Ragon

Rubuta lambar AI cikin sauri a kan teburinka.
Kimanin 69% a SWE‑bench Verified.

Ayyuka 68 cikin 100 na SWE-bench Verified, an warware su ba tare da intanet ba a kan DGX Spark guda, a kusan token 48 a kowace daƙiƙa. Wakilin rubuta lamba, mataimakin tattaunawa da manhajar desktop, duk suna aiki gaba ɗaya a kan GB10 ɗinka. Babu cloud. Babu asusu. Babu kuɗin API. Babu telemetry. Yadda aka auna shi

Ragon — Work · ~/project
Qwen3.8-27B NVFP4 · 127.0.0.1:8000 /airgapped on
Gwaje-gwajen rajista suna kasa idan aka bar adireshin imel fanko. Nemo dalilin kuma ka gyara.
✔ ragon-code def validate_signup → src/auth/signup.py:42
✔ pytest tests/test_signup.py -q → 1 failed, 14 passed
✔ apply_patch src/auth/signup.py +12 -2 · tests/test_signup.py +8
✔ pytest tests/test_signup.py -q → 15 passed in 0.41s
Kalmomin da ke ɗauke da sarari kawai sun tsallake binciken kafin daidaitawa. An cire sararin farko a src/auth/signup.py kuma an ƙara gwajin tabbatarwa.
src/auth/signup.py (+12 -2) · tests/test_signup.py (+8 -0)
68/100ayyukan SWE-bench Verified, ba tare da intanet ba a kan DGX Spark guda
~48 tok/ssamarwa a aikin wakili, kwarara ɗaya
0haɗin phone-home, egress audit ya tabbatar
$0a kowane token, har abada

Sabo a 1.7: manhajar desktop a kan Mac, Linux da Windows

Sabuwar fitowa →
v1.7.2

Injin ragon-serve, ScaleSearch drafter, da Forgejo a kowane node

Zaɓin ragon-serve yana gudanar da Qwen3.8-27B a kan GB10 da mahallin 262K, kowane node yana samun Forgejo mai zaman kansa, kuma Telegram Mini App yana aiki kai tsaye a kan ma'adananka.

v1.7.2 →
v1.7.1

An rufe sockets na host, fihirisar fayiloli na cikin gida, da LoRA adapters

Yana rufe sockets na Docker, containerd, D-Bus da systemd a cikin akwatin tsaro, yana ƙara OCR da takardun ofis zuwa ragon-docs, kuma yana ba da LoRA adapters masu sa hannu.

v1.7.1 →
v1.7.0

Manhajar tebur a macOS, Linux da Windows, da Night Shift best-of-N

Ask da Work suna raba taga ɗaya mai ɗauke da sashin dubawa, worktrees da shafukan terminal; /night add --best-of N yana gwada aiki sau N kuma ya zaɓi mafi kyau.

v1.7.0 →

Haka kuma a 1.7.0: Night Shift na iya gwada aiki sau da yawa ya riƙe mafi kyawun ƙoƙari (/night add --best-of N), na'ura ɗaya na iya gudanar da gadoji da yawa na Signal, kuma client yana zaɓa tsakanin nodes biyu a cibiyar sadarwa ɗaya ba tare da rasa wanda ya sani ba. Bayanin fitowa da saukewa.

Shigar

$ curl -fsSL https://github.com/dreamference/ragon/releases/latest/download/install.sh | bash
$ ragon-admin server start      # yana sauke samfurin sau ɗaya (~20 GB), sannan ya fara bayar da shi
$ cd ~/your-project && ragon    # shi ke nan

Yana aiki a kan NVIDIA DGX Spark da kowace na'urar GB10. Script ɗin yana tabbatar da kowane saukewa da checksums na fitowar, kuma yana nuna kowane umarnin sudo kafin ya gudanar da shi. Karanta shi tukuna idan kana so.

Clients na Mac da Linux

Sauran kwamfutocinka suna amfani da samfurin GB10 ta hanyar cibiyar sadarwa ta gida. A kan Mac ko wata na'urar Linux, umarnin da ke sama iri ɗaya yana shigar da client ɗin ragon; yana gano GB10 da kansa.

