."[46] Because the file is a notice and not the full license text, GitHub's license field for the repository reads "Other" instead of Apache-2.0; the README and the documentation both state Apache 2.0.[1][47][2] The change replaced a copyleft license with a permissive one. These licenses cover the application code; the terms attached to model weights are a separate matter, discussed under open-source AI and open weights.

Related components carry their own licenses. The former engine repository cortex.cpp is Apache 2.0 and archived, and the models that the company publishes on Hugging Face under the names Jan, Jan-nano and Lucy are tagged Apache 2.0.[48][49]

Architecture

Application shell

Jan began as an Electron application. Version 0.6.0 added a build on Tauri, a framework that pairs a web front end with a Rust back end, and retired the Electron builds; GitHub now reports Rust as the repository's main language, and the README credits llama.cpp, Tauri and Scalar.[32][1][47] Chats, settings, logs and downloaded models are kept in a local data folder.[50]

Inference engines

The engine underneath Jan has been replaced several times.

  • Nitro (2023 to May 2024). The README of October 2023 said that Jan ran Nitro, "an open source, C++ inference engine", in the background, for GGUF and TensorRT model formats.[51] An extension for NVIDIA's TensorRT-LLM followed in version 0.4.9.[52]
  • Cortex (May 2024 to mid-2025). Version 0.4.13 replaced Nitro with cortex-cpp, which the company also offered as a standalone local API platform.[28][48] The cortex.cpp repository is now archived, with a notice that "This repository is no longer actively maintained."[48]
  • llama.cpp directly (since mid-2025). Version 0.6.6 migrated Cortex models to a llama.cpp extension, and 0.6.8 removed the remaining Cortex code.[34][22] From 0.8.0 the engine runs as one router process that loads and unloads models on demand, where earlier versions started a separate server per model.[42] Up to 0.8.4 Jan downloaded a llama.cpp backend to suit the machine; since 0.8.5 the engine is compiled into the installer, runs in a supervised worker process and updates only with the app.[44][3] Users who need a different llama.cpp build are told to run their own llama-server and connect it as a custom endpoint.[3]
  • MLX (since February 2026). Version 0.7.7 added MLX as a second backend on macOS.[39] It is built on mlx-swift-lm, needs macOS 14 and Apple silicon, and is labelled experimental: the documentation lists missing embeddings, no separate reasoning output, and some newer architectures that fail to load.[4]

Version 0.8.0 also exposed llama.cpp's Multi-Token Prediction for models that carry the extra prediction heads; the release notes class it as speculative decoding.[42][53]

Models and the Hub

Local models for the llama.cpp engine are GGUF files; the MLX engine uses MLX-format models.[54][4] Jan's Hub tab is connected to the Hugging Face model hub: a user can browse suggested models, paste a Hugging Face repository ID into the search bar, choose a quantization level, or start from a model page on Hugging Face through its "Use this model" menu. The Hub labels each download as fitting, possibly slow, or not fitting the detected hardware, and existing GGUF files can be imported by linking to them in place.[54][3] General guidance on choosing models by memory budget is in Best Local and On-Device LLMs.

Local API server

A local API server arrived in version 0.4.5 in January 2024.[27] It listens on 127.0.0.1:1337 by default with the /v1 prefix that OpenAI's API uses, so that software written for the OpenAI API can be pointed at a local model. An API key is optional on the loopback address and mandatory when the server is bound to 0.0.0.0; further settings cover trusted hosts, CORS and whether MCP tools run on the server side, which is off by default.[5] Version 0.7.6 added the Anthropic Messages API format, and 0.7.7 let the server route to remote models as well as local ones.[38][39] A command-line tool, jan, shipped inside the desktop app from 0.7.8 to 0.8.4; as documented for those versions it served a model at localhost:6767/v1 and could start coding agents such as Claude Code against it.[55] Since 0.8.5 the desktop app no longer bundles it, and the separately installed jan command is Jan Agent, a terminal agent that reads and edits files and runs commands in a project folder; the documentation labels it a preview distributed through a nightly channel.[44][93]

