LFM3-VL-2B-BrowserAgent (GGUF Q8_0 + mmproj)

LiquidAI LFM3-VL-2B fine-tuned to operate BrowserOS and BrowserOS neo through their real MCP tool surfaces — snapshot → act → verify loops, batched form filling, tab/session management, error recovery, and the request_human_help discipline for logins, one-time codes, captchas, and money approvals. The vision tower is untouched by this text-only fine-tune and ships as the standard mmproj, so the model keeps its image understanding for screenshot-grounded browsing work.

Training data

1,553 training rows (+173 held-out eval), fully synthetic, built from the open-source browseros-ai/BrowserOS repo (tool definitions, the repo's own neo agent skill, and docs) and the live tools/list of a running BrowserOS neo (25 tools). No personal information — the dataset was screened before training. Task families: inbox triage, unsubscribing, social posting, CRM updates, expense filing with receipt upload, flight search, PDF archiving, settings changes, history reopen, tab groups, saved skills, parallel multi-tab research, two-task sessions, next-call decisions from page snapshots, error recovery, and human-help cases.

Recipe

LoRA (rank 8, alpha 10) via mlx-vlm SFT from the bf16 base, lr 2e-4, batch 4, max seq 2048, 1 epoch (389 iters). Held-out eval: val loss 0.102, token accuracy 97.8%. Adapter fused back into the base, converted to GGUF with llama.cpp's LFM3 converter (vision tower exported as mmproj), quantized to Q8_0. Verified in llama.cpp: loads and emits schema-valid BrowserOS neo tool calls from snapshot prompts.

Files

• lfm3-vl-2b-browseragent-1f00-Q8_0.gguf — the language model.
• mmproj-lfm3-vl-2b-browseragent-1f00-F16.gguf — the vision projector.

Notes

• Converted and verified with current llama.cpp. Note that some bundled llama.cpp builds (e.g. LM Studio's current engine pack) cannot load the lfm3_vl arch yet — use a current llama.cpp build, or LM Studio once its engine adds lfm3_vl support.
• Base model: LiquidAI/LFM3-VL-2B (see its card for license terms).