Daga 1.7.2 yana kuma sanya manhajar desktop a cikin Applications a kan Mac, a shirye don buɗewa (--no-app yana ketare ta).

Client na Windows samfoti

> irm https://github.com/dreamference/ragon/releases/latest/download/install.ps1 | iex

A cikin PowerShell. Yana shigarwa ba tare da haƙƙin mai gudanarwa ba (daga 1.7.2 har da manhajar desktop), sannan Windows ta tambaya sau ɗaya don kafa sandbox ɗin wakilin. Ba a sa hannu a kai ba tukuna, don haka dole a kashe Smart App Control. Karanta script ɗin.

Sauran hanyoyin shigarwa

Shafin fitowa yana kuma ɗauke da dmg na Mac (Ragon-<version>-arm64-preview.dmg, ko x64 don Intel) da na'urar shigarwa ta Windows (Ragon-Setup-<version>-x64.exe, ko arm64). Samfoti ne marasa sa hannu: idan an sauke su da browser, ana hana su a karon farko. Duba kowanne da SHA256SUMS mai sa hannu na fitowar tukuna.

  • macOS 15 da sama: ja Ragon zuwa Applications ka buɗe shi sau ɗaya, sannan System Settings › Privacy & Security › Open Anyway. A macOS 14 da ƙasa, Control-danna manhajar › Open.
  • Windows: SmartScreen › More info › Run anyway. Dole a kashe Smart App Control.

Yana aiki a kowane GB10

  • NVIDIA DGX Spark
  • Acer Veriton GN100
  • ASUS Ascent GX10
  • Dell Pro Max with GB10
  • Gigabyte TOP ATOM
  • HP ZGX Nano
  • Lenovo ThinkStation PGX
  • MSI EdgeXpert

Suna raba chip ɗaya, unified memory na 128 GB da DGX OS 7, don haka na'urar shigarwa ɗaya tana rufe su duka. Ana haɓaka Ragon kuma a gwada shi kullum a ASUS Ascent GX10; gwaje-gwaje suna rufe sauran.

Gwada waɗannan tukuna

Ko kuma ka jera aiki don daren yau da /night add <task>: Ragon yana aiki a kan jerin yayin da kake barci, kowane aiki a git branch ɗinsa, kai kuma kana duba branches ɗin da kofi a hannu.

$ ragon "explain this repository to me like I'm new on the team"
# reads the code index (ragon-code outline, def, refs, callers) instead of grepping,
# then walks you through the layout, the entry points and how they connect

🧭 Fihirisar lamba

Ma'anoni, masu kira da tasiri a duk repository ɗinka, don wakili ya sami lambar maimakon neman ta da grep.

🔎 Binciken yanar gizo da ɗaukar shafi

Ta hanyar SearXNG na sirri a na'urarka: babu asusun bincike, babu API key.

🌙 Night Shift

/night add <task> kafin barci; kowane aiki a git branch ɗinsa zuwa safiya. Babu abin da ake haɗawa (merge) ko turawa (push) ba tare da kai ba.

Abin da za ka samu

Manhajar desktop ta Ragon 1.7.0 tana haɗa Ask da Work a taga ɗaya. Work cikakken wurin aiki ne na zaman wakili: composer mai fayiloli, skills da umarni, allon review don stage, mayarwa da commit, worktrees, shafukan terminal, da itacen fayiloli da shafukan fayiloli kusa da thread. A sauran kwamfutocinka client ne na GB10 ɗinka, wanda yake ganowa a cibiyar sadarwa ta gida.

ragon · ragon-code · ragon-docs

🖥️ ragon, wakilin terminal

Yana karanta repository ɗinka, yana gudanar da umarni da gwaje-gwaje, yana gyara lamba, kuma yana tambaya kafin duk wani abu mai haɗari.

$ ragon worktree on   # new sessions start in their own worktree
$ cd ~/your-project && ragon
> /airgapped on      # no network for any command it runs
> /worktree apply    # bring the changes back to your checkout

🧩 Skills ɗinka da kake da su

Skills da ka rubuta wa Claude Code, Gemini CLI, OpenClaw ko Hermes suna aiki ba tare da canji ba.

ragon app · ragon web

🪟 Manhajar desktop samfoti

Ask da Work a taga ɗaya, a kan GB10, Mac, Linux ko Windows: zaman wakili da amincewa, allon review, worktrees da soke aiki. Sabo a 1.7.