Remote providers

Jan is not limited to local inference. Remote OpenAI models were supported from version 0.4.3, and the documentation now lists OpenAI, Anthropic, Google, Groq, Mistral and OpenRouter as built-in providers, with Hugging Face added as a provider in 0.6.6.[26][54][35] Any other endpoint that speaks the OpenAI or Anthropic wire format can be added as a custom provider, which is also how Jan connects to servers such as vLLM, Ollama or a separately run llama.cpp.[56] When a cloud model is used, the conversation is sent to that provider.[57]

MCP, tools and extensions

MCP support appeared behind an experimental switch in June 2025 and became a regular feature in version 0.6.9 that August.[33][36] Jan acts as the MCP host: the user adds servers, and each tool call needs approval in a panel that shows the arguments, unless the user has saved a standing permission.[6] The documentation warns that a compromised tool or a malicious prompt could lead to unintended actions, a risk discussed under prompt injection.[6] Since 0.8.4 web search and page fetching are built-in tools that need no MCP server.[43]

Features inside the app are organised as extensions. The llama.cpp engine, MLX on macOS, downloads, assistants, retrieval-augmented generation and the vector store are built in. Third-party extensions run in sandboxed Web Workers with permissions granted at install time, there is no central extension store, and the documentation states that the sandbox "is not a guarantee against malicious code".[58]

Several agent features have been withdrawn or kept out of stable builds. An integration with the OpenClaw agent shipped in 0.7.8 and was withdrawn in 0.7.9, and "Cowork", a mode in which Jan plans a task, edits files and runs commands, is available only in nightly builds as of the 0.8.8 release.[40][41][59][44]

Jan models

Menlo Research also trains small models intended to run inside the app. All are fine-tunes of Alibaba's Qwen3 family and are published with Apache 2.0 tags. The dates below are the creation dates of the Hugging Face repositories.[49]

ModelRepository createdSizeBase modelStated purpose
Jan-nano2025-06-104BQwen3-4BResearch tasks with MCP tools [60]
Jan-nano-128k2025-06-254BJan-nanoSame, with a native 128k context window [61]
Lucy2025-07-181.7BQwen3-1.7BWeb search on phones and CPU-only machines [62]
Jan-v12025-08-084BQwen3-4B-Thinking-2507Agentic reasoning and tool use in the app [63]
Jan-v1-edge2025-08-271.7BQwen3-1.7B, distilled from Jan-v1On-device web search [64]
Jan-v2-VL (low, med, high)2025-11-068BQwen3-VL-8B-ThinkingVision-language agent for multi-step tasks [65]
Jan-v2-VL-max2025-12-1630BQwen3-VL-30B-A3B-ThinkingLarger version of Jan-v2-VL [66]
Jan-v3-4B-base-instruct2026-01-194BQwen3-4B-Instruct-2507Distilled base model for fine-tuning [67]
Jan-Code-4B2026-03-024BJan-v3-4B-base-instructCoding assistance and sub-agent work [68]
Jan-v3.5-4B2026-03-204BJan-v3-4B-base-instructMathematical reasoning, with a set persona [69]

Expanded article table

The Jan-nano technical report (arXiv, June 28, 2025) describes training with reinforcement learning with verifiable rewards in place of supervised fine-tuning and reports 83.2% on SimpleQA; the Lucy paper (August 1, 2025) reports 78.3% for the 1.7B model, and the Jan-v1 model card reports 91.1%.[70][71][63] These are the developer's own figures, measured with the model connected to web-search tools through MCP, so they are not comparable with SimpleQA scores for models answering from memory.[60] The company's handbook also lists Ichigo, a speech model family, among its work.[72] These are small language models meant for on-device use; Jan 0.7.6 began offering a Jan model during first-run setup.[38]

Platforms and hardware

PlatformRequirements stated by the developerGPU acceleration
WindowsWindows 10 or later; x64 CPU with AVX2, or ARM64; 8 GB RAM minimumNVIDIA through CUDA 13; AMD, Intel Arc and NVIDIA through Vulkan [73][3]
macOSmacOS 13.6 or later on Apple silicon; Intel Macs not supportedMetal, automatic [50][3]
Linuxx64 CPU with AVX2; Flatpak recommended, .deb and AppImage also offeredNVIDIA through CUDA 13; AMD, Intel Arc and NVIDIA through Vulkan [74][3]

