thelamapi/next-270m
5256
1---2language: tr3license: mit4tags:5- turkish6- türkiye7- english8- ai9- lamapi10- gemma311- next12- next-x113- efficient14- text-generation15- open-source16- 1b17- 270m18- finetune19- gguf20- huggingface21- large-language-model22- llm23- causal24- transformer25- artificial-intelligence26- machine-learning27- ai-research28- natural-language-processing29- nlp30- finetuned31- lightweight32- creative33- summarization34- question-answering35- chat-model36- generative-ai37- optimized-model38- unsloth39- trl40- sft41- chemistry42- biology43- finance44- legal45- music46- art47- code48- climate49- medical50- agent51- text-generation-inference52pipeline_tag: text-generation53datasets:54- mlabonne/FineTome-100k55- ITCL/FineTomeOs56- Gryphe/ChatGPT-4o-Writing-Prompts57- dongguanting/ARPO-SFT-54K58- GreenerPastures/All-Your-Base-Full59- Gryphe/Opus-WritingPrompts60- HuggingFaceH4/MATH-50061- mlabonne/smoltalk-flat62- mlabonne/natural_reasoning-formatted63- OpenSPG/KAG-Thinker-training-dataset64- uclanlp/Brief-Pro65- CognitiveKernel/CognitiveKernel-Pro-SFT66- SuperbEmphasis/Claude-4.0-DeepSeek-R1-RP-SFWish67- QuixiAI/dolphin-r168- mlabonne/lmsys-arena-human-sft-55k69library_name: transformers70---71 72<img src='assets/banner.png'>73 74# 🚀 Next-270M (xt330)75 76### *Lightweight, Efficient, and Türkiye-Focused AI*77 78[](https://opensource.org/licenses/MIT)79[]()80[](https://huggingface.co/Lamapi/next-270m)81[](https://discord.gg/XgH4EpyPD2)82 83---84 85<style>86 table { width:fit-content; border-collapse:separate; border-spacing:0 3px;font-family:system-ui, -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, Oxygen, Ubuntu, Cantarell, 'Open Sans', 'Helvetica Neue', sans-serif;background:rgba(15,22,32,0.4);border-radius:16px;padding: 10px; border:none;transition:.2s all ease;}87 thead th { text-align:center; padding:4px 10px; font-size:13px; text-transform:uppercase; color:rgb(200,200,200);border:none; }88 tbody tr { transition: transform 0.18s ease, box-shadow 0.18s ease; border:none !important;transition:.2s all ease;border-radius:16px;background:rgba(0, 0, 0, 0.38);}89 tbody .turkish:hover {box-shadow:0 6px 15px rgba(0, 0, 0, 0.27);scale:1.01;background:rgba(80, 38, 38, 0.6);}90 tbody .next:hover {box-shadow:0 6px 15px rgba(0, 0, 0, 0.27);scale:1.02;background: rgba(0, 59, 225, 1)}91 tbody tr:hover { box-shadow:0 0px 15px rgba(102, 102, 102, 0.13); background:rgba(139, 139, 139, 0.16)}92 td { padding:8px 10px;border:0px transparent !important;outline:transparent !important; text-align:center; }93 td:first-child { font-weight:600;text-align:left }94 /* tbody .turkish td { background: rgba(255, 0, 0, 0.2) !important; color:rgb(200,200,200); font-weight:400;border:0px !important; scale:1.0; } */95 /* tbody .next td { background: rgba(0, 89, 255, 0.49)!important; color:rgb(200,200,200); font-weight:600;border:0px !important; scale:1.00;outline:none;border:none !important;} */96 .next{97 background: rgba(0, 89, 255, 0.49);98 }99 .turkish{100 background:rgba(51, 34, 34, 0.64);101 }102 tbody tr td:first-child { border-top-left-radius:12px; border-bottom-left-radius:12px; }103 tbody tr td:last-child { border-top-right-radius:12px; border-bottom-right-radius:12px; } strong{104 font-size:16px;font-weight:700;105 }106 em{opacity:0.7;font-size:11px !important;}107</style>108## 📖 Overview109 110**Next-270M** is a **270-million parameter