File size: 2,523 Bytes
389a167
 
 
 
 
 
 
 
 
 
 
 
 
26474ab
389a167
26474ab
389a167
 
 
 
 
26474ab
 
389a167
 
 
 
 
 
26474ab
 
 
389a167
 
 
 
 
 
 
 
 
 
 
 
 
26474ab
389a167
26474ab
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
---
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`](https://huggingface.co/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`](https://pi.dev/) coding agent harness (`read`, `bash`, `edit`, `write`).

## Links

- Code, training & eval stack: https://github.com/lyffseba/pingi
- Smoke-test checkpoint (`pingi-smo`): https://huggingface.co/lyffseba/pingi-smo
- Companion architecture research (`xai`): https://huggingface.co/spaces/lyffseba/xai · https://huggingface.co/datasets/lyffseba/xai-studies

## 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](https://huggingface.co/lyffseba/pingi-smo)) |
| **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`](https://huggingface.co/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.