Fast AI coding on your desk.
About 69% on SWE‑bench Verified.

New in 1.7: the desktop app on a Mac, Linux and Windows

Mightling 1.7.0's desktop app holds Ask and Work in one window. Work is a full workbench for agent sessions: a composer with files, skills and commands, a review pane to stage, revert and commit, worktrees, terminal tabs, and a file tree and file tabs beside the thread. On your other computers it is a client of your GB10, which it finds on the local network.

$ sudo apt install ./mightling_1.7.0_arm64.deb

Also in 1.7.0: Night Shift can try a task several times and keep the best attempt (/night add --best-of N), one machine can run several Signal bridges, and a client chooses between two nodes on one network without losing the one it knows. Release notes and downloads.

SWE-bench Verified, measured on one DGX Spark

On 2026-10-08, Mightling 1.5.1's agent with Qwen3.8-27B resolved 68 of 100 SWE-bench Verified tasks on one DGX Spark, offline: a fixed sample of 100 tasks, one attempt per task, default mode. Scaled to all 500 tasks, that is an estimated ≈69% (95% band 63.9–73.4%). This is our own measurement, on our own machine and on a subset, not a leaderboard entry.

Read it as company, not a ranking. On the same 100 tasks Mightling resolved as many as Qwen3-Coder-480B and GLM-4.6, models many times its size, but every row uses a different agent, different hardware and, for the leaderboard rows, the official x86 test images, where ours ran on community arm64 images. In refine mode a first session studies the task without changing anything and a fresh session does it: it resolved one task more than the default (69 against 68, each arm winning tasks the other lost), at about 2.2 times the time per task, so it stays opt-in (ling --refine).

1. Mightling: 100 tasks, one attempt each; one DGX Spark (GB10, 128 GB), model served locally, no cloud; default mode measured 2026-10-08, refine mode 2026-10-09. Estimate from the SWE-bench experiments per-task data.
2. 500 tasks, one attempt; hardware not stated. Source: experiments, 20250805_openhands-Qwen3-Coder-480B-A35B-Instruct. The same model with mini-SWE-agent: 57/100, 55.4%.
3. 500 tasks, one attempt, agent undisclosed; hardware not stated. Source: experiments, 20250930_zai_glm4-6.
4. 500 tasks, one attempt; hardware not stated. Source: experiments, 20260217_mini-v2.0.0_kimi-k2-5-high.
5. 500 tasks, one attempt; hardware not stated. Source: experiments, 20260217_mini-v2.0.0_deepseek-3-2-high.
6. 500 tasks each; hardware not stated. Four submissions: 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%). An earlier mini-SWE-agent run scored 74/100, 74.4%.
7. 500 tasks, one attempt; hardware not stated. Source: experiments, 20251120_livesweagent_gemini-3-pro-preview. The same model with mini-SWE-agent: 72/100, 74.2%.
8. The best published 20–40B model on the leaderboard. 500 tasks; test-time scaling uses two or more attempts per task; hardware not stated. Sources: test-time scaling, single attempt.
9. 500 tasks, one attempt; hardware not stated. Source: experiments, 20251209_mini-v1.17.2_devstral-small-2512.
10. GitHub's own figure, April 2025: "agent mode achieves a pass rate of 56.0% on SWE-bench Verified with Claude 3.7 Sonnet". Number of tasks, attempts and hardware not stated; not on the leaderboard. Source: GitHub blog.
11. Anthropic publishes SWE-bench Verified for its Claude models, run in a simple scaffold of its own (a bash tool and an edit tool), not for the Claude Code product. Sources: Claude Code, Anthropic's SWE-bench method.
12. Cursor publishes results on its own benchmark, Cursor Bench, not SWE-bench Verified. Source: Cursor blog, Composer.

All leaderboard numbers: swebench.com and the SWE-bench/experiments repository; "same 100 tasks" is each submission's published per-task results restricted to the tasks Mightling ran.

Try these first

$ ling "explain this repository to me like I'm new on the team"
$ ling "the tests are failing: find out why and fix them"
$ ling "add input validation to the signup form, with tests"

Or queue work for tonight with /night add <task>: Mightling works through the queue while you sleep, each task on its own git branch, and you review the branches over coffee.

Your code never leaves your desk — and you can prove it

Cloud coding agents send your source code to someone else's servers. Mightling sends it nowhere: the model runs on your GB10, and nothing in Mightling phones home. No usage analytics, no metrics, no sign-in, no "anonymous" crash reports. That makes it usable for code you're not allowed to upload: client work, regulated data, unreleased products.

$ ling-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

Why Mightling

What you get

Mightling and cloud coding agents

Runs on every GB10

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

They share the chip, the 128 GB of unified memory and DGX OS 7, so one installer covers them all. Mightling is developed and tested daily on the ASUS Ascent GX10; the others are covered by tests.

Paper

FAQ