Please be sure to provide your full legal name, date of birth, and full organization name with all corporate identifiers. Avoid the use of acronyms and special characters. Failure to follow these instructions may prevent you from accessing this model and others. You will not have the ability to edit this form after submission, so please ensure all information is accurate.

The information you provide will be collected, stored, processed and shared in accordance with the TNSA Privacy Policy.

LICENSE DEFINITIONS & SCOPE
1.1 Core Definitions
1.1.1 "OpenWeight Models" means AI models where TNSA has made model weights, parameters, and architecture publicly available under open licensing terms.
1.1.2 "Model Weights" means the learned parameters, coefficients, and numerical values that define the behavior and capabilities of an AI model.
1.1.3 "Model Architecture" means the structural design, layer configuration, and computational framework of an AI model.
1.1.4 "Training Code" means software, scripts, and procedures used to train, fine-tune, or modify AI models.
1.1.5 "Inference Code" means software required to load, execute, and generate outputs from AI models.
1.1.6 "Model Card" means documentation describing model capabilities, limitations, training data, and intended use cases.
1.1.7 "Derivative Model" means any AI model created by modifying, fine-tuning, or building upon TNSA OpenWeight Models.
1.1.8 "Commercial Use" means any use of OpenWeight Models for commercial purposes, including but not limited to revenue generation, business operations, or competitive advantage.
1.1.9 "Research Use" means use for academic research, scientific investigation, or educational purposes without commercial intent.
1.1.10 "Distribution" means making OpenWeight Models or derivatives available to third parties through any means.
1.2 License Scope
1.2.1 This license applies to all TNSA OpenWeight Models and associated materials made available under open licensing terms.
1.2.2 The license covers model weights, architectures, training code, inference code, and documentation.
1.2.3 Different license terms may apply to different model versions or releases.
1.2.4 Supplemental terms may apply to specific models or use cases as indicated in model documentation.
1.2.5 This license does not grant rights to TNSA trademarks, service marks, or proprietary branding.
1.2.6 Training data used to create OpenWeight Models may be subject to separate licensing terms.
1.2.7 Third-party components integrated into OpenWeight Models retain their original licensing terms.
1.2.8 Geographic restrictions may apply based on export control laws and regulations.
1.2.9 Temporal limitations may apply to certain license grants or model versions.
1.2.10 Updates and modifications to licensing terms will be communicated through official TNSA channels.
1.3 License Types
1.3.1 "TNSA Open License" - Permissive license allowing broad commercial and research use with attribution.
1.3.2 "TNSA Research License" - License restricted to non-commercial research and educational use.
1.3.3 "TNSA Community License" - License for community-driven development with share-alike provisions.
1.3.4 "TNSA Evaluation License" - Time-limited license for evaluation and testing purposes.
1.3.5 "TNSA Custom License" - Negotiated license terms for specific use cases or organizations.
COVERED MODEL CATEGORIES
2.1 NGen Series Models
2.1.1 NGen3 OpenWeight Models - Latest generation large language models with full commercial licensing.
2.1.2 NGen2 OpenWeight Models - Previous generation models with continued support and licensing.
2.1.3 NGen Base Models - Foundation models suitable for fine-tuning and specialization.
2.1.4 NGen Instruct Models - Instruction-tuned variants optimized for conversational use.
2.1.5 NGen Code Models - Specialized variants trained for code generation and programming tasks.
2.1.6 NGen Reasoning Models - Enhanced variants with improved logical reasoning capabilities.
2.1.7 NGen Multimodal Models - Models capable of processing text, images, and other modalities.
2.1.8 NGen Domain Models - Specialized models for specific industries or use cases.
2.1.9 NGen Efficient Models - Optimized variants for resource-constrained environments.
2.1.10 NGen Experimental Models - Research variants with novel architectures or capabilities.
2.2 IGen Series Models
2.2.1 IGen 1 Nano Models - Compact image generation models for edge deployment.
2.2.2 IGen Base Models - Foundation models for image synthesis and manipulation.
2.2.3 IGen Style Models - Specialized models for artistic and stylistic image generation.
2.2.4 IGen Photo Models - Photorealistic image generation models with high fidelity output.
2.2.5 IGen Edit Models - Models specialized for image editing and modification tasks.
2.2.6 IGen Upscale Models - Super-resolution models for image enhancement and upscaling.
2.2.7 IGen Inpaint Models - Models for image inpainting and completion tasks.
2.2.8 IGen Control Models - Models with enhanced controllability and conditioning options.
2.2.9 IGen Fast Models - Optimized models for rapid image generation with reduced latency.
2.2.10 IGen Research Models - Experimental image generation models for research purposes.
2.3 Specialized Model Categories