Expanded article table

The bundled engine falls back to the CPU when no supported GPU is found, and NVIDIA cards that cannot run CUDA 13 (the GTX 10 series, for example) use Vulkan.[3][44] Jan is also distributed through the Microsoft Store and Flathub.[47] On an Intel Mac, version 0.8.5 and later still open and work with remote providers, but local models do not load; the release notes advise staying on 0.8.4 for that use.[44] Support for AMD's ROCm was added as an experimental backend on Linux in 0.8.2; since the engine was bundled in 0.8.5 the documentation directs AMD users to Vulkan or to a separately run ROCm build of llama.cpp.[75][73][3]

Privacy and telemetry

Jan's privacy page says the app collects nothing until the user agrees ("Zero data collection until you say so"), that the choice is made at first launch and can be changed in Settings, and that with permission it counts active users and retention against a random identifier. The page says chats, files, prompts and model choices are not tracked, and that analytics are processed by PostHog's EU service.[57] The formal policy states that the company "does not collect any data until you explicitly allow tracking" and relies on consent under the GDPR.[7] These are the developer's own statements about its software.

The analytics setup has changed over time. Release notes record the Umami analytics tool being integrated in version 0.4.6 (February 2024) and removed in 0.5.15 (February 2025), and version 0.5.12 introduced the explicit opt-in and opt-out controls.[76][77][30]

Adoption

DateFigureSource
2024-07-301 million downloadsCompany blog [16]
2025-01-0124.7K GitHub stars, up from 0.5K a year earlierRuna Capital ROSS Index [19]
2026-11-17"7.4M+ downloads"jan.ai homepage [78]
2026-11-1745,755 GitHub stars, 3,178 forksGitHub API [1]

Expanded article table

Download totals are self-reported and are not broken down by platform or version; on the retrieval date the homepage showed both the 7.4 million counter and an older "Over 6 million downloads" heading.[78]

Comparison with other local AI tools

The table below sets Jan beside LM Studio, Ollama, GPT4All and Open WebUI. The table records what each project's own license file and documentation say as of November 2026. KoboldCpp, another program in the same family, is covered in its own article.

ToolApplication licenseInterfaceLocal inference engine
JanApache 2.0; source public [46]Desktop app, local API server, separate CLI [2]llama.cpp built in; MLX on Apple silicon (experimental) [3][4]
LM StudioProprietary terms that prohibit modification and reverse engineering; free for home and work use since July 2025 [79][80]Desktop app, local API server, command-line tool [81]llama.cpp; MLX on Apple silicon [81]
OllamaMIT; source public [82]Command line, REST API and, since July 2025, a desktop app [83]llama.cpp listed as supported backend [83]
GPT4AllMIT; source public [84]Desktop app and Python client [85]llama.cpp implementations [85]
Open WebUIOpen WebUI License (BSD-3-Clause plus a branding clause, from v0.6.6); source public, described by its own docs as not OSI-approved [86]Self-hosted web interface [87]None of its own; connects to Ollama or OpenAI-compatible APIs [87]

Expanded article table

Among the desktop applications in the table, Jan and GPT4All publish their source under permissive licenses and LM Studio does not publish the source of its application. Ollama and Open WebUI occupy different positions: the first is mainly a model runner with an API, the second a browser front end that needs a separate engine. The tools are also combined in practice. Jan's documentation shows Ollama as a custom endpoint, and the 0.8.8 release notes mention Open WebUI connecting to Jan's local API.[56][14] Open WebUI's own comparison page, written by that project, presents Jan as a personal desktop tool without multi-user features such as single sign-on or role-based access control, and suggests running the two together.[88]

Security issues, limitations and criticism

Vulnerabilities in Cortex. On April 2, 2025 Snyk Labs published four flaws in Cortex.cpp, then Jan's engine: an arbitrary file write through path traversal (CVE-2025-2446), an out-of-bounds read in the GGUF parser (CVE-2025-2439), a command injection (CVE-2025-2445) and missing cross-site request forgery protection (CVE-2025-2447). The local server had no authentication, so a web page visited by the user could reach it; Snyk chained three of the bugs into a one-click information leak and used the command injection to reach code execution. Snyk reported the issues on February 18, 2025 and, according to SecurityWeek, all were addressed by March 6.[89][90] Jan's changelog for version 0.5.16 refers to security fixes "found in Cortex".[31]

