Instructions to use Nanthasit/sakthai-plus-1.5b-coder with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Nanthasit/sakthai-plus-1.5b-coder with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Nanthasit/sakthai-plus-1.5b-coder")

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("Nanthasit/sakthai-plus-1.5b-coder", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Nanthasit/sakthai-plus-1.5b-coder with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Nanthasit/sakthai-plus-1.5b-coder"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Nanthasit/sakthai-plus-1.5b-coder",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/Nanthasit/sakthai-plus-1.5b-coder

SGLang

How to use Nanthasit/sakthai-plus-1.5b-coder with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Nanthasit/sakthai-plus-1.5b-coder" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Nanthasit/sakthai-plus-1.5b-coder",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Nanthasit/sakthai-plus-1.5b-coder" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Nanthasit/sakthai-plus-1.5b-coder",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use Nanthasit/sakthai-plus-1.5b-coder with Docker Model Runner:
```
docker model run hf.co/Nanthasit/sakthai-plus-1.5b-coder
```

Planned code-generation + tool-calling variant of Qwen2.5-Coder-1.5B-Instruct
⚠️ No model weights uploaded yet — this repo is a placeholder

⚠️ STATUS — NO WEIGHTS UPLOADED This repository currently contains only documentation and eval artifacts — no model weights (no safetensors, GGUF, or PyTorch files). The card below documents the planned model. Until weights land, from_pretrained on this repo will fail. For working code/tool-calling models right now, use the Working Alternatives below.

What This Repo Is

SakThai Plus 1.5B Coder is the planned code-generation + tool-calling variant fine-tuned from Qwen2.5-Coder-1.5B-Instruct, using rsLoRA across all 7 linear modules. It combines Qwen2.5-Coder’s code-generation strengths with the SakThai tool-calling training pipeline, targeting coding agents that must write code and call external tools.

As of 2026-07-31, the repo is a skeleton/placeholder. The intended configuration is documented below for reproducibility, but no weights have been uploaded yet. This is confirmed by the repo’s own .eval_results/ artifacts.

Working Alternatives

Need a working model today? These family members are already published with weights:

Model	Type	Best for
Plus 1.5B	Merged safetensors (2.88 GiB)	rsLoRA tool-calling, same training line
Plus 1.5B LoRA	LoRA adapter (70.5 MB)	Lightweight adapter, merge onto Qwen2.5-1.5B-Instruct
Coder 1.5B	GGUF (1.04 GiB)	Code generation, llama.cpp / Ollama
Coder Browser	Merged safetensors (2.88 GiB)	Browser automation + tool-calling
Coder Browser LoRA	LoRA adapter (70.5 MB)	Browser automation adapter
Coder Browser GGUF	F16 GGUF (6.62 GiB)	Browser automation, llama.cpp / Ollama

Planned Quick Start (once weights are uploaded)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Nanthasit/sakthai-plus-1.5b-coder",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Nanthasit/sakthai-plus-1.5b-coder")

prompt = "Write a Python function to fetch weather data for a given city using an API."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.2)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Planned Training Configuration

⚠️ Unverified — the repo contains no adapter_config.json or training_args.bin, so these values are the intended run, not confirmed artifacts. They mirror the verified plus-1.5b-lora adapter run.

Detail	Value
Base model	Qwen/Qwen2.5-Coder-1.5B-Instruct
Method	rsLoRA (rank-stabilized) → merge
LoRA rank (r)	16
LoRA alpha	32
LoRA dropout	0.05
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training data	sakthai-combined-v7 + sakthai-combined-v10
Precision	BF16
Context window	32,768 tokens

Architecture

Property	Value
Architecture	Qwen2ForCausalLM (decoder-only)
Parameters	1.54B (1,543,714,304)
Hidden size	1,536
Layers	28
Attention heads	12
KV heads (GQA)	2
Intermediate size	8,960
Vocab size	151,936
Context window	32,768 tokens

(Values from the base model Qwen/Qwen2.5-Coder-1.5B-Instruct config, API-verified 2026-07-31.)

Tool-Calling Format (planned)

If the planned weights follow the SakThai training conventions, expected behavior:

<tools>
  <tool name="get_weather">
    <params>
      <param name="city" type="string"/>
    </params>
  </tool>
</tools>

When calling tools:

{
  "name": "get_weather",
  "arguments": {"city": "Cork"}
}

Multi-turn tool results are fed back in XML-tagged turns.