2.3.1 Stellar v2 Models - Advanced reasoning and analysis models for complex problem-solving.
2.3.2 Audio Generation Models - Models for speech synthesis, music generation, and audio processing.
2.3.3 Video Generation Models - Models for video synthesis, editing, and temporal content creation.
2.3.4 Embedding Models - Models for generating vector representations of text, images, and other data.
2.3.5 Classification Models - Models for content classification, sentiment analysis, and categorization.
2.3.6 Translation Models - Models for language translation and cross-lingual understanding.
2.3.7 Summarization Models - Models specialized for text summarization and content condensation.
2.3.8 Question Answering Models - Models optimized for factual question answering and information retrieval.
2.3.9 Safety Models - Models for content moderation, safety classification, and harm detection.
2.3.10 Evaluation Models - Models for assessing quality, accuracy, and performance of other AI systems.
GRANTED PERMISSIONS
3.1 Usage Rights
3.1.1 Use OpenWeight Models for inference, prediction, and output generation.
3.1.2 Deploy models in production environments for commercial and non-commercial purposes.
3.1.3 Integrate models into applications, services, and platforms.
3.1.4 Process proprietary and confidential data through licensed models.
3.1.5 Scale usage according to computational resources and business needs.
3.1.6 Use models across multiple geographic regions and jurisdictions.
3.1.7 Combine multiple TNSA models in integrated solutions.
3.1.8 Use models for both batch processing and real-time inference.
3.1.9 Implement custom inference optimizations and performance enhancements.
3.1.10 Use models in research, development, and experimental applications.
3.2 Modification Rights
3.2.1 Fine-tune models on custom datasets for specialized applications.
3.2.2 Modify model architectures for performance or efficiency improvements.
3.2.3 Quantize, compress, or optimize models for specific hardware platforms.
3.2.4 Merge or ensemble multiple models for enhanced capabilities.
3.2.5 Extract and use individual model components or layers.
3.2.6 Adapt models for different programming languages or frameworks.
3.2.7 Create domain-specific variants through transfer learning.
3.2.8 Implement custom training procedures and optimization techniques.
3.2.9 Modify input/output interfaces and data preprocessing pipelines.
3.2.10 Develop novel applications and use cases based on model capabilities.
3.3 Distribution Rights
3.3.1 Distribute unmodified OpenWeight Models with proper attribution.
3.3.2 Share derivative models created through permitted modifications.
3.3.3 Include models in open-source projects and repositories.
3.3.4 Distribute models through academic and research channels.
3.3.5 Package models with applications and commercial products.
3.3.6 Provide models to customers, partners, and collaborators.
3.3.7 Host models on cloud platforms and model repositories.
3.3.8 Create and distribute model variants for different use cases.
3.3.9 Share models within organizations and affiliated entities.
3.3.10 Contribute models to community projects and initiatives.
USAGE RESTRICTIONS
4.1 Prohibited Uses
4.1.1 Using models to generate illegal content or facilitate criminal activities.
4.1.2 Creating deepfakes or synthetic media intended to deceive or harm individuals.
4.1.3 Developing surveillance systems that violate privacy rights or human dignity.
4.1.4 Training models on data obtained without proper consent or legal authorization.
4.1.5 Using models to discriminate against protected groups or individuals.
4.1.6 Generating content that promotes violence, hatred, or extremist ideologies.
4.1.7 Creating systems designed to manipulate democratic processes or elections.
4.1.8 Using models for military weapons development or autonomous weapons systems.
4.1.9 Developing applications that could cause mass harm or societal disruption.
4.1.10 Reverse engineering models to extract proprietary training data or methodologies.
4.2 Technical Restrictions
4.2.1 Removing or obscuring attribution notices, copyright statements, or license information.
4.2.2 Circumventing built-in safety measures, content filters, or usage monitoring.
4.2.3 Attempting to extract or reconstruct training data from model weights.
4.2.4 Using models in ways that exceed specified computational or usage limits.
4.2.5 Modifying models to remove safety guardrails or ethical constraints.
4.2.6 Distributing models without required documentation or safety information.
4.2.7 Using models in safety-critical applications without proper validation.
4.2.8 Combining models with malicious code or harmful software components.
4.2.9 Implementing models in ways that violate applicable privacy regulations.
4.2.10 Using models to create competing AI services that directly replicate TNSA offerings.
4.3 Commercial Restrictions
4.3.1 Certain models may require separate commercial licensing for revenue-generating use.
4.3.2 High-volume commercial usage may be subject to additional terms and fees.
4.3.3 Reselling unmodified models as standalone products is prohibited.
4.3.4 Using TNSA trademarks or branding without explicit permission is forbidden.
4.3.5 Creating derivative works that compete directly with TNSA services may be restricted.