Local API exposure. Later releases kept tightening the same surface. Version 0.8.5 fixed an advisory (GHSA-x6p8-7cp8-c3p6) under which a server bound to 0.0.0.0 accepted any Host header, and 0.8.8 was a dedicated security release that added origin checks for browser requests, path checks on file operations, and limits on the agent's file tools. Its notes also say of the CORS switch: "In earlier versions it was shown but had no effect."[44][14] On October 29, 2026, the day after that release, the project published eleven security advisories on GitHub, two of them rated critical: a one-click remote code execution in versions up to 0.8.7, in which content from a tool or MCP server could navigate the app's main window to an attacker's page, and a code-execution flaw through unsanitized HTML in Model Hub content, which had been fixed in 0.7.6.[94]

Regressions and dropped platforms. The engine change in 0.8.5 was followed a day later by a hotfix, 0.8.6, for an AppImage that crashed on every model load on some Linux distributions, a failure on NVIDIA GTX 16 and RTX 20 series cards with older drivers, and a slowdown with one model family.[53] The 0.8.5 release ended local inference on Intel Macs, which its notes say "have not been supported since v0.8.0", and removed the option to download CUDA 11 or 12 backends.[44] The MLX backend remains experimental, and the 0.8.8 notes list automatic context compaction for remote models as a known gap.[4][14]

Reviews. In a June 2026 review for It's FOSS, Bhuwan Mishra called Jan "a polished, genuinely usable desktop app" and credited the move to Tauri, but returned to Ollama. On a CPU-only AMD laptop he measured 6.62 and 7.35 tokens per second with Ollama against about 5 and 3 with Jan, wrote that "Jan AI provided slow inference time as compared to Ollama on the same hardware", and reported that the application sometimes froze his system.[91] This is one tester's result on one machine. In a Hacker News discussion of Jan in August 2025, one commenter called Jan "basically the OSS version of LM Studio", while others reported that one chat blocked another, that the Tauri interface felt clunky on Linux, and that connecting to an Ollama endpoint failed for them.[92]

