command-center/omp-docs/pi/models.md

type, title, description, tags, timestamp

type	title	description	tags	timestamp
reference	Pi Models Configuration	Complete reference for models.json — provider definitions, model specs, and API endpoints for the Frank Program.	pi models providers llamacpp unverified	2026-08-28

Pi Models Configuration (models.json)

models.json defines the LLM providers pi connects to. Each provider specifies a base URL, API type, authentication, and available models.

Location: /home/dank/.pi/agent/models.json

---

Providers

Current providers in models.json (verified 2026-08-27, after the 2026-08-18 service swap):

llamacpp (dankserver)

Property	Value
Base URL	http://192.168.0.235:7770/v1
API	openai-completions
Auth	None (EMPTY)
Host	[[dankserver

Models

Model ID	Context Window	maxTokens	Reasoning	Input Types
frank-default-small	262,144	262,144	Yes	Text, image

The server is upstream llama.cpp (b10121) serving HauhauCS Qwen3.6-35B-A3B-Uncensored-Aggressive Q6_K_P with 36 MoE experts on CPU, ngram-mod spec, 4 slots of ~175K context. The server serves the loaded model regardless of the requested model ID. Requests with any model value succeed.

llamacpp (dankclaw)

Property	Value
Base URL	http://192.168.0.143:7770/v1
API	openai-completions
Auth	None (EMPTY)
Host	[[dankclaw

2026-08-22: frank-default-large moved from :52625 to :7770 on Dankclaw (operator request: the primary inference server sits on the standard llama.cpp port). The OpenWebUI quadlet (OPENAI_API_BASE_URL) was re-pointed to :7770 in the same change.

Models

Model ID	Context Window	maxTokens	Reasoning	Input Types
frank-default-large	262,144	262,144	Yes	Text, image

The server is the ROCmFPX llama.cpp runtime serving kingjones777's Qwen3.8-Flash-Next Uncensored ROCmFP4 STRIX_LEAN GGUF — a 177B-parameter, 56B-MoE model sharded into 3 Q4_0-ROCmFP4 GGUF files (98.3 GB total) with a standalone BF16 vision embedding project (0.85 GB). Speculative decoding uses --spec-type ngram-mod because the ROCmFPX quantization drops the 4B MTP head. The model is loaded from /home/dank/models/kingjones-Qwen3.8-Flash-Next-Uncensored-ROCmFP4-STRIX_LEAN/; the original Q4_0-ROCmFP4-STRIX variant remains as a .backup directory. Context is 262144 (native max), 1 slot. Concurrent requests queue on the single slot; the server never preempts a running generation. It exposes reasoning_content in deepseek reasoning format and supports multimodal (vision) inference via the embedded mmproj.

2026-08-27: Upgraded from the original Qwen3.8-Flash-Next Q4_0-ROCmFP4-STRIX GGUF to the Uncensored STRIX_LEAN variant with vision support and ngram-mod speculative decoding.

Thinking Levels

Pi's default thinking level is configured globally in settings.json as "defaultThinkingLevel": "medium". The model entries map Pi's max level to the llama.cpp-supported xhigh wire value. Pi shows max in the thinking selector from the global settings, but sends reasoning_effort: "xhigh" to both local llama.cpp endpoints. thinkingLevelMap can translate a UI level without changing the global default. Verified 2026-07-22: getSupportedThinkingLevels returns off,minimal,low,medium,high,xhigh and the live server accepts reasoning_effort: "xhigh" and returns reasoning_content (verified 2026-08-21).

Mechanism (pi-ai models.js + openai-completions.js):

• getSupportedThinkingLevels: extended levels xhigh/max appear only when thinkingLevelMap defines them (string = supported, null = hidden, omitted = not supported).
• Default request path (no compat.thinkingFormat): sends reasoning_effort = thinkingLevelMap[level] ?? level when supportsReasoningEffort is true (default for this provider).

Token Budgets (maxTokens)

Verified in the pi source, 2026-08-16 (@earendil-works/pi-coding-agent 0.84.1 + pi-ai):

• maxTokens is an optional model field in models.json. When omitted, pi defaults the model object to 16,384 (provider-composer.js).
• The value is metadata. It drives the model-picker "max out" display and length-error recovery (isRecoverableLength).
• Normal chat turns do not send max_tokens. The request builder sends the field only when a per-request option is set (openai-completions.js buildParams: if (options?.maxTokens)). The agent loop passes no such option.
• The only pi calls that send max_tokens are context compaction and branch summarization. They are deliberately budgeted.

Model Selection Strategy

graph LR
    A[User Request] --> B{Complexity?}
    B -->|Large context / long docs| C[frank-default-large<br/>CIRU llama.cpp dankclaw :7770, 262K ctx]
    B -->|Fast / interactive| D[frank-default-small<br/>llama.cpp dankserver :7770, 262K ctx]

The CIRU endpoint (dankclaw) serves the 177B 56B-MoE model with
262144 context, ngram-mod speculative decoding, 1 slot. Concurrent
requests queue on the single slot. The upstream
llama.cpp endpoint (dankserver) serves the 35B A3B MoE model with 36
experts on CPU and 262K context per slot.
## Live Defaults

`settings.json` picks the main session provider and model (verified 2026-08-19):

| Scope | Provider | Model |
|-------|----------|-------|
| Main session | `llamacpp (dankclaw)` | `frank-default-large` |
| Subagent default | `frank-default-small` (llamacpp dankserver) | ~262K ctx per slot, ngram-mod spec |
| Subagent overrides | scout, researcher | `frank-default-small` |
| Subagent overrides | planner, worker, reviewer, oracle, context-builder | `frank-default-large` (llamacpp dankclaw) |

The subagent model IDs (`frank-default-small`, `frank-default-large`) resolve against the new provider set in `models.json`. The current mappings are `max -> xhigh` for both local models; the global and subagent defaults remain `max`.

---

## See Also

- [[settings|settings]] — Main configuration file
- [[dankserver|dankserver]] — Primary llama.cpp server
- [[dankclaw|dankclaw]] — Secondary llama.cpp server