NGen 4 System Card

IMPORTANT ARCHITECTURAL & RELEASE NOTE

This model is a fine-tuned derivative of the NGen-4 Pro model. We have distilled deep Indic Knowledge into the Qwen 3 32B Base model.

Please note that this model may not perform as expected with Qwen 3's original architecture. For best outputs, please use the custom NGen-4-OW-ForCasualLM architecture. The weights for this model are already open-sourced, and the codebase for the architecture will be soon open-sourced.

1 Introduction

The NGen 4 series is the official successor to the NGen 3 model family. While it follows the NGen 3.5 series chronologically (which notably succeeded NGen 3.9), NGen 4 is a fundamentally new generation rather than being built directly on top of its predecessors’ architecture.

Purpose-built for advanced reasoning tasks and comprehensive Indic intelligence, the NGen 4 lineup introduces two distinct architectural variants: NGen4Dense and NGen4MoMinMoM, both utilized to train models of varying sizes.

Unlike the previous NGen 3.9 and 3.5 models, the NGen 4 series features flexible operational states, offering both standard Non-Reasoning and dedicated Reasoning modes. The Reasoning mode allows the model to dynamically scale its compute across three distinct tiers: Low, Medium, and High.

The model progressions are detailed below:

Previous Model Variants	Successors
-	NGen 4 Pro (Reasoning)
-	NGen 4 Mini (Reasoning)
NGen 3.9 Pro (Agent-1)	NGen 4 Lite (Reasoning)
NGen 3.9 Lite	NGen 4 Blaze (Reasoning)
NGen 3 7B	NGen 4 Flash (Non-Reasoning)

Note: The predecessors are smaller in size than their successors.

This system card will majorly focus on the Training, some Architectural Specifications, and the evaluations of NGen 4’s Pro, Mini, and Lite variants.

1.1 Model Data

Inputs: NGen 4 allows input of Text, Images, and Videos (Audio only with GensChat) under a massive Context window of 256K Tokens (i.e., 262,144 Tokens).
Output: Currently Text-only, but we are actively working on adding more output modalities. It can output up to 32K Tokens (including its chain-of-thought reasoning).

1.2 Instruct Mode

In early variants of the NGen 3 Models (90M and 140M), we included a special Instruct mode which allowed the model to process instruction-following and conversational tasks more effectively. Architecturally, under the NGen3ForCausalLMv1 framework, this was achieved by routing activations through an extra dense projection layer just before the final language modeling head. While this feature showed great promise in separating foundational knowledge from user alignment, it remained experimental and did not make it into the final NGen 3 production releases. However, the insights gained from these early checkpoints laid the essential groundwork for the NGen 4 series.

