pingi / README.md
lyffseba's picture
Rename model card to PINGI
26474ab verified
|
Raw History Blame Contribute Delete
2.52 kB
metadata
base_model: thinkingmachines/Inkling-2-Mini
license: apache-2.0
tags:
  - pi
  - inkling-2-mini
  - tinker
  - harbor
  - rl
  - agents
  - coding

PINGI — Pi-Native Inkling-2-Mini

PINGI is an ongoing project adapting thinkingmachines/Inkling-2-Mini (26.4B total / 3.4B active / 5.4B effective across its 3-pass looped MoE core, 1M context, multimodal, Apache 2.0) specifically for the pi coding agent harness (read, bash, edit, write).

Links

Current Status

Stage Status Checkpoint / Target
Smoke (pingi-smo) Completed (3 steps, local Docker) Published on Tinker: tinker://e7bba24a-6366-5c8c-93b8-21e624a621a6:train:0/sampler_weights/final (record)
Pilot (pingi-pilot) Next (640 rollouts, scale800 + termgrade, Modal) Pending
MVP / Full (pingi) Planned (1,600–5,120 rollouts + paired held-out gate) Pending

Note: Thinking Machines currently disables raw checkpoint file downloads for thinkingmachines/Inkling-2-Mini (403). Trained checkpoints are published directly via Tinker (publish_checkpoint_from_tinker_path).

Reference Comparison Benchmark

Our primary apples-to-apples target is harness-hub/Qwen3.8-27B-pi-adapted (same pi harness, same scale800-harbor + termgrade-train-harbor dataset mixture, thinking OFF, LoRA r=32):

Model Active / Total Thinking Terminal-Bench 2.1 (pi) SWE-bench Pro hard-51 (pi)
Qwen/Qwen3.8-27B (base) 27B / 27B off 39.5% 39.2%
harness-hub/Qwen3.8-27B-pi-adapted 27B / 27B off 55.8% (+16.3 pts) 64.7% (+25.5 pts)
thinkingmachines/Inkling-2-Mini (official card) 3.4B (5.4B eff.) / 26.4B default 71.6% (pi v0.83.0) 77.9% (mini-swe-agent)
PINGI (lyffseba/pingi) 3.4B (5.4B eff.) / 26.4B none (default) pending held-out pingi-eval pending held-out pingi-eval

Paired held-out scores will be added here only after pingi-eval compare_with_base=True passes the release gate.