Devotd Github
NewInference & serving

ishekkhal/bitrun v0.2.0-beta — Adds a Terminal Cockpit for 1-Bit Models

ishekkhal/bitrun released v0.2.0-beta, adding a terminal dashboard Cockpit plus real-time throughput and power displays for 1-bit and 1.58-bit ternary models.

Repositoryishekkhal/bitrun ↗

ishekkhal/bitrun released v0.2.0-beta, adding a terminal dashboard Cockpit plus real-time throughput and power displays for 1-bit and 1.58-bit ternary models.

GitHub description: Zero-compile runner and terminal Cockpit for 1-bit / 1.58-bit ternary LLMs on CPU. Real-time wattage telemetry, 70+ t/s on laptops, and 8B model support (macOS, Windows, Linux)

stars
★ 6
Forks
2
Language
Go
License
MIT
Latest release
v0.2.0-beta
Read original source ↗View on GitHub ↗

ishekkhal/bitrun released v0.2.0-beta. This release, contained in the October 15, 2026 GitHub commit 0af499c, is chiefly about attaching a full-screen terminal dashboard Cockpit to the runner that executes 1-bit and 1.58-bit ternary weight (weights with only three possible values: -1, 0, 1) models on the CPU. The change spans 21 files, with 2,760 lines added and 156 lines deleted.

The Cockpit is a terminal UI built with Go and Bubble Tea. The left pane keeps the conversation history across multiple turns, while the right pane displays, in real time, generation speed, prompt processing speed, time to first token (TTFT), context usage, per-P-core and per-E-core utilization, the SIMD engine name, package power, energy per token, cumulative session energy, and the power connection status. The point of this tool is that the inference and power measurement screens are combined in one place.

The 'modern 1-bit engine' named in the commit message comes in as a 1,158-line patch file, patches/0004-bitrun-modern-engine-b1395.patch, applied at build time. The macOS build script was newly added, and the Linux and Windows build, packaging, and install scripts were also modified. According to the README badge the license is MIT, and items that had been excluded in .gitignore — docs/ and personal session records (progress.md, memory.md) — were removed.

The evidence is limited. The Cockpit screen shown in the README is a UI example displaying a generation speed of 73.7 tok/s, package power of 11.7 W, and energy per token of 0.1587 J on the basis of bonsai-1.7b, and in this commit the existing benchmark tables for Apple Silicon, Windows, and Linux were deleted. The 70+ t/s and 8B model support stated in the repository description also cannot have their measurement conditions confirmed from this change alone. v0.2.0-beta is a worthwhile tool for people who want to see local CPU inference with their own eyes, but to judge it on performance the missing benchmarks have to be filled in separately.