How it works
The model answers from its tools and never from memory alone. It runs with six tools:
Table with columns: Tool, What it returns| Tool | What it returns |
|---|
search_docs(query) | The 3 best sections of the public texts of texel.run: the docs, GPU rates, model prices, the roadmap, the changelog, the receipts pages, the blog, and the support notes |
get_credits | The credits of the wallet signed in: the balance, free outputs left today, what a refund would send, and deposits that wait for the chain |
list_instances | The GPU instances of the wallet that run now or ended in the last 24 hours |
list_outputs | The images, videos, and voices the wallet made in the last 24 hours, with their status and price |
list_receipts | The newest receipts of the wallet, and whether each hour is sealed on Robinhood Chain |
open_ticket(kind, summary, details) | Sends a bug report or a complaint to the Texel team, with the wallet attached |
The wallet tools only read. The model cannot buy, refund, rent, or move funds, so it tells the user where to do it. This repo holds the tool schemas (tools.json). Give the model a short system prompt that says what Texel is, the date and time, and whether a wallet is signed in ("No wallet is signed in." or "The user is signed in with the wallet 0x1a2b…9f3e."), as in the example below.
Training
- Data: 960 support chats for training (1,367 examples, 4.3M tokens) and 72 held out for evaluation. Each chat starts from a scenario: a question on the docs, a check of a made-up wallet, a bug report or complaint, a scam or another request to refuse, or small talk. A simulated user (DeepSeek V4.1 Flash) wrote the user's side, and an open-weight teacher (Qwen3.8) answered as Texel Support with the real tools. Of 1,073 chats, 41 were dropped by checks for numbers or links that no tool gave, credits mixed up with dollars, and answers that skipped a tool.
- Sources: only the public pages of texel.run, read as a signed-out visitor on 11 Nov 2026.
- Method: LoRA with rank 32 and alpha 32 on the MLP of all 32 layers and on q, k, v, and o of the 8 full-attention layers, for 58.2M trainable weights (0.61%). bf16, learning rate 2e-4 with a cosine schedule, batches of 16, 1.2 epochs (103 steps), sequences up to 8,192 tokens, thinking off. The loss covers only what the assistant writes after the last user message.
- Run: 55 minutes of training at 1,313 tokens a second, with a peak of 21.3 GB on the A40. The eval loss fell from 0.674 to 0.600. The receipt covers the whole rental (13,681 GPU seconds, 2.2422 credits): setup, training, evaluation, and the time until the instance was ended.
- Software: Unsloth 2026.11.3, Transformers 5.17.0, PEFT 0.21.2, PyTorch 2.14.1, flash-linear-attention 0.6.0.
Evaluation
The 72 held-out chats were played again from their first message with the same tools and made-up wallets. A judge model (DeepSeek V4 Pro) graded each first answer against what the tools returned.
Table with columns: Qwen3.5-9B, Texel Support 9B | Qwen3.5-9B | Texel Support 9B |
|---|
| Correct | 62.5% | 77.8% |
| Partly correct | 22.2% | 12.5% |
| Wrong | 15.3% | 9.7% |
| Calls the same tools as the reference answer | 47.2% | 79.2% |
| Passes the rule checks | 97.2% | 98.6% |
Use
Load the base with AutoModelForImageTextToText, since the adapter's weights sit under its language_model.
import jsonimport torchfrom peft import PeftModelfrom transformers import AutoModelForImageTextToText, AutoTokenizer base = AutoModelForImageTextToText.from_pretrained( "unsloth/Qwen3.5-9B", dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(base, "texellabs/texel-support-9b")tokenizer = AutoTokenizer.from_pretrained("texellabs/texel-support-9b") system = ( "You are Texel Support, the assistant in the chat of texel.run. Texel is the onchain GPU " "cloud on Robinhood Chain: a wallet signs in, buys Texel credits, makes images, video, and " "voice in the Playground, and rents NVIDIA GPUs by the second.\n\n" "Today is Thursday, November 12, 2026, 09:00 UTC. No wallet is signed in.")tools = json.load(open("tools.json"))messages = [ {"role": "system", "content": system}, {"role": "user", "content": "How much is an A40 per hour on Texel?"},]prompt = tokenizer.apply_chat_template( messages, tools=tools, add_generation_prompt=True, enable_thinking=False, tokenize=False)inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)output = model.generate(**inputs, max_new_tokens=600, do_sample=False)print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:]))
The model calls a tool in the XML format of Qwen3.5 (<tool_call><function=search_docs>...). Run the tool, add its result as a tool message, and generate again until the model answers.
Limits
- It writes in English.
- Prices, GPUs, and the roadmap of Texel change, so serve it with
search_docs over the live site, not over a copy.
- In the evaluation, about 1 answer in 10 was still wrong, mostly price units or a guess where the tools had no answer.
- The wallet tools call the gateway of texel.run for the wallet signed in, so this repo has their schemas, not the gateway.
License
Apache 2.0, the license of the base model.
texel.run · X · GitHub