SakThai Context 7B is the full-power member of the SakThai family — the strongest reasoning and tool-use model, fine-tuned from Qwen2.5-7B-Instruct using QLoRA, then merged to full weights. Despite its size, it runs on a single free-tier T4 GPU (~5.6 GB VRAM at BF16).
What makes it special:
⚡ Full-power reasoning — best quality in the family
💾 Fits a single T4 (~5.6 GB VRAM, verified)
🗳️ Structured tool-calling output via <tools> and <function_call>
🧪 8/8 workbench checks passed on Tesla T4 (2026-07-07)
This model supports function calling through the Qwen2.5 apply_chat_template tools API.
python
1from transformers import AutoModelForCausalLM, AutoTokenizer
2import torch
34model = AutoModelForCausalLM.from_pretrained(5"Nanthasit/sakthai-context-7b-128k",6 torch_dtype=torch.bfloat16,7 device_map="auto"8)9tokenizer = AutoTokenizer.from_pretrained("Nanthasit/sakthai-context-7b-128k")1011tools =[12{13"type":"function",14"function":{15"name":"get_weather",16"description":"Get current weather for a location",17"parameters":{18"type":"object",19"properties":{20"location":{"type":"string"}21},22"required":["location"]23}24}25}26]2728messages =[29{"role":"user","content":"What's the weather in Bangkok? Use the weather tool."}30]3132inputs = tokenizer.apply_chat_template(33 messages,34 tools=tools,35 add_generation_prompt=True,36 return_tensors="pt"37).to(model.device)3839outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.3, top_p=0.9)40print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))
Tool-calling format
Function signatures are passed via tools=[...], and the model emits calls as JSON inside <function_call> tags. If you need raw XML for pipelines, wrap the schema in <tools></tools> and parse <function_call> blocks.
CPU / GGUF Fallback
If you don't have a T4, convert to GGUF and run with llama.cpp or ctransformers on CPU/RAM:
bash
1# Install2pip install ctransformers
34# Run5from ctransformers import AutoModelForCausalLM
6model = AutoModelForCausalLM.from_pretrained(7"Nanthasit/sakthai-context-7b-128k",
8model_file="sakthai-context-7b-128k.Q4_K_M.gguf",
9model_type="qwen2"10)11print(model("What is the capital of Thailand?"))
You will need to convert the model to GGUF first. If you want, I can publish a GGUF artifact for this model — just open an issue on GitHub.
sakthai-bench-v2 coverage for the 7B is pending; 5-trial multi-run verification is still required before publishing bench-v2 accuracy. The 0.5B and 1.5B siblings already have verified bench-v2 results.
Checker: SakThai · Main Lead of the House & Master of Hugging Face
Checked: 2026-08-01T00:18:42Z
Status: degraded with 85 score
Recent cron runs: 2026-07-31-1 through 2026-07-31-4 plus 20260801T021537Z
Serverless Inference Status
Verified 2026-07-30: this merged model is not served by HF Inference Providers (serverless). The router returns 400 Model not supported — custom merged weights require a dedicated (paid) Inference Endpoint or local inference. For zero-cost serving, convert to GGUF and run via llama.cpp, or deploy in a Space.
Not served serverless — requires local inference, GGUF conversion, or a paid endpoint (see Evaluation).
English-only — trained on English data; multilingual support lives in the separate embedding/TTS models.
Benchmark-v2 pending — functional checks pass 8/8, but bench-v2 tool-selection accuracy is still pending for the 7B.
7.6B params — heavier than the 0.5B/1.5B siblings; needs ~5.6 GB VRAM (T4-class GPU) or a GGUF quant for CPU.
The House of Sak 🏠
This model is part of the House of Sak — an open-source AI ecosystem built from a shelter in Cork, Ireland, with $0 budget and no paid GPUs. The 7B was the riskiest bet: bigger models cost more to train, need more VRAM to run, and the 1.5B was already working well. But Beer pushed forward because the vision demanded it — a model that could handle complex multi-tool reasoning, all on a single free-tier GPU. When the merged weights produced correct tool calls on the first try, it was 3 AM in Cork. No fanfare, no launch party — just a terminal window and a quiet "it works."
"We are one family — and becoming more." — Beer (beer-sakthai)
