AATensorPlay

TensorMind-1.5-preview

README

License: apache-2.0

Overview

TensorMind 1.5 Preview is a lightweight Chinese/English conversational model for local experimentation, instruction-following research, and continued fine-tuning. It ships as a standard Transformers package with Safetensors weights, tokenizer, generation config, and chat template—no custom model code is required.

Highlights

Compact deployment: 536,941,568 parameters in a 1.00 GiB BF16 weight file.
Standard runtime: native Qwen3ForCausalLM support in Transformers.
Chat ready: bundled bilingual chat template with optional non-thinking generation.
Reproducible evaluation: exact score data, vector figures, and rendering source are included.
Open release: Apache-2.0 licensed weights for research and application prototyping.

Quick start

bash
pip install "transformers>=4.51.0" accelerate torch

python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "AATensorPlay/TensorMind-1.5-preview"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "用三句话解释什么是强化学习。"},
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=True,
        temperature=0.7,
        top_p=0.8,
    )

new_tokens = output[0, inputs.input_ids.shape[1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

For deterministic decoding, set do_sample=False and omit temperature and top_p.

Model details

Table with columns: Item, Value
Item	Value
Architecture	`Qwen3ForCausalLM`
Parameters	536,941,568
Precision	BF16
Layers	32
Hidden size	1,024
Intermediate size	4,096
Attention heads / KV heads	16 / 8
Vocabulary	32,768
Position configuration

¹ max_position_embeddings=10240 is the exported configuration ceiling. This preview has not been evaluated on Needle, RULER, LongBench, or another effective long-context benchmark, so it should not be interpreted as a validated long-context quality claim.

Evaluation

Scores use one matched protocol: lm-eval 0.4.12, SGLang 0.5.14, zero-shot evaluation, full datasets, batch size 48, and fixed seeds.

Table with columns: Model, CMMLU, AGIEval-CN, A-CLUE, C-Eval, TMMLU+, Macro
Model	CMMLU	AGIEval-CN	A-CLUE	C-Eval	TMMLU+	Macro
TensorMind 1.5 Preview	24.8834	32.3822	24.7282	23.2541	24.7272	25.9950

These are standalone release scores. Compare them with other models only when the harness, prompts, datasets, and decoding protocol are matched. The exact values are available in benchmark-results.json, with an additional benchmark scorecard.

Training snapshot

Table with columns: Item, Value
Item	Value
Final alignment packing length	512 tokens
Packed train / eval sequences	8,429 / 422
Optimization steps	50
Per-device train / eval batch	28 / 28
Peak learning rate	5e-7
Precision	BF16
Train loss	2.2760
Eval loss, start → finish	2.16935 → 2.16928

Intended use

Good fits include:

local and edge-oriented language-model experiments;
Chinese/English chat prototypes;
instruction-tuning and alignment research;
reproducible inference, quantization, and serving studies.

This preview is not a safety-tuned production assistant, a high-stakes decision system, or a substitute for domain experts. Evaluate it for your task before deployment.

Limitations

At 0.5B scale, the model can hallucinate, miss instructions, and produce incorrect reasoning or arithmetic.
Benchmark scores are protocol-specific and do not establish superiority over other releases.
Effective long-context quality has not been measured; the position setting is a configuration value only.
Safety, multilingual breadth, tool use, and production robustness have not received comprehensive evaluation.

Checksums

Table with columns: File, SHA-256
File	SHA-256
`model.safetensors`	`b99f6c0448831a6cb2012b93793dd72c9e64e94deee15fd8c4d8206c056817d9`
`tokenizer.json`	`2b31db3618982e4b4425535b563b4f12f8b62f82114abf3ef67e363a310bc44b`
`chat_template.jinja`	`ed2c59f23b9eb551f51d870b3a2aa7f9e431d424f8e34f27e4012049e85ba814`
`config.json`	`bf6d70f9520550eec8aef7961b3f240a06b69a46e8c2bd273b8cb18d61a494f1`

License

TensorMind 1.5 Preview is released under the Apache License 2.0.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

AATensorPlay

Model Tree

Base

this model

Input Modalities

Text

Output Modalities