GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
SmartWhatt
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
SmartWhatt
Model Tree
Qwen/Qwen3-VL-2B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Thai language modelling. Against the released base model, on held-out text the run never trained on:
| measurement | base | this model |
|---|---|---|
| Thai Wikipedia, plain continuation loss | 1.7292 | 1.5334 (−11.3%) |
| FLORES Thai, plain continuation loss | 1.8975 | 1.8321 (−3.4%) |
| held-out loss over the training mixture | 1.5186 | 1.2590 (−17.1%) |
It also beats the base model on every one of the seven training corpora.
This is a base model, not an assistant. The run trained on raw continuation, so instruction following eroded, and every capability benchmark fell:
| benchmark | items | base | this model |
|---|---|---|---|
| Belebele Thai, reading comprehension | 900 | 0.6689 | 0.4311 |
| MMMU validation | 900 | 0.4100 | 0.3011 |
| CV-Bench 2D and 3D | 2,638 | 0.8108 | 0.4215 |
| HumanEval, pass@1 | 164 | 0.2744 | 0.0610 |
Multiple choice was scored by the likelihood of the option letter, not by generated text, so this is not an answer-format effect. The damage was done in the first 1.8B true tokens, and the 1.5B-token cosine decay did not undo it.
theta = base + 0.3 * (this - base).docs/post_training_plan.md of the training
repository.| base | Qwen/Qwen3-VL-2B-Instruct, revision 89644892… |
| mixture | wangchanlion_curated 35%, sea_vl_crawling 18%, cambrian_737k 14%, mammoth_vl_sea 12%, mathpile 10%, opc_code 10%, thai_handwriting 1% capped at 5 epochs |
| stages | 3.0B at 4096 constant, 1.5B annealed, 0.5B at 8192 |
| optimizer | Muon 5e-4 on hidden matrices, AdamW 1e-5 on embeddings and norms |
| hardware | one node, four A100-40GB, about 485 share-hours |
| checkpoint | tokens-5200796360.pt, step 444,960, stage 3 (long context) |
| tokens at export | 5.0B true (the tracker total reads 5,200,796,360, which carries the documented +200M label offset) |
Mixture sampling is token-proportional: a weighted fair queue picks the dataset with the largest token deficit, so the realized mixture converges on effective tokens rather than rows.
model.safetensors holds the parameters in the released class's layout, checked to load with
zero missing and zero unexpected keys. provenance.json records the base revision, the
checkpoint, and the counters at export time.
| evaluation at export | mean held-out loss 1.1414 over 128 rows at the stage's sequence length |