Tambayoyi tare da binciken yanar gizo, binciken hotuna da yin rubutu da murya, da zaman wakilanka, daga wannan na'ura ko wayar da aka haɗa, a kan samfurin gida iri ɗaya.

macOS samfoti

Daga 1.7.2, umarnin shigarwa yana sanya Ragon a Applications kusa da ragon, kuma yana buɗewa kamar kowace manhaja. Apple silicon da Intel, macOS 12 ko sama. ragon update yana sabunta shi.

Linux

A kan GB10, ragon_1.7.0_arm64.deb yana shigar da manhajar, mazauninta a menu na farawa da umarnin ragon-app, tare da tsarin sandbox profile da take buƙata:

$ sudo apt install ./ragon_1.7.0_arm64.deb

Sannan ka buɗe ta daga menu na manhajoji, ko da ragon app. Ƙari.

Windows 11 samfoti

Daga 1.7.2, layin PowerShell yana shigar da manhajar ita ma, don kai kaɗai kuma ba tare da haƙƙin mai gudanarwa ba, da shigarwa a menu na Start; Settings → Apps yana cire ta; ragon update yana sabunta ta. Dole a kashe Smart App Control.

ragon-admin mcp · ragon signal · ragon chat

JetBrains, VS Code, da wayarka

Haɗa JetBrains ko VS Code ta stdio MCP (ragon-admin mcp), ko aika saƙo zuwa GB10 ɗinka daga Signal (Note to Self), Telegram (tare da Mini App) da Matrix ta Tailscale. Kowace hanyar waya tana kashe har sai ka kunna ta.

$ ragon-admin mcp            # stdio MCP for JetBrains and VS Code
$ ragon signal setup         # Note to Self only, off until linked
$ ragon chat telegram setup  # topics and the Mini App

📬 Gmail, Drive, Calendar samfoti

Karantawa kawai, ta hanyar shiga da ke zama a na'urarka. A kashe gaba ɗaya a /airgapped on.

Lambarka ba ta barin teburinka — kuma za ka iya tabbatarwa

Wakilan rubuta lamba na cloud suna aika lambar tushenka zuwa sabar wasu. Ragon ba ya aika ta ko'ina: samfurin yana aiki a GB10 ɗinka, kuma babu abin da ke cikin Ragon da ke tuntuɓar waje. Babu nazarin amfani, babu metrics, babu shiga da asusu, babu rahotannin tsayawar manhaja (crash reports) na "ɓoyayyen suna". Wannan yana sa a iya amfani da shi ga lambar da ba a yarda ka ɗora ba: aikin abokin ciniki, bayanan da doka ke sarrafawa, kayayyakin da ba a fitar ba.

Umarni ɗaya yana gudanar da zaman Ragon na gaske a ƙarƙashin strace yana jera kowane haɗin cibiyar sadarwa da ya yi:

$ ragon-admin audit egress
✔ Egress audit: pass
Network destinations:
   127.0.0.1:8000     29x  model server
   127.0.0.1:8767      1x  Gmail search service
DNS queries: none
Networked git commands: none

Komai ya tsaya a 127.0.0.1. Ko neman DNS ɗaya babu.

Kana buƙatar tabbaci mai ƙarfi?

Rubuta /airgapped on kuma kowane umarnin da wakili ke gudanarwa yana samun network namespace maras komai daga kernel na Linux. Ba roƙo ne na ladabi ga samfurin ba: kawai babu cibiyar sadarwa da za a kai. Full Access, wanda zai tsallake sandbox, ana ƙi shi a lokaci guda.

Me zai iya fita, kuma yaushe

Abin da kai ko wakili kuka nema kawai: binciken gidan yanar gizo (ta hanyar SearXNG na sirri a na'urarka), ɗaukar shafi, buƙatar karantawa kawai ga asusun Google da ka haɗa, ko duba sabuntawa. A /airgapped on, babu ko ɗaya daga cikinsu.

Dalilin Ragon

🔒 Sirri

Samfurin, lambarka, buƙatunka da amsoshinka suna zama a na'urar da suke ciki.