Continue reading

More to read

Topics:Developer ToolsLarge Language ModelsOpen Source AI

References

  1. ^1 ^2 ^3 ^4 ^5 ^6GitHub. "janhq/jan" (repository metadata: description, creation date, stars, forks, language, license field). GitHub REST API, retrieved November 17, 2026. api.github.com/...jan
  2. ^1 ^2 ^3Jan. "Overview". Jan documentation. docs.jan.ai/desktop
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Jan. "llama.cpp Engine". Jan documentation. GitHub, janhq/jan. github.com/...llama-cpp.mdx
  4. ^1 ^2 ^3 ^4 ^5Jan. "MLX". Jan documentation. GitHub, janhq/jan. github.com/...mlx.mdx
  5. ^1 ^2Jan. "Local API Server". Jan documentation. GitHub, janhq/jan. github.com/...api-server.mdx
  6. ^1 ^2 ^3Jan. "Model Context Protocol". Jan documentation. GitHub, janhq/jan. github.com/...mcp.mdx
  7. ^1 ^2 ^3Jan. "Jan Privacy Policy" (last updated January 16, 2025). GitHub, janhq/jan. github.com/...privacy-policy.mdx
  8. ^1 ^2Jan. "Homebrew Computer Company" (About page as of tag v0.5.9, November 2024). GitHub, janhq/jan. github.com/...about
  9. ^1 ^2Menlo Research. "Homebrew is now Menlo Research". Menlo Research blog, March 25, 2025. menlo.ai/...homebrew-to-menlo
  10. ^1 ^2Jan. Commit 5c1b0c2, "license". GitHub, janhq/jan, September 6, 2023. github.com/...0c2e2ecd5e402f056decf7e8a56b61300a77
  11. ^1 ^2Jan. Commit 8860734, "Change License and update README". GitHub, janhq/jan, October 15, 2023. github.com/...7345c386062b2137bb2c98e08de5627175d8
  12. ^1 ^2Jan. Commit ec612fc, "Change license to AGPL". GitHub, janhq/jan, October 25, 2023. github.com/...2fc583c01a08a46e6bb049cbda97fa421f02
  13. ^1 ^2Jan. Commit e8ca7f3, "chore: Jan's code is now under the Apache license (#5042)". GitHub, janhq/jan, May 20, 2025. github.com/...7f3c1bc5eee962b49ee46982e83e217ed928
  14. ^1 ^2 ^3 ^4 ^5Jan. "Jan v0.8.8: security fixes, llama.cpp 0.7.0, and Trusted Hosts for the local API". GitHub releases, October 28, 2026. github.com/...v0.8.8
  15. ^1 ^2Jan. "0.2.0". GitHub releases, October 13, 2023. github.com/...v0.2.0
  16. ^1 ^2Menlo Research. "Jan is now part of something bigger: Homebrew". Menlo Research blog, July 30, 2024. menlo.ai/...introducing-homebrew
  17. ^Jan. README as of tag v0.5.14, January 2025. GitHub, janhq/jan. github.com/...README.md
  18. ^Flathub. "Install Jan on Linux" (app listing, ai.jan.Jan). flathub.org/...ai.jan.Jan
  19. ^1 ^2Runa Capital. "Top trending open-source startups in 2024". ROSS Index. runacap.com/...annual-2024
  20. ^Jan. "Who We Are". Jan Team Handbook. GitHub, janhq/jan. github.com/...who-we-are.mdx
  21. ^1 ^2Menlo Research. "What Is Menlo". Menlo Handbook. menlo.ai/...about
  22. ^1 ^2Jan. "0.6.8". GitHub releases, August 14, 2025. github.com/...v0.6.8
  23. ^Jan. "0.7.3". GitHub releases, November 13, 2025. github.com/...v0.7.3
  24. ^Jan. "Releases". GitHub, janhq/jan (release list with publication dates). github.com/...releases
  25. ^Jan. "Changelog". jan.ai. jan.ai/changelog
  26. ^1 ^2Jan. "Faster inference across: Mac, Windows, Linux, and GPUs" (v0.4.3). Jan changelog, December 21, 2023. github.com/...faster-inference-across-platform.mdx
  27. ^1 ^2Jan. "Local API server" (v0.4.5). Jan changelog, January 29, 2024. github.com/...2024-01-29-local-api-server.mdx
  28. ^1 ^2Jan. "0.4.13". GitHub releases, May 20, 2024. github.com/...v0.4.13
  29. ^Jan. "Jan has Stable, Beta and Nightly versions" (v0.5.7). Jan changelog, October 24, 2024. github.com/...2024-10-24-jan-stable.mdx
  30. ^1 ^2Jan. "Jan gives you full control over your privacy" (v0.5.12). Jan changelog, December 30, 2024. github.com/...2024-12-30-jan-new-privacy.mdx
  31. ^1 ^2Jan. "Security fixes and UI improvements" (v0.5.16). Jan changelog, March 14, 2025. github.com/...2025-03-14-jan-security-patch.mdx
  32. ^1 ^2 ^3Jan. "0.6.0". GitHub releases, June 19, 2025. github.com/...v0.6.0
