Asisten produktivitas Indonesian-first untuk Word, Excel, dan PowerPoint. Berjalan lokal, tanpa mengirim dokumen ke luar mesin.
An Indonesian-first productivity assistant for Word, Excel and PowerPoint. Runs locally, without sending documents off the machine.
Ini adalah profil runtime di atas Qwen3.5-9B (system prompt, template, parameter) — bukan hasil fine-tune. Bobotnya adalah Qwen3.5, Apache 2.0.
This is a runtime profile over Qwen3.5-9B (system prompt, template, parameters) — not a fine-tune. The weights are Qwen3.5, Apache 2.0.
38/40
Indonesian voice
Indonesian voice
19/20
Edit-contract JSON
Edit-contract JSON
0
Jawaban kosong
Empty answers
6,4s
Latency p50
Latency p50
Model ini menghasilkan reasoning panjang bila thinking tidak dimatikan. Endpoint OpenAI-compatible milik Ollama mengabaikan think: false dan chat_template_kwargs.enable_thinking, sehingga jawaban bisa kosong tanpa pesan error.
Terukur pada satu tugas edit dokumen
The model produces long reasoning when thinking is not turned off. Ollama's OpenAI-compatible endpoint ignores think: false and chat_template_kwargs.enable_thinking, so the answer can come back empty with no error.
Measured on one document-edit task
/v1/chat/completions
512s
21.808 karakter reasoning, jawaban kosong.
21,808 characters of reasoning, empty answer.
/api/chat think: false
2s
Jawaban benar.
Correct answer.
curl http://localhost:11434/api/chat -d '{
"model": "ghifidanukusumo/tantular",
"think": false,
"stream": false,
"options": { "temperature": 0.2, "num_predict": 1200 },
"messages": [
{ "role": "user", "content": "Ringkas memo ini dalam tiga butir: ..." }
]
}'
Bila aplikasi Anda memakai bentuk OpenAI, terjemahkan di sisi klien: max_tokens menjadi options.num_predict, dan selalu kirim think: false.
If your application speaks the OpenAI shape, translate client-side: max_tokens becomes options.num_predict, and always send think: false.
ollama pull ghifidanukusumo/tantular
| Tag | Kuantisasi | Ukuran | RAM | Untuk |
|---|---|---|---|---|
| Tag | Quantization | Size | RAM | For |
latest · q8-0.5 | Q8_0 | 10 GB | 16 GB+ | Default |
q4-0.4 | Q4_K_M | 6,6 GB | < 16 GB | Rollback, atau mesin dengan RAM lebih kecil |
lite | Q4_K_M 4B | 3,4 GB | < 16 GB | Untuk RAM di bawah 16 GB |
latest · q8-0.5 | Q8_0 | 10 GB | 16 GB+ | Default |
q4-0.4 | Q4_K_M | 6.6 GB | < 16 GB | Rollback, or a machine with less RAM |
lite | Q4_K_M 4B | 3.4 GB | < 16 GB | For machines under 16 GB RAM |
Diukur pada 60 item internal (40 voice, 20 edit-contract) melalui /api/chat dengan think: false, temperature 0, satu permintaan pada satu waktu, GPU Apple lokal.
Measured on 60 internal items (40 voice, 20 edit-contract) through /api/chat with think: false, temperature 0, one request at a time, on a local Apple GPU.
| Q8_0 — latest | Q4_K_M — q4-0.4 | |
|---|---|---|
| Q8_0 — latest | Q4_K_M — q4-0.4 | |
| Indonesian voice | 38 dari 40 | 37 dari 40 |
| Edit-contract JSON | 19 dari 20 | 18 dari 20 |
| Jawaban kosong | 0 | 0 |
| Latency p50 | 6,4 detik | 3,8 detik |
| Latency p95 | 25,5 detik | 15,3 detik |
| Indonesian voice | 38 of 40 | 37 of 40 |
| Edit-contract JSON | 19 of 20 | 18 of 20 |
| Empty answers | 0 | 0 |
| Latency p50 | 6.4 s | 3.8 s |
| Latency p95 | 25.5 s | 15.3 s |
Q8 menjadi default karena Q4_K_M kehilangan kualitas; Q8 menyamai hasil bobot bf16 pada kedua evaluasi. Q4 lebih cepat dan lebih kecil — pilih Q4 bila kecepatan lebih penting daripada satu-dua item ketelitian.
Angka latency berasal dari satu mesin dan bukan janji kinerja. Permintaan paling lambat pada pengukuran kami mencapai 30 detik.
Q8 is the default because Q4_K_M loses quality; Q8 matches the bf16 weights on both evaluations. Q4 is faster and smaller — pick it when speed matters more than one or two items of accuracy.
The latency figures come from one machine and are not a performance promise. The slowest request in our measurements reached 30 seconds.
01
Pada instruksi multi-edit, model kadang mengeluarkan satu find melebihi batas 200 karakter alih-alih beberapa edit terarah. Terjadi pada Q4, Q8, maupun bf16 — 1 dari 20 item.
02
Evaluasi memakai dokumen sintetis, bukan dokumen Office nyata.
03
Ini profil, bukan fine-tune: kemampuan dasarnya adalah kemampuan Qwen3.5-9B.
01
On multi-edit instructions the model sometimes emits one find past the 200-character limit instead of several targeted edits. Occurs on Q4, Q8 and bf16 alike — 1 of 20 items.
02
The evaluation uses synthetic documents, not real Office documents.
03
This is a profile, not a fine-tune: the underlying capability is Qwen3.5-9B's.
Apache 2.0
Apache 2.0, mengikuti Qwen3.5.
Qwen3.5-9B
Tantular Office adalah system prompt, template, dan parameter di atas bobot itu. Tidak ada bobot yang dilatih.
Apache 2.0
Apache 2.0, following Qwen3.5.
Qwen3.5-9B
Tantular Office is a system prompt, template and parameters over those weights. No weights were trained.
Mulai dengan ollama pull ghifidanukusumo/tantular, atau jelajahi Tantular Office untuk pengalaman AI yang terintegrasi dengan produktivitas sehari-hari.
Indonesian-first. Local-first. Dibangun untuk privasi dan kedaulatan data.
Start with ollama pull ghifidanukusumo/tantular, or explore Tantular Office for AI built into everyday productivity work.
Indonesian-first. Local-first. Built for privacy and data sovereignty.