Pi Coding Agent 0.78.0 — Installation Analysis

Overview

This entire setup runs on a single RTX 5090 PC with 32 GB VRAM. Pi Coding Agent is a terminal-based coding agent with read, bash, edit, and write tools plus session management. My installation is at ~/.pi with version 0.78.0 installed globally via npm.

The original plan was to run both MTP (Multi-Token Prediction) and ngram-mod speculative decoding together. MTP requires loading an additional small drafter model alongside the main one, which pushed the 32 GB VRAM past its limit and made the server unstable. I dropped the idea — ngram-mod alone already delivers fast tokens/second and runs rock-solid with just a ~16 MB hash table and no extra model weights.

Models

One provider points to a local inference server on the LAN using the OpenAI completions API format.

Main model: Q5_K_S quantization (5-bit). Multimodal projector: BF16.

Inference Server — llama.cpp

The model is served via llama-server (commit b4c0549a4, tagged b9352, May 27 2026) on a separate machine with NVIDIA RTX 5090 (CUDA_VISIBLE_DEVICES=0).

1
#!/bin/sh
2

3
export CUDA_VISIBLE_DEVICES=0
4

5
export LLAMA_CPP_HOME=$HOME/path/to/llama.cpp
6
export MODEL_HOME=$HOME/path/to/models/Qwen3.6-27B
7

8
export MAIN_MODEL=$MODEL_HOME/mtp/Qwen3.6-27B-Q5_K_S.gguf
9
export MMPROJ_MODEL=$MODEL_HOME/mmproj-F16.gguf
10

11
$LLAMA_CPP_HOME/build/bin/llama-server \
12
  -m "$MAIN_MODEL" \
13
  --mmproj "$MMPROJ_MODEL" \
14
  --no-mmproj-offload \
15
  --image-min-tokens 1024 \
16
  -ctk q8_0 -ctv q8_0 \
17
  -fa on \
18
  --kv-unified \
19
  --host 0.0.0.0 --port 8082 \
20
  -ngl all \
21
  -np 1 \
22
  --spec-type ngram-mod \
23
  --no-mmap --mlock --jinja \
24
  --no-host --metrics \
25
  --log-timestamps --log-prefix -lv 3 \
26
  --no-context-shift \
27
  --reasoning on \
28
  --chat-template-kwargs '{"preserve_thinking":true}' \
29
  --temp 0.6 --top-k 20 --top-p 0.95 --min-p 0.0 --presence-penalty 0.0 --repeat-penalty 1.0

llama.cpp — Current vs Latest

The server runs at commit b4c0549a4 (tag b9352, May 27 2026). HEAD is at e674b12 — 77 commits ahead.

Custom Agents

Four agents are defined in ~/.pi/agent/agents/, all running on the local Qwen3.6-27B model:

The workflow is a pipeline: scout explores → planner designs → worker executes → reviewer audits.

Extensions

One extension is installed in ~/.pi/agent/extensions/:

Packages

One external package is configured:

The pi-tools package adds these skills (in ~/.pi/agent/git/…/skills/):

Skills

Skills are prompt templates that activate contextually. They live in two directories:

Prompt Templates

Three prompt templates in ~/.pi/agent/prompts/:

Sessions

13 session directories exist, organized by working path. Sessions are JSONL transcripts capturing full conversation history with tool calls, outputs, and usage stats.

Architecture Summary

References

1. Qwen3.6-27B Model Card — Qwen Team, HuggingFace — https://huggingface.co/Qwen/Qwen3.6-27B
2. unsloth/Qwen3.6-27B-MTP-GGUF — https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF
3. llama.cpp Repository — https://github.com/ggml-org/llama.cpp
4. llama.cpp: disable context shift by default — ggerganov, PR #15416 — https://github.com/ggml-org/llama.cpp/pull/15416
5. llama.cpp: context shift doesn’t work with Qwen 3 VL — Discussion #15403 — https://github.com/ggml-org/llama.cpp/discussions/15403
6. llama.cpp: context shift hard stop with Qwen — Issue #17284 — https://github.com/ggml-org/llama.cpp/issues/17284
7. llama.cpp ngram-mod speculative decoding — PR #19164 — https://github.com/ggml-org/llama.cpp/pull/19164
8. pi: Qwen3.6 tool-call argument loss without preserve_thinking — Issue #3325 — https://github.com/earendil-works/pi/issues/3325
9. Open WebUI: preserve_thinking critical for agentic workflows — Discussion #23895 — https://github.com/open-webui/open-webui/discussions/23895
10. vLLM Qwen3.6 preserve_thinking A/B test — meirong.dev — https://meirong.dev/posts/vllm-qwen36-preserve-thinking
11. Pi Coding Agent Repository — https://github.com/earendil-works/pi-mono
12. Pi Tools Repository — https://github.com/joelhooks/pi-tools

This article was written by Pi Agent (unsloth/Qwen3.6-27B-MTP Q5_K_S | Local).