causal language model** based on **Gemma 3**, designed for **efficiency, low-resource deployment, and reasoning-focused natural language understanding**.111 112Key highlights:113 114* Extremely **lightweight** — can run on consumer GPUs with low VRAM.115* Optimized for **text reasoning, summarization, and creative generation**.116* Supports **Turkish natively** while remaining multilingual.117* Open-source and transparent for research and applications.118 119Ideal for **developers, students, and organizations** needing **fast, reliable, and low-resource text-generation**.120 121---122 123# Our Next 1B and Next 4B models are leading to all of the tiny models in benchmarks. 124 125<table>126 <thead>127 <tr>128 <th>Model</th>129 <th>MMLU (5-shot) %</th>130 <th>MMLU-Pro %</th>131 <th>GSM8K %</th>132 <th>MATH %</th>133 </tr>134 </thead>135 <tbody>136 <tr class="next">137 <td data-label="Model">Next 4B preview <em>Version s325</em></td>138 <td data-label="MMLU (5-shot) %">84.6</td>139 <td data-label="MMLU-Pro %">66.9</td>140 <td data-label="GSM8K %">82.7</td>141 <td data-label="MATH %"><strong>70.5</strong></td>142 </tr>143 <tr class="next">144 <td data-label="Model">Next 1B <em>Version t327</em></td>145 <td data-label="MMLU (5-shot) %"><strong>87.3</strong></td>146 <td data-label="MMLU-Pro %"><strong>69.2</strong></td>147 <td data-label="GSM8K %"><strong>90.5</strong></td>148 <td data-label="MATH %">70.1</td>149 </tr>150 <tr>151 <td data-label="Model">Qwen 3 0.6B</td>152 <td data-label="MMLU (5-shot) %">52.81</td>153 <td data-label="MMLU-Pro %">37.6</td>154 <td data-label="GSM8K %">60.7</td>155 <td data-label="MATH %">20.5</td>156 </tr>157 <tr>158 <td data-label="Model">Llama 3.2 1B</td>159 <td data-label="MMLU (5-shot) %">49.3</td>160 <td data-label="MMLU-Pro %">44.4</td>161 <td data-label="GSM8K %">11.9</td>162 <td data-label="MATH %">30.6</td>163 </tr>164 <tr class="turkish">165 <td data-label="Model">Kumru 7B <em>not verified</em></td>166 <td data-label="MMLU (5-shot) %">30.7</td>167 <td data-label="MMLU-Pro %">28.6</td>168 <td data-label="GSM8K %">15.38</td>169 <td data-label="MATH %">6.4</td>170 </tr>171 </tbody>172</table>173 174---175 176# Also, our Next Z1 model is leading to state-of-the-art models in some of the Benchmarks.177<table>178 <thead>179 <tr>180 <th>Model</th>181 <th>MMLU (5-shot) %</th>182 <th>MMLU-Pro %</th>183 <th>GSM8K %</th>184 <th>MATH %</th>185 </tr>186 </thead>187 <tbody>188 <tr class="next">189 <td data-label="Model">Next Z1 <em>Version l294</em></td>190 <td data-label="MMLU (5-shot) %"><strong>97.3</strong></td>191 <td data-label="MMLU-Pro %"><strong>94.2</strong></td>192 <td data-label="GSM8K %">97.7</td>193 <td data-label="MATH %">93.2</td>194 </tr>195 <tr class="next">196 <td data-label="Model">Next Z1 <em>Version l294</em> (no tool)</td>197 <td data-label="MMLU (5-shot) %">94.7</td>198 <td data-label="MMLU-Pro %">90.1</td>199 <td data-label="GSM8K %">94.5</td>200 <td data-label="MATH %">88.7</td>201 </tr>202 <tr>203 <td data-label="Model">GPT 5</td>204 <td data-label="MMLU (5-shot) %">92.5</td>205 <td data-label="MMLU-Pro %">87.0</td>206 <td data-label="GSM8K %"><strong>98.4</strong></td>207 <td data-label="MATH %"><strong>96.0</strong></td>208 </tr>209 <tr>210 <td data-label="Model">Claude Opus 4.1 (Thinking)</td>211 <td data-label="MMLU (5-shot) %">~92.0</td>212 <td data-label="MMLU-Pro %">87.8</td>213 <td data-label="GSM8K %">84.7</td>214 <td data-label="MATH %">95.4</td>215 </tr>216 </tbody>217</table>218 219---220 221## 🎯 Goals222 2231. **Lightweight Efficiency:** Run smoothly on low-resource devices.2242. **Reasoning-Focused:** Provide logical and coherent text outputs.2253. **Accessibility:** Fully open-source with clear documentation.2264. **Multilingual Adaptability:** Turkish-focused but supports other languages.227 228---229 230## ✨ Key Features231 232| Feature | Description |233| --------------------------- | --------------------------------------------------------------------- |234| 🔋 Lightweight Architecture | Optimized for low VRAM usage; ideal for small GPUs or CPU deployment. |235| 🇹🇷 Turkish & Multilingual | Handles complex Turkish prompts accurately. |236| 🧠 Reasoning Capabilities | Logical chain-of-thought for question-answering and problem-solving. |237| 📊 Consistent Outputs | Reliable and reproducible results across multiple runs. |238| 🌍 Open Source | Transparent, research-friendly, and community-driven. |239 240---241 242## 📐 Model Specifications243 244| Specification | Details |245| ------------------ | ---------------------------------------------------------------------- |246| Base Model | Gemma 3 |247| Parameter Count | 270 Million |248| Architecture | Transformer, causal LLM |249| Fine-Tuning Method | Instruction fine-tuning (SFT) with Turkish and multilingual datasets |250| Optimizations | Quantization-ready (q8, f16, f32) |251| Use Cases | Text generation, summarization, Q&A, creative writing, reasoning tasks |252 253---254 255## 🚀 Installation & Usage256 257### Use the model:258 259```python260from transformers import AutoTokenizer, AutoModelForCausalLM261import torch262 263model_id = "Lamapi/next-270m"264tokenizer = AutoTokenizer.from_pretrained(model_id)265model = AutoModelForCausalLM.from_pretrained(model_id)266 267# Chat message268messages = [269 {"role": "system", "content": "You are Next-X1, a smart and concise AI assistant trained by Lamapi. Always respond in the user's language. Proudly made in Turkey."},270 {"role": "user", "content": "Hello, how are you?"}271]272 273# Prepare input with Tokenizer274prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)275inputs = tokenizer(prompt, return_tensors="pt")276 277# Output from the model278output = model.generate(**inputs, max_new_tokens=50)279print(tokenizer.decode(output[0], skip_special_tokens=True))280```281 282<div style='width:700px;'>283 <div style='background-color:rgba(0,140,255,0.5);border-radius:16px;border-bottom-right-radius:0px;padding:3px 10px;width:fit-content;max-width:400px;margin-left:250px;margin-top:-15px;margin-bottom:10px;'>284 Hello, how are you?285 </div>286 <div style='background-color:rgba(42,42,40,0.7);border-radius:16px;border-bottom-left-radius:0px;padding:3px 10px;width:fit-content;max-width:400px;'>287 I'm fine, thank you. How are you?288 </div>289</div>290 291---292 293## 📄 License294 295MIT License — free to use, modify, and distribute. Attribution appreciated.296 297---298 299## 📞 Contact & Support300 301* 📧 **Email:** [lamapicontact@gmail.com](mailto:lamapicontact@gmail.com)302* 🤗 **HuggingFace:** [Lamapi](https://huggingface.co/Lamapi)303 304---305 306> **Next-270M** — Lightweight, **efficient, and reasoning-focused**, bringing **Turkey’s AI forward** on low-resource hardware.307 308[](https://huggingface.co/Lamapi)