  33. ^1 ^2Jan. "Jan v0.6.3 brings new features and models!". Jan changelog, June 26, 2025. github.com/...2025-06-26-jan-nano-mcp.mdx
  34. ^1 ^2Jan. "0.6.6". GitHub releases, July 31, 2025. github.com/...v0.6.6
  35. ^1 ^2Jan. "Jan v0.6.6: Enhanced llama.cpp integration and smarter model management". Jan changelog, July 31, 2025. github.com/...2025-07-31-llamacpp-tutorials.mdx
  36. ^1 ^2Jan. "Jan v0.6.9: Image support, stable MCP, and powerful model tools". Jan changelog, August 28, 2025. github.com/...2025-08-28-image-support.mdx
  37. ^Jan. "Jan v0.7.0: Jan Projects". Jan changelog, October 2, 2025. github.com/...2025-10-02-jan-projects.mdx
  38. ^1 ^2 ^3Jan. "Jan v0.7.6: Unified Interface & Jan V3 Model". Jan changelog, January 27, 2026. github.com/...-27-jan-unified-ui-anthropic-api.mdx
  39. ^1 ^2 ^3Jan. "Jan v0.7.7: Native MLX Support & UI Refresh". Jan changelog, February 11, 2026. github.com/...2026-02-11-jan-mlx-ui-refresh.mdx
  40. ^1 ^2Jan. "Jan v0.7.8: OpenClaw Integration, Jan CLI & Auto Context Management". Jan changelog, March 11, 2026. github.com/...6-03-11-jan-openclaw-cli-context.mdx
  41. ^1 ^2Jan. "Agents" (OpenClaw integration notice). Jan documentation. GitHub, janhq/jan. github.com/...openclaw.mdx
  42. ^1 ^2 ^3Jan. "Jan v0.8.0: Multi-Token Prediction, llama.cpp Router Mode & Inline MCP Approval". Jan changelog, May 22, 2026. github.com/...2026-05-22-jan-v0.8.0.mdx
  43. ^1 ^2Jan. "Jan v0.8.4: Native Web Search, Per-Model Chat Templates & a Backend Settings Store". Jan changelog, July 21, 2026. github.com/...2026-07-21-jan-v0.8.4.mdx
  44. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Jan. "0.8.5". GitHub releases, October 8, 2026. github.com/...v0.8.5
  45. ^Jan. Commit 87f3b16, "chore: bundle license to app". GitHub, janhq/jan, August 26, 2025. github.com/...b168c9170b8371e736e9113fef9aba7de505
  46. ^1 ^2Menlo Research. "LICENSE". GitHub, janhq/jan. github.com/...LICENSE
  47. ^1 ^2 ^3Jan. "Jan - Open-source ChatGPT replacement" (README). GitHub, janhq/jan. github.com/...README.md
  48. ^1 ^2 ^3Menlo Research. "janhq/cortex.cpp" (archived repository and README). GitHub. github.com/...cortex.cpp
  49. ^1 ^2Hugging Face. Model listings for the "janhq" and "Menlo" organizations (repository creation dates and license tags). Hugging Face Hub API, retrieved November 17, 2026. huggingface.co/...models
  50. ^1 ^2Jan. "Mac" (installation guide). Jan documentation. GitHub, janhq/jan. github.com/...mac.mdx
  51. ^Jan. README as of October 15, 2023 (commit 8860734). GitHub, janhq/jan. github.com/...README.md
  52. ^Jan. "Nitro-Tensorrt-LLM Extension" (v0.4.9). Jan changelog, March 19, 2024. github.com/...-19-nitro-tensorrt-llm-extension.mdx
  53. ^1 ^2Jan. "Jan v0.8.6: hotfix for v0.8.5". GitHub releases, October 9, 2026. github.com/...v0.8.6
  54. ^1 ^2 ^3Jan. "Managing Models". Jan documentation. GitHub, janhq/jan. github.com/...manage-models.mdx
  55. ^Jan. "Jan CLI". Jan documentation. GitHub, janhq/jan. github.com/...cli.mdx
  56. ^1 ^2Jan. "Custom Endpoints". Jan documentation. GitHub, janhq/jan. github.com/...custom-endpoint.mdx
  57. ^1 ^2Jan. "Jan's Privacy Approach". Jan documentation. docs.jan.ai/...privacy
  58. ^Jan. "Extensions". Jan documentation. GitHub, janhq/jan. github.com/...extensions
  59. ^Jan. "Cowork". Jan documentation. GitHub, janhq/jan. github.com/...cowork.mdx
  60. ^1 ^2Menlo Research. "Jan-Nano: An Agentic Model" (model card). Hugging Face. huggingface.co/...Jan-nano
  61. ^Menlo Research. "Jan-Nano-128k: Empowering deeper research through extended context understanding" (model card). Hugging Face. huggingface.co/...Jan-nano-128k
  62. ^Menlo Research. "Lucy: Edgerunning Agentic Web Search on Mobile with a 1.7B model" (model card). Hugging Face. huggingface.co/...Lucy
  63. ^1 ^2Jan. "Jan-v1: Advanced Agentic Language Model" (model card). Hugging Face. huggingface.co/...Jan-v1-4B
