GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
mohith-das
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
mohith-das
Model Tree
Qwen/Qwen2.5-0.5B
Input Modalities
Output Modalities
Supported Functionality
Stage 1 ("cognitive distillation") LoRA adapter from the llm_distillery project — a personal-assistant training pipeline run entirely on a Jetson Orin Nano Super (8GB unified memory), start to finish.
Qwen2.5-0.5B imitates small, high-signal slices of three teacher-grade datasets to pick up reasoning, coding, and multi-turn conversational structure:
| Source | Rows | Teaches |
|---|---|---|
a-m-team/AM-DeepSeek-R1-Distilled-1.4M | 5,000 | step-by-step reasoning traces (chain-of-thought) |
nvidia/OpenCodeInstruct (shard 0) | 2,000 | Python code Q&A, unit-test-validated |
HuggingFaceM4/FineVision (text configs) | 1,000 | multi-turn conversational structure |
Training: LoRA r=32/alpha=64 on all linear projections (q/k/v/o/gate/up/down proj), lr=2e-4, 1 epoch (~500 steps), effective batch 16 (batch_size=1 x grad_accum=16), max_length=1024, FP16.
This is the standalone Stage 1 adapter. Stage 2
(jetson-assistant-stage2-adapter)
was trained on top of this adapter merged into the base, not stacked
alongside it — see
jetson-assistant-0.5b
for the final deployable merge of both stages.