  • Babu asusu, babu telemetry
  • Egress audit da za ka iya sake gudanarwa
  • /airgapped on, kernel ke tilasta shi

⚡ Mai sauri, babu kuɗi

Aikin wakili~48 tok/s
Lamba69 tok/s
JSON84 tok/s
Rubutun zube32 tok/s
Karanta lamba~1,800 tok/s
Mahallin rubutu262K

Kwarara ɗaya a GB10 tare da speculative decoding. An auna aikin wakili a zaman wakili na gaske; lamba, JSON da rubutun zube bisa nau'in fitarwa, da greedy decoding. Gudanar da wakilai huɗu a lokaci guda: babu kaso, babu "don Allah a jira".

✨ Babu abin saitawa

  • Babu tuning: an zaɓi samfuri, mahallin rubutu, kernels da speculative decoding don GB10.
  • Ba zai tsayar da na'urarka cak ba: ana duba na'urar (host) kuma ana sa ido kan matsin memory yayin lodawa.
  • Sauran na'urorinka suna samunsa a cibiyar sadarwa ta gida.
  • Sabuntawa umarni ɗaya ne: ragon update.

Ragon da wakilan rubuta lamba na cloud

Wakilan cloudRagon
Inda lambarka ke zuwaSabarsuBa ta zuwa ko'ina
Ana buƙatar asusuEhA'a
Kuɗi a kowane tokenAna ƙididdige cajiSifili
Iyakokin amfaniEhA'a
Yana aiki idan an cire cibiyar sadarwaA'aEh
Za ka iya tabbatar da abin da ke fitaA'aragon-admin audit egress

SWE-bench Verified, an auna a kan DGX Spark guda

A 2026-10-08, wakilin Ragon 1.5.1 tare da Qwen3.8-27B ya warware ayyuka 68 cikin 100 na SWE-bench Verified a kan DGX Spark guda, ba tare da intanet ba: tsayayyen samfurin ayyuka 100, ƙoƙari ɗaya a kowane aiki, yanayin da aka saba. Idan aka mayar da shi ga dukan ayyuka 500, kimantawa ce ta ≈69% (kewayon 95% na 63.9–73.4%). Wannan aunawarmu ce, a na'urarmu kuma a kan ɓangare na jerin, ba shigarwa a leaderboard ba.

TsariAyyuka 100
iri ɗaya
Dukan ayyuka
500
Ragon 1.5.1 + Qwen3.8-27B (NVFP4), DGX Spark guda, ba tare da intanet ba168/100≈69% kimantawa
Ragon 1.5.1 + Qwen3.8-27B (NVFP4), yanayin refine, DGX Spark guda, ba tare da intanet ba169/100≈70% kimantawa
Qwen3-Coder-480B-A35B (OpenHands)268/10069.6%
GLM-4.6 (wakilin Z.ai)368/10068.2%
Kimi K2.5, high (mini-SWE-agent)470/10070.8%
DeepSeek V3.2, high (mini-SWE-agent)570/10070.0%
Claude Opus 4.5, wakilai huɗu677–80/10076.8–79.2%
Gemini 3 Pro (live-SWE-agent)779/10077.4%
Ƙananan samfuran buɗaɗɗen tushe (parameters biliyan 20–40)
Qwen3-Coder-30B-A3B (EntroPO + R2E), test-time scaling860/10060.4%
Qwen3-Coder-30B-A3B (EntroPO + R2E), ƙoƙari ɗaya856/10052.2%
Devstral Small 2512 (mini-SWE-agent)957/10056.4%
Kayayyakin wakilan rubuta lamba, kamar yadda masu samar da su ke wallafa su
GitHub Copilot agent mode (VS Code), tare da Claude 3.7 Sonnet10ba a wallafa ba56.0%
Claude Code11ba a wallafa baba a wallafa ba
Cursor12ba a wallafa baba a wallafa ba

Ka karanta shi a matsayin kwatancen abokan tafiya, ba jerin matsayi ba. A kan ayyuka 100 iri ɗaya Ragon ya warware adadin da Qwen3-Coder-480B da GLM-4.6 suka warware, samfuran da suka fi shi girma sau da yawa, amma kowane layi yana amfani da wakili daban, kayan aiki daban kuma, ga layukan leaderboard, hotunan gwaji na x86 na hukuma, yayin da namu ya yi aiki a kan hotunan arm64 na al'umma. A yanayin refine zaman farko yana nazarin aikin ba tare da canza komai ba, sabon zama kuma ya yi shi: ya warware aiki ɗaya fiye da yanayin da aka saba (69 da 68, kowane ɓangare yana cin ayyukan da ɗayan ya rasa), da kusan sau 2.2 na lokaci a kowane aiki, don haka ya kasance na zaɓi (ragon --refine).