  64. ^Jan. "Jan-v1-edge: Distilled for Edge, Built for Web Search" (model card). Hugging Face. huggingface.co/...Jan-v1-edge
  65. ^Jan. "Jan-v2-VL: Multimodal Agent for Long-Horizon Tasks" (model card, Jan-v2-VL-high). Hugging Face. huggingface.co/...Jan-v2-VL-high
  66. ^Jan. "Jan-v2-VL: Multimodal Agent for Long-Horizon Tasks" (model card, Jan-v2-VL-max-FP8). Hugging Face. huggingface.co/...Jan-v2-VL-max-FP8
  67. ^Jan. "Jan-v3-4B-base-instruct: a 4B baseline model for fine-tuning" (model card). Hugging Face. huggingface.co/...Jan-v3-4B-base-instruct
  68. ^Jan. "Jan-Code-4B: a small code-tuned model" (model card). Hugging Face. huggingface.co/...Jan-code-4b
  69. ^Jan. "Jan-v3.5-4B: The first Jan personality" (model card). Hugging Face. huggingface.co/...Jan-v3.5-4B
  70. ^Dao, Alan; Vu, Dinh Bach. "Jan-nano Technical Report". arXiv:2506.22760, June 28, 2025. arxiv.org/...2506.22760
  71. ^Dao, Alan; Vu, Dinh Bach; Nguyen, Alex; Buppodom, Norapat. "Lucy: edgerunning agentic web search on mobile with machine generated task vectors". arXiv:2508.00360, August 1, 2025. arxiv.org/...2508.00360
  72. ^Jan. "What is Jan?". Jan Team Handbook. GitHub, janhq/jan. github.com/...what-jan-is.mdx
  73. ^1 ^2Jan. "Windows" (installation guide). Jan documentation. GitHub, janhq/jan. github.com/...windows.mdx
  74. ^Jan. "Linux" (installation guide). Jan documentation. GitHub, janhq/jan. github.com/...linux.mdx
  75. ^Jan. "Jan v0.8.2: Faster Startup, AMD ROCm/HIP on Linux & Resumable Downloads". Jan changelog, June 1, 2026. github.com/...2026-06-01-jan-v0.8.2.mdx
  76. ^Jan. "0.4.6". GitHub releases, February 5, 2024. github.com/...v0.4.6
  77. ^Jan. "0.5.15". GitHub releases, February 18, 2025. github.com/...v0.5.15
  78. ^1 ^2Jan. "Jan - Open-Source ChatGPT Replacement" (homepage). jan.ai, retrieved November 17, 2026. jan.ai
  79. ^LM Studio. "LM Studio App Terms of Service" (version dated August 23, 2026). lmstudio.ai/app-terms
  80. ^LM Studio. "LM Studio is free for use at work". LM Studio blog, July 8, 2025. lmstudio.ai/...free-for-work
  81. ^1 ^2LM Studio. "Welcome to LM Studio Docs!". lmstudio.ai/...app
  82. ^Ollama. "LICENSE" (MIT License). GitHub, ollama/ollama. github.com/...LICENSE
  83. ^1 ^2Ollama. "README". GitHub, ollama/ollama. github.com/...README.md
  84. ^Nomic. "LICENSE.txt". GitHub, nomic-ai/gpt4all. github.com/...LICENSE.txt
  85. ^1 ^2Nomic. "GPT4All README". GitHub, nomic-ai/gpt4all. github.com/...README.md
  86. ^Open WebUI. "Open WebUI License". Open WebUI documentation. docs.openwebui.com/license
  87. ^1 ^2Open WebUI. "README". GitHub, open-webui/open-webui. github.com/...README.md
  88. ^Open WebUI. "Open WebUI & Jan". Open WebUI documentation. docs.openwebui.com/...jan
  89. ^Onitza-Klugman, Raul. "In Localhost We Trust: Exploring Vulnerabilities in Cortex.cpp, Jan's AI Engine". Snyk Labs, April 2, 2025. labs.snyk.io/...ities-in-cortex-cpp-jans-ai-engine
  90. ^Arghire, Ionut. "Vulnerabilities Expose Jan AI Systems to Remote Manipulation". SecurityWeek, April 2, 2025. securityweek.com/...systems-to-remote-manipulation
  91. ^Mishra, Bhuwan. "I Tried This Open Source ChatGPT Alternative on Linux, But Went Back to Ollama". It's FOSS, June 2, 2026. itsfoss.com/jan-ai
  92. ^Hacker News. "Jan - Ollama alternative with local UI" (discussion thread), August 9, 2025. news.ycombinator.com/item
  93. ^Jan. "Overview". Jan Agent documentation. GitHub, janhq/jan. github.com/...agent
  94. ^Jan. "Security Advisories". GitHub, janhq/jan, retrieved November 17, 2026. github.com/...advisories

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

1 revision · v1 · 4,775 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Cite this page: AI Wiki. "Jan." aiwiki.ai, updated 17 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/jan_ai

Edit article