lc little-coder github.com/itayinbarr/little-coder ↗
v1.5.1

lc

a coding agent built for small models

> quick start github repo
// an 8 GB laptop walked onto the official terminal-bench 2.0 leaderboard.

and walked out ranked above
gemini cli + gemini 2.5 pro.

this is the laptop next to you, with the wifi off.

// challenger
qwen3.6-35b-a3b + little-coder
35B-A3B · 3B active · 8 GB laptop GPU · offline · no API
0 7.5 15 22.5 30
0.0% ± 3.2
vs
// frontier cloud agent
Gemini 2.5 Pro + Gemini CLI
frontier-scale · cloud-served · undisclosed params · API metered
0 7.5 15 22.5 30
0.0% ± 2.9
fig. 1 — terminal-bench 2.0 · 89 tasks × 5 trials · official leaderboard led by +5.0 points
// one command, one local model

install.

~/projects · zsh
$
resolving dependencies...
fetching @earendil-works/pi-coding-agent
100%
✓ installed little-coder 1.5.1 · node 22.19+
$
fetching Qwen3.6-35B-A3B · UD-Q4_K_M · 22.1 GB
experts → RAM, attention → GPU · 3.8 GB vram
✓ llama.cpp listening on 127.0.0.1:8888
$
lc little-coder v1.5.1 · offline
Context: 0.4K/32K (1%) · model: llamacpp/qwen3.6-35b-a3b
/help for commands · ctrl-c to quit
>
// a sample session

how it feels.

you write english. it reads, edits, and runs
against your working tree — all done locally.

~/fizzbuzz · little-coder
>
running · llamacpp/qwen3.6-35b-a3b · 3.8 GB vram
→ Glob **/*
✓ → 3 lines (50 chars)
→ Read .docs/instructions.md
✓ → 3 lines (183 chars)
→ Read fizzbuzz.py
✓ → 2 lines (37 chars)
→ Edit fizzbuzz.py
✓ → 11 lines (255 chars)
→ Bash python -m pytest fizzbuzz_test.py -v
✓ → 6 lines (313 chars)
All tests pass.
Context: 4.2K/32K (13%) · ~48 msgs until recommended new session · model: llamacpp/qwen3.6-35b-a3b
// community

talk to us
where the code lives.

no discord, no slack, no forum subscription — just the github issues tab. file a bug, ask how to serve a model on your hardware, and you'll usually get a worked-through answer.

> browse issues ↗ open an issue
0
stars on github
0
issues opened
0
issues resolved
discussions issues pulls
[closed]
does it work with lm studio?
#6 · opened by AtaaEddin · 4 replies · resolved
[closed]
exceeding the thinking budget causes the agent to stop running
#8 · opened by diagonalfish · 3 replies · resolved
[closed]
not starting on windows
#23 · opened by Kaiohshin · 2 replies · resolved
[closed]
how are subagents spawned?
#19 · opened by rahimnathwani · 2 replies · resolved
[closed]
context handling with llama-server
#3 · opened by cmhamiche · 2 replies · resolved