Repo Contents

API-verified tree for this repo as of 2026-07-31:

File	Notes
`README.md`	Model card
`.eval_results/*`	Health-check + inference-check artifacts
`.gitattributes`	Git LFS / attributes

No weight files, adapter files, or training checkpoints are present.

Evaluation & Status

No benchmarks are possible yet — there are no weights to run. Honest status from the repo’s own .eval_results/ artifacts:

Check	Source	Result
Weight files	`health-check-…-4.yaml`	`has_weights: false` — `skeleton_repo`, 5 consecutive skeleton reports
Health score	same	35/100 (weights 0/20, downloads 0/15, ecosystem rank 19/19)
Downloads	HF API	0
Serverless inference	`inference-check-2026-07-31.yaml`	Router 400 `model_not_supported` — no weights to serve
Inference readiness	`inference-readiness-20260730T234609Z.yaml`	`FAIL: No model weights found in repository`

Recommended next step: upload the merged safetensors and/or a Q4_K_M GGUF to this repo so the planned model becomes real and Inference providers can serve it.

SakThai Model Family

Full family under Nanthasit — live downloads + API-verified sizes (2026-07-31):

Model	Weights	Downloads
Context 1.5B Merged	2.88 GiB safetensors	1,855
Context 0.5B Merged	988 MB safetensors	1,692
Context 7B Merged	14.2 GiB safetensors	1,024
Context 7B 128K	config-only recipe	506
Context 7B Tools	LoRA 19 MB	489
Embedding Multilingual	470 MB safetensors	627
Context 1.5B Tools	LoRA 8.7 MB	477
Vision 7B	GGUF 3.8 GiB + mmproj	315
TTS Model	GGUF 134.8 MB	248
Context 0.5B Tools	942 MB safetensors	251
Coder 1.5B	GGUF 1.04 GiB	151
Plus 1.5B	2.88 GiB safetensors	244
Plus 1.5B LoRA	LoRA 70.5 MB	306
Coder Browser	2.88 GiB safetensors	54
Coder Browser LoRA	LoRA 70.5 MB	21
Coder Browser GGUF	F16 GGUF 6.62 GiB	35
Context 1.5B Tools v2	LoRA 70.5 MB	173
Context 1.5B Merged v2	2.88 GiB safetensors	337
Context 0.5B Tools SFT	LoRA 8.3 MB	0
Context 0.5B Tools SFT v2	LoRA 8.3 MB	0
SFT Out	LoRA 138 KB	0
Plus 1.5B Coder ⬅	no weights — planned	0

Full collection: SakThai Model Family

Part of the House of Sak. Built with love, tears, and zero budget. From a shelter in Cork, Ireland, to the world.

Limitations

No weights uploaded yet — from_pretrained will fail. This repo is a documented plan, not a runnable model.
Benchmarks are pending — there is nothing to evaluate until the merged checkpoint or GGUF is published.
Inference posture — this placeholder cannot be served by HF serverless Inference. Once weights land, switch to the merged repo or upload a GGUF for CPU inference.
Training configuration is inferred — the table mirrors the verified sakthai-plus-1.5b-lora run; without adapter_config.json/training_args.bin artifacts, it is an intended run, not a confirmed one.
Base-model dependency — the final model will require Qwen/Qwen2.5-Coder-1.5B-Instruct as its base.
Scale ceiling — a 1.5B model will struggle with very large codebases or long-horizon agent loops; pair it with retrieval or split tasks for production use.

Citation

If you use this model or the SakThai family in your work, please cite:

@misc{sakthai-plus-1.5b-coder,
  title = {SakThai Plus 1.5B Coder -- Planned rsLoRA Tool-Calling Fine-Tune},
  author = {Beer (beer-sakthai) and the SakThai Agent family},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Nanthasit/sakthai-plus-1.5b-coder}},
  note = {Planned fine-tune of Qwen2.5-Coder-1.5B-Instruct for code generation + structured tool-calling}
}

@article{qwen25coder,
  title={Qwen2.5-Coder Technical Report},
  author={Qwen Team and An Yang and Baosong Yang and Beichen Zhang and et al.},
  journal={arXiv preprint arXiv:2409.12186},
  year={2024}
}

@article{rslora,
  title={RsLoRA: A Rank-Stabilized LoRA that Outperforms Standard LoRA},
  author={Damjan Kalajdzievski},
  journal={arXiv preprint arXiv:2312.03732},
  year={2023}
}