GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ResonatexIntegratedTechnologies
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
ResonatexIntegratedTechnologies
Model Tree
Input Modalities
Output Modalities
Supported Functionality
ResonateX T1 125M Talking is a compact decoder-only Transformer language model designed for English text generation and conversational experimentation.
Unlike a fine-tune of an existing pretrained language model, the model weights were trained from scratch.
Training follows a progressive curriculum that begins primarily with general language modeling and gradually shifts toward conversational data. The final stage is heavily focused on dialogue.
The project is intended as a lightweight platform for experimenting with:
ResonateX T1 is an experimental research model. Formal benchmark evaluation is still pending.
| ResonateX T1 | |
|---|---|
| Parameters | 125.268M |
| Architecture | Decoder-only Transformer |
| Transformer Layers | 16 |
| Hidden Size | 768 |
| MLP Size | 2,048 |
| Query Heads | 12 |
| KV Heads | 4 |
| Attention | Grouped-Query Attention |
| Vocabulary | 32,004 tokens |
| Training Context | 1,024 tokens |
| Training Tokens | 724.2M |
| Tokens / Parameter | 5.78 |
| Training Steps | 22,100 |
| Primary Language | English |
| Model Format | Safetensors / Transformers |
The model uses Grouped-Query Attention (GQA) with 12 query heads and 4 key/value heads.
Input embeddings and the language-model output projection use tied weights.
ResonateX T1 uses a compact Llama-style causal Transformer architecture.
text
Architecture Decoder-only TransformerParameters 125.268MLayers 16Hidden Size 768Intermediate Size 2048Query Heads 12Key/Value Heads 4Vocabulary 32,004Training Context 1,024Attention GQAObjective Causal Language ModelingThe architecture was selected to provide a balance between model capacity and computational efficiency, making the model practical for experimentation without requiring billion-parameter-scale hardware.
ResonateX T1's model weights were initialized from scratch.
The tokenizer is based on the TinyLlama tokenizer vocabulary, with additional special tokens used by the conversational training format.
No pretrained TinyLlama model weights were used to initialize ResonateX T1.
text
RESONATEX T1 V6E COMPLETE Parameters : 125.268 MTPU workers : 1Steps : 22,100Tokens : 724.2 MTokens/parameter : 5.78Throughput : 53,225 tok/sTotal time : 3.83 hTraining was performed using a Google Cloud TPU v6e environment through Google Colab.
Training used a combination of foundation text and conversational examples.
HuggingFaceTB/smollm-corpus — Cosmopedia v2
Used primarily for general language modeling and development of the model's underlying English language representation.
HuggingFaceH4/ultrachat_200k — train_sft
Used for conversational and instruction-oriented training.
For dialogue examples, the training objective masks non-assistant portions of the conversation so that the loss is primarily applied to the assistant response tokens.
This encourages the model to learn the mapping:
text
context + user message → assistant responserather than simply learning to reproduce every token in the conversation.
Instead of using a fixed mixture throughout training, ResonateX T1 progressively increases its exposure to dialogue.
| Training Progress | Stage | Dialogue Sampling |
|---|---|---|
| 0–45% | FOUNDATION | ~3% |
| 45–65% | LANGUAGE | ~10% |
| 65–80% | MIXED | ~25% |
| 80–92% | CHAT | ~55% |
| 92–100% | DIALOGUE | ~90% |
The majority of early training is devoted to general text.
The goal is to establish basic language modeling capability before strongly specializing the model toward conversation.
Conversational examples become more frequent while foundation training remains dominant.
The model receives a substantially larger proportion of dialogue while continuing to train on general text.
Conversational data becomes the majority of sampled training examples.
The final stage uses approximately 90% dialogue sampling, providing a final conversational specialization pass at a relatively low learning rate.
During the final training run, throughput stabilized at approximately 53K tokens/second after compilation and warm-up.
Selected points from the run:
| Progress | Stage | Tokens | Training Loss | Throughput |
|---|---|---|---|---|
| 1.33% | FOUNDATION | 0.8M | 8.507 | 4,470 tok/s |
| 3.61% | FOUNDATION | 16.4M | 4.013 | 51,314 tok/s |
| 45.00% | LANGUAGE | 320.3M | 2.188 | 53,404 tok/s |
| 82.07% | CHAT | 592.3M | 1.940 | 53,359 tok/s |
| 95.75% | DIALOGUE | 693.0M | 1.866 | 53,338 tok/s |
| 100% | COMPLETE | 724.2M | — | 53,225 tok/s avg. |
Training loss is included as training telemetry and should not be interpreted as a benchmark of downstream model quality.
ResonateX T1 is designed primarily for experimentation with:
Conversational generation
Short English conversations and assistant-style responses.
General text generation
Short-form continuation, descriptions, explanations, and creative text.
Instruction following
Basic instruction-oriented prompts represented by the conversational training distribution.
Research
The relatively small parameter count makes the model useful for experiments where repeatedly training or modifying a multi-billion-parameter model would be unnecessarily expensive.
