tellama

Tellama models for local Android AI

한국어 안내 · Download and setup · Measurements · Tellama website

Tellama runs language models locally on Android. This repository provides GGUF model files, their licenses, quantization recipes and evaluation records. Start with TM Qwen3 4B Q4_K_M if you want the model used in Tellama's current high-end-device validation. It supports text conversations; Korean and English were included in our checks.

Current status, October 7, 2026: the Tellama 1.3.11 (41) app baseline was frozen as gold-v1.3.11 on October 5. This is a verified reference build, not a Google Play release announcement. Long response waits and a historical intermittent wrong answer remain unresolved. No model weights changed in this documentation update.

Choose a model

The model download size is not the total RAM needed to run it. Tellama also needs memory for the runtime, context and operating system. The measured device is Samsung Galaxy Z Fold6 SM-F956N, Android 16. Other phones, FAST mode and GPU inference are not covered by this validation.

Get started

In Tellama: open Models, choose TM Qwen3 4B Q4_K_M, complete the download and verification, and review the DEEP-mode notice. Then open Chat and start a conversation. The current release scope uses supplied, verified TM models; it does not provide arbitrary model imports or external AI API connections.

With another GGUF runtime: download the exact file below. Use a runtime that supports Qwen3 and its chat template. Other runtimes are outside Tellama's device validation; this repository does not provide a Transformers-format checkpoint.

hf download LaraAI-Labs/tellama \
  models/qwen3-4b-tellama-thinking-q4_k_m.gguf \
  --revision fa04ad644a423cc3ab5b38029a371c456ae7ecb0 \
  --local-dir ./tellama-models

See setup and checksum instructions, including Windows verification and memory troubleshooting. Download the named file rather than the entire repository, which also contains the historical model.

What to expect

Ordinary TM chat inference runs on the device after download. Model downloads, web research and optional AI response reports use network access. When submitting a report, the user reviews and consents to the content sent. See the privacy policy and 한국어 개인정보처리방침.

This distribution uses internal reasoning before displaying its final answer. In a 30-minute Fold6 battery test, visible-answer waits ranged 14.6–199.8 seconds, with a 37.8-second median across 41 uninterrupted observations. These are answer-visibility times, not native first-token latency (TTFT), and not a forecast for every prompt. Short questions can still take a long time.

An earlier signed-app test produced one wrong arithmetic follow-up: 15 → 16 → 18, where 17 was expected. Later passing runs do not establish that this failure is fixed. Check important answers independently. See test conditions and results before treating a small synthetic score as general accuracy.

Exact artifact and provenance

Tellama performed quantization and packaging, without fine-tuning or calibration. Qwen/Alibaba Cloud retains authorship of the base model; this work does not imply Qwen endorsement. Attribution, recipe, artifact manifest and qualification status are available in this repository. Historical qualified metadata describes the recorded model-screening gates, not unconditional app release approval.

Feedback

Use the Community tab for model questions and reproducible issues. Include the model filename, app/runtime version, device and Android version, and a short non-sensitive example with expected and actual results. Do not post private conversations, credentials or device identifiers.

For AI response reports and report deletion requests, contact [email protected]. For deletion, include the report receipt ID instead of resending the conversation.

Documentation update

October 7, 2026: added Korean guidance, pinned download and checksum steps, current app-baseline status, later Fold6 results and known limitations. Existing GGUF files, licenses, recipes and original evaluation records are preserved. Additional performance tuning remains on hold.