mohitkrconnect/heed-tasks-qwen3-0.6b
Text Generation • 0.6B • Updated • 83
Heed: on-device voice notes to tasks
Sub-1B models that turn voice notes into task lists on an Android phone, offline. 0.6B = 80% of Claude Sonnet on a sealed test.
Note The app model. Sealed test: 75.6 F1 = 80% of Claude Sonnet, +20 over GPT-5.4 mini; 5.1 s per note on a phone CPU.
Heed - 80% of Sonnet on a phone
🎙How a 0.6B on-device model reached 80% of Claude Sonnet
Note Blog: how a 0.6B phone model reached 80% of Claude Sonnet, and how the LLM judge was checked against a human.
mohitkrconnect/heed-tasks-qwen3-0.6b-seed0
Text Generation • 0.6B • Updated • 62Note Same recipe, seed 0: shows seed noise (77.5 vs 79.4 validation).
mohitkrconnect/heed-tasks-qwen3-0.6b-exp054
Text Generation • 0.6B • Updated • 60Note Previous app model, before targeted data. Trained as written + plain text.
mohitkrconnect/heed-tasks-only-qwen3.5-0.8b
Text Generation • 0.8B • Updated • 62Note Not shipped: best early tasks-only model, but no official LiteRT-LM converter.
mohitkrconnect/heed-tasks-only-qwen3-0.6b
Text Generation • 0.6B • Updated • 56Note First model on the phone (tasks only, older labels).
mohitkrconnect/heed-full-output-json-qwen3-0.6b
Text Generation • 0.6B • Updated • 48Note One model writes task, bucket, time and reminder (JSON). Reminders cost 8-10 points on real notes, so they were dropped.
mohitkrconnect/heed-full-output-lines-qwen3-0.6b
Text Generation • 0.6B • Updated • 47Note Negative result: a shorter lines format, 30% fewer tokens but 2 points worse than JSON.
mohitkrconnect/heed-full-output-json-4ep-qwen3-0.6b
Text Generation • 0.6B • Updated • 48Note The paired JSON baseline for the lines-format test.
mohitkrconnect/heed-tasks-only-long-qwen3-0.6b
Text Generation • 0.6B • Updated • 44Note Tasks only, one consistent label set + 2,697 long notes: real notes 60.9 -> 81.7.
mohitkrconnect/heed-simnotes-e2-qwen3.5-0.8b
Text Generation • 0.8B • Updated • 38Note Negative result: 1,109 simulated notes in my style gave no gain beyond seed noise.
mohitkrconnect/heed-simnotes-soup-qwen3.5-0.8b
Text Generation • 0.8B • Updated • 36Note A weight average of two 0.8B runs: fewest errors per note, same share fully right.
mohitkrconnect/heed-voice-notes-to-tasks
Viewer • Updated • 8.76k • 44Note Synthetic notes + labels + every system's answers and judge verdicts on the public test notes, with a scoring script.