Install the required packages:
bash
pip install -U transformers torch accelerateLoad the model:
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "ResonatexIntegratedTechnologies/ResonateX-T1-125M-Talking" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype="auto", device_map="auto") model.eval()python
prompt = "The most interesting thing about space is" inputs = tokenizer( prompt, return_tensors="pt").to(model.device) with torch.no_grad(): output = model.generate( **inputs, max_new_tokens=100, do_sample=True, temperature=0.8, top_p=0.9, repetition_penalty=1.1 ) print( tokenizer.decode( output[0], skip_special_tokens=True ))The model received dedicated conversational training.
A simple starting configuration for generation is:
python
generation_config = { "max_new_tokens": 128, "do_sample": True, "temperature": 0.8, "top_p": 0.9, "repetition_penalty": 1.1}These values are starting points, not benchmark-optimized generation parameters.
Generation behavior can change substantially with temperature, nucleus sampling, repetition penalty, prompt format, and context length.
| Parameter | Starting Value |
|---|---|
| Temperature | 0.8 |
| Top P | 0.9 |
| Repetition Penalty | 1.1 |
| Max New Tokens | 128 |
For more deterministic output, try lowering temperature.
For more varied output, try increasing it moderately.
ResonateX T1 has not yet been evaluated on a comprehensive standardized benchmark suite.
For that reason, this model card intentionally does not claim scores for:
Future releases may add reproducible benchmark results.
Until then, the model should be treated as an experimental compact language model rather than a benchmark-validated general-purpose assistant.
ResonateX T1 is a 125M-parameter model trained on approximately 724M tokens.
It has substantially less capacity and training compute than modern billion-parameter language models.
Users should expect limitations including:
The model was trained primarily for English.
Its training context length was 1,024 tokens. Behavior on substantially longer contexts has not been established by this training run.
ResonateX T1 has not undergone comprehensive safety alignment, red-team evaluation, or production safety validation.
Training-time filtering and conversational training should not be interpreted as a complete safety system.
The model may generate:
It should not be relied upon without additional validation for medical, legal, financial, safety-critical, or similarly high-stakes applications.
Applications using the model should implement safeguards appropriate to their use case.
ResonateX T1 is intended to remain an evolving experimental project.
Potential future experiments include:
The original T1 checkpoint can serve as a reference point for comparing future versions.
The Transformers release uses:
text
config.jsongeneration_config.jsonmodel.safetensorstokenizer.jsontokenizer_config.jsonmodel.safetensors contains the canonical inference weights.
Training optimizer states and internal training checkpoints are not required for standard inference.
The canonical ResonateX T1 release uses the Hugging Face Transformers / Safetensors format.
A GGUF conversion can be provided separately for runtimes compatible with the GGUF ecosystem, including LM Studio and llama.cpp-based applications.
GGUF builds should be considered derived inference versions of the canonical Safetensors checkpoint.
ResonateX T1 can be used as a starting checkpoint for further experimentation.
Examples include:
text
ResonateX T1 │ ├── Continued Pretraining │ ├── Conversational SFT │ ├── Domain Specialization │ ├── Personality Tuning │ ├── Preference Optimization │ └── Quantized / GGUF VariantsKeeping the original T1 checkpoint unchanged is recommended when comparing experimental descendants.
This repository is released under the Apache License 2.0.
Please review the license terms as well as any applicable terms associated with the datasets used during training before redistribution or commercial deployment.
If you use ResonateX T1 in research, experiments, or derivative projects, you can reference the model as:
bibtex
@software{resonatex_t1_125m, title = {ResonateX T1 125M Talking}, author = {ResonateX Integrated Technologies}, year = {2026}, description = {A compact 125M-parameter conversational language model trained from scratch}}Small enough to experiment with.
Large enough to learn from.
125.27M parameters · 724.2M tokens · 22,100 steps
Architecture Decoder-only TransformerParameters 125.268MLayers 16Hidden Size 768Intermediate Size 2048Query Heads 12Key/Value Heads 4Vocabulary 32,004Training Context 1,024Attention GQAObjective Causal Language ModelingRESONATEX T1 V6E COMPLETEParameters : 125.268 MTPU workers : 1Steps : 22,100Tokens : 724.2 MTokens/parameter : 5.78Throughput : 53,225 tok/sTotal time : 3.83 himport torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "ResonatexIntegratedTechnologies/ResonateX-T1-125M-Talking" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype="auto", device_map="auto") model.eval()prompt = "The most interesting thing about space is" inputs = tokenizer( prompt, return_tensors="pt").to(model.device) with torch.no_grad(): output = model.generate( **inputs, max_new_tokens=100, do_sample=True, temperature=0.8, top_p=0.9, repetition_penalty=1.1 ) print( tokenizer.decode( output[0], skip_special_tokens=True ))generation_config = { "max_new_tokens": 128, "do_sample": True, "temperature": 0.8, "top_p": 0.9, "repetition_penalty": 1.1}config.jsongeneration_config.jsonmodel.safetensorstokenizer.jsontokenizer_config.jsonResonateX T1 │ ├── Continued Pretraining │ ├── Conversational SFT │ ├── Domain Specialization │ ├── Personality Tuning │ ├── Preference Optimization │ └── Quantized / GGUF Variants@software{resonatex_t1_125m, title = {ResonateX T1 125M Talking}, author = {ResonateX Integrated Technologies}, year = {2026}, description = {A compact 125M-parameter conversational language model trained from scratch}}