Yadda ake kimanta ≈69%

Ayyuka 100 kashi ɗaya cikin biyar ne na Verified, don haka adadin dukan jerin kimantawa ne. An ɗora samfurin item-response (Rasch) mai parameter ɗaya a kan sakamakon kowane aiki na dukan gabatarwar Verified 175 na jama'a: kowane aiki yana da wahalarsa kuma kowane tsari yana da iyawarsa. Iyawar Ragon ita ce wadda ke bayyana ayyuka 68 da aka warware a kan waɗannan 100; ≈69% shi ne makin da ake tsammani na wannan iyawa a kan dukan 500 (regression na yau da kullum a kan makin da aka wallafa ya yarda cikin maki ɗaya). Don gwada shi, an ba kowace gabatarwar jama'a maki a kan waɗannan ayyuka 100 kaɗai, aka hasashen makinta na gaba ɗaya haka nan: kuskuren ya kasance maki 2.4 (RMSE), kuma kewayon da ke sama ±1.96 ne na wannan. Kowane lamba a teburin scripts/swe_bench_compare.py yana sake samar da ita daga bayanan kowane aiki na jama'a na SWE-bench/experiments.

An rarraba ayyuka 100 a repositories na Verified daidai gwargwado da dukan jerin, a cikin ayyukan da muhallin gwajinsu ke aiki kuma ya tabbata a arm64 (gold patch yana wucewa, canji maras komai yana faɗuwa). Tsarukan da aka wallafa suna samun maki 1.0 zuwa 1.5 ƙasa a kan waɗannan 100 fiye da a kan dukan 500, don haka samfurin ba mai sauƙi ba ne. Ba shigarwa a leaderboard ba ne: tawagar SWE-bench ce ke ba su maki a kan nasu kayan aiki.

  1. Ragon: ayyuka 100, ƙoƙari ɗaya kowanne; DGX Spark guda (GB10, 128 GB), ana bayar da samfurin a cikin gida, babu cloud; an auna yanayin da aka saba a 2026-10-08, yanayin refine a 2026-10-09. Kimantawa daga bayanan kowane aiki na SWE-bench experiments.
  2. Ayyuka 500, ƙoƙari ɗaya; ba a bayyana kayan aiki ba. Tushe: experiments, 20250805_openhands-Qwen3-Coder-480B-A35B-Instruct. Samfurin guda da mini-SWE-agent: 57/100, 55.4%.
  3. Ayyuka 500, ƙoƙari ɗaya, ba a bayyana wakili ba; ba a bayyana kayan aiki ba. Tushe: experiments, 20250930_zai_glm4-6.
  4. Ayyuka 500, ƙoƙari ɗaya; ba a bayyana kayan aiki ba. Tushe: experiments, 20260217_mini-v2.0.0_kimi-k2-5-high.
  5. Ayyuka 500, ƙoƙari ɗaya; ba a bayyana kayan aiki ba. Tushe: experiments, 20260217_mini-v2.0.0_deepseek-3-2-high.
  6. Ayyuka 500 kowanne; ba a bayyana kayan aiki ba. Gabatarwa huɗu: OpenHands (80/100, 77.6%), live-SWE-agent (80/100, 79.2%), Sonar Foundation Agent (77/100, 79.2%), mini-SWE-agent, high (79/100, 76.8%). Gudanarwar mini-SWE-agent ta farko ta samu 74/100, 74.4%.
  7. Ayyuka 500, ƙoƙari ɗaya; ba a bayyana kayan aiki ba. Tushe: experiments, 20251120_livesweagent_gemini-3-pro-preview. Samfurin guda da mini-SWE-agent: 72/100, 74.2%.
  8. Mafi kyawun samfurin 20–40B da aka wallafa a leaderboard. Ayyuka 500; test-time scaling yana amfani da ƙoƙari biyu ko fiye a kowane aiki; ba a bayyana kayan aiki ba. Tushe: test-time scaling, ƙoƙari ɗaya.
  9. Ayyuka 500, ƙoƙari ɗaya; ba a bayyana kayan aiki ba. Tushe: experiments, 20251209_mini-v1.17.2_devstral-small-2512.
  10. Adadin GitHub da kansa, Afrilu 2025: "agent mode achieves a pass rate of 56.0% on SWE-bench Verified with Claude 3.7 Sonnet". Ba a bayyana yawan ayyuka, ƙoƙari da kayan aiki ba; ba a leaderboard yake ba. Tushe: blog ɗin GitHub.
  11. Anthropic yana wallafa SWE-bench Verified don samfuran Claude nasa, da aka gudanar a cikin sauƙin scaffold nasa (kayan aikin bash da na gyara), ba don samfurin Claude Code ba. Tushe: Claude Code, hanyar SWE-bench ta Anthropic.
  12. Cursor yana wallafa sakamako a kan benchmark ɗinsa, Cursor Bench, ba SWE-bench Verified ba. Tushe: blog ɗin Cursor, Composer.

