---
family: lfm2.5
license: lfm-open-license-v1.0
model_id: lfm2.5-1.2b-thinking-int4
source_url: https://huggingface.co/litert-community/LFM2.5-1.2B-Thinking
task: text-generation
---

# lfm2.5-1.2b-thinking-int4

| | |
|---|---|
| **Task** | text-generation |
| **Family** | lfm2.5 |
| **Source** | https://huggingface.co/litert-community/LFM2.5-1.2B-Thinking |
| **License** | lfm-open-license-v1.0 |

## Conversion

| | |
|---|---|
| **Tool** | litert-torch 0.9.1 |
| **Command** | `python convert_lfm25.py LiquidAI/LFM2.5-1.2B-Thinking out_lfm25_12b_fp --fp && python ../minicpm_work/quantize_litertlm.py apply out_lfm25_12b_fp/model.litertlm lfm25_int4.litertlm --recipe wi4b32_wi8 --algo octav` |
| **Quantization** | int4 blockwise-32 + OCTAV linears, int8 embedding, convs float |

## Artifacts

| File | SHA-256 | Size (MB) |
|---|---|---|
| `LFM2.5-1.2B-Thinking_int4.litertlm` | `f4f091503ef33c9ad70e2a58f9cbad03a71b3d4d9fe4266b684e8889ec371669` | 701.919 |

## Performance

No benchmark data yet.

## Device runs (NPU / LiteRT-LM)

Model-level on-device results — measured behavior of this exact artifact on the recorded device, only meaningful together with that environment. Throughput is conditional on the measured prompt length (Prompt (tokens); '-' = the source stated none): rows differing only there are different measurements, not re-runs. Full records (failure classes, log evidence): `data/device_runs/0.16.0/2026-08-24/lfm2.5-1.2b-thinking-int4__galaxy-s26.json`, `data/device_runs/0.16.1/2026-09-01/lfm2.5-1.2b-thinking-int4__raspberry-pi-5.json`, `data/device_runs/0.17.0/2026-09-06/lfm2.5-1.2b-thinking-int4__mac-studio-m4-max.json`.

| Device | Accelerator | Status | Full delegation | Output match | Prompt (tokens) | Latency p50 (ms) | Prefill tok/s | Decode tok/s | TTFT (ms) | Peak mem (MB) | Environment | Date | Provenance |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| galaxy-s26 | gpu | load_failed | no | - | - | - | - | - | - | - | Galaxy S26 (SM-S942Q) · Qualcomm SM8850 · litert-lm 0.16.0 · Android 16 | 2026-08-24 | measured |
| mac-studio-m4-max | cpu | pass | - | - | - | - | - | - | - | - | Mac Studio (M4 Max) · litert-lm 0.17.0 | 2026-09-06 | measured |
| raspberry-pi-5 | cpu | pass | - | - | 256 | - | 54.83 | 9.25 | 4776.8 | 1450.0 | Raspberry Pi 5 Model B Rev 1.1 · Broadcom BCM2712 · litert-lm 0.16.1 · Linux-6.18.34+rpt-rpi-2712-aarch64-with-glibc2.41 | 2026-09-01 | measured |

## Pitfalls

- This artifact cannot use a GPU delegate. It is litert-torch 0.9.1 lineage, whose ShortConv patch emits GATHER_ND and INT64 ops that GPU delegates reject; the delegate takes 536 of 579 operations and engine creation then aborts. The count was measured on the Instruct sibling of the same lineage (Pixel 8a and Galaxy S26, litert-lm v0.16.0); this file has not been separately gated. The repo also ships LFM2.5-1.2B-Thinking_int4_gpu.litertlm, a litert-torch 0.9.3 re-export that delegates fully — that file is a separate card.
- The 0.9.1 exporter needs the ShortConv prefill-pad fix that convert_lfm25.py applies: the stock block saves its conv state from the padded columns of a prefill chunk, corrupting the first generated token of nearly every reply. It is easy to miss — the model recovers after about one token and GSM8K still parses answers, it just loses roughly 20 points.
- Quantize convs at export time only. Post-hoc ALL_SUPPORTED int8 through ai-edge-quantizer kills the conv layers (no output); post-hoc recipes must stay on linears and the embedding (wi8fc, wi4b32_wi8).
- litert-lm >= 0.15 needs an ExecutorMetadata section for this hybrid: files exported before that run on 0.14 but fail at inference on 0.15 with 'missing some output TensorBuffers'. The published files were repaired in place on 2026-08-04.
- GSM8K (greedy, 0-shot CoT, n=100, max-tokens 2048 — a thinking model needs the budget) is 72 for this int4 file against a bf16 reference of 81; the int8 sibling scores 77. The reasoning stream arrives on the thought channel and the final answer follows after </think>, so a harness that reads the raw stream must strip it.

---

_Rendered by `edge-card` from `card.json` (the source of truth); edit the inputs, not this file._