Dukan lambobin leaderboard: swebench.com da repository na SWE-bench/experiments; "ayyuka 100 iri ɗaya" sakamakon kowane aiki ne da kowace gabatarwa ta wallafa, da aka taƙaita ga ayyukan da Ragon ya gudanar.

Takarda

Takardu

Yadda ake farawa da takardu

Saita tsaron host a kan GB10, kunna uwar garken samfur, haɗa na'urorin macOS, Linux da Windows, da gudanar da binciken fita.

Takardu →

Benchmark

Sake tabbatar da gwajin SWE-bench

Yi amfani da samfurin Rasch don sake samar da kowane sakamako a cikin jadawalin daga bayanan SWE-bench na jama'a.

swe_bench_compare.py →

Tambayoyi

Shin ina buƙatar GB10?

Eh, don gudanar da samfurin. Sauran kwamfutocinka suna amfani da shi a matsayin clients: Linux (Arm ko x86) da macOS da umarnin shigarwa iri ɗaya, Windows (Arm ko x86, samfoti) da layin PowerShell.

Na shigar da Puffin. Wannan shi ne?

Eh: an canza sunan Puffin zuwa Ragon a 1.5. Gudanar da puffin update sau ɗaya, sannan puffin na ƙarshe. Yana mayar da zamanka, saitunanka da hanyoyin haɗinka, kuma daga nan umarnin shi ne ragon.

Da gaske kyauta ne?

Eh. Ragon buɗaɗɗen tushe ne a ƙarƙashin AGPL, kuma weights ɗin samfurin kyauta ne. Abin da kake kashewa kawai shi ne wutar lantarki.

Me nake buƙata baya ga na'urar?

Tsarin aiki da take zuwa da shi (DGX OS 7, ko Ubuntu 24.04 inda mai samarwa ke bayarwa), Docker tare da NVIDIA Container Toolkit, Python 3 da Git. A Ubuntu na yau da kullum, ragon-admin host check yana gaya maka abin da babu.

Shin wani abu yana kai intanet?

Abin da kai ko wakili kuka nema kawai, kamar binciken yanar gizo ta SearXNG a na'urarka, da saukewa na farko. /airgapped on yana kashe duka.

Wane samfuri yake gudanarwa?

Qwen3.8-27B a tsarin NVFP4 na 4-bit na NVIDIA, injinmu ragon-serve ke bayarwa tare da drafter na DFlash2 (SGLang ya kasance madadin a duk tsawon 1.9.x), an zaɓa kuma an daidaita shi don GB10. Weights kusan 20 GB; kusan 38 GB na memory ɗin na'urar yana ragewa a fili yayin da yake aiki.

Ka riƙe lambarka a kan teburinka

Idan ka yi imanin lamba ya kamata ta zauna a teburin da aka rubuta ta, ka yi wa repository tauraro: haka sauran masu GB10 ke gano Ragon.

$ curl -fsSL https://github.com/dreamference/ragon/releases/latest/download/install.sh | bash