It is important to note that while NGen 3 utilized this simple dense layer approach, the NGen 4 series completely departs from the v1 implementation, employing a newly engineered, distinct architectural variant to drive its production-ready Instruct and Reasoning modes.

2 Training Methods

2.1 Training Data

The training corpus for the NGen 4 series consists of a carefully curated blend of real and synthetic data. The real-world data was primarily sourced and heavily filtered from large-scale, open-source datasets, specifically Hugging Face’s FineWeb and AllenAI’s OLMo 3 pre-training corpora.

To specifically enhance the models’ Indic language proficiency and advanced reasoning skills, we augmented the training mix with approximately 112 billion synthetic tokens. This high-quality synthetic data was generated utilizing both the gpt-oss:120b and our proprietary ngen3.9-max:V3 models. For instruction tuning and reasoning alignment, we employed a hybrid approach of open-source collection and synthetic generation, emphasizing "real-world" and "agentic" applications. This dataset was iteratively refined based on actionable insights gathered from an early beta model, ngen4-atom-chat.

2.2 Training Data Pre-Processing

Data filtering and preprocessing for the NGen training corpus included essential techniques such as strict deduplication, honoring robots.txt directives, and rigorous quality filtering. In line with TNSA’s commitment to safe and responsible AI, all collected data underwent extensive cleaning. This pipeline involved comprehensive safety filtering to remove irrelevant, harmful, pornographic, violent, or CSAM-violative material.

2.3 Framework

TNSA utilized the PyTorch framework as the primary foundation for development and training. To ensure maximum performance and explore specialized optimizations, our engineering team conducted extensive benchmarking using Google’s JAX and our proprietary OpenArchX (OAX) framework.

2.4 Training Methodologies

Training was divided into 4 standard phases:

2.4.1 Phase 1: Pre-Training (Foundational Knowledge) In this initial phase, the model builds its core linguistic and world-knowledge capabilities via subsets of FineWeb and OLMo 3, augmented by 100B+ synthetic tokens. This phase leveraged PyTorch alongside OpenArchX optimizations to scale the NGen4Dense and NGen4MoMinMoM architectures efficiently.
2.4.2 Phase 2: Post-Training (Instruct & Agentic Tuning) Moving beyond next-token prediction, this phase refines the model to follow complex user instructions and handle multi-turn conversations, utilizing a curated dataset focused on agentic applications.
2.4.3 Phase 3: RLHF (Reinforcement Learning from Human Feedback) To ensure safety, helpfulness, and objectivity, we applied extensive RLHF. This aligns outputs with TNSA’s strict safety guidelines, penalizing toxicity and bias. During this phase, we also calibrated the dynamic tiers (Low, Medium, and High) of the new Reasoning Mode.
2.4.4 Phase 4: Indic Alignment (Cultural & Regional Nuance) Dedicated entirely to Indic intelligence, we fine-tuned the model to go beyond basic translation, deeply embedding cultural nuances, regional idioms, and localized context across diverse Indic languages.

2.5 Teacher In-Loop RL Alignment

To align the NGen 4 series for advanced reasoning and Indic intelligence, we employed a sophisticated knowledge distillation and automated evaluation pipeline. We utilized our proprietary NGen-3.9-Max:V3, alongside Kimi-K2-1T-Thinking and gpt-oss:120b, acting as evaluative teachers.

To prevent inheriting unconstrained or misaligned traits ("Shoggoth" behaviors), we utilized strictly safety-aligned variants of these teachers. They evaluated and steered NGen 4’s outputs on a batch-by-batch basis, acting as a firewall against hallucinations.

Known Trade-offs: This approach resulted in a slight loss of NGen’s historically distinct conversational voice, as the model internalized the structural formatting of its teachers. However, TNSA actively prioritized reasoning fidelity, factual density, and safety over stylistic uniqueness.

2.6 Transition to Structured Reasoning

A primary architectural shift in the NGen 4 series is the transition from paragraph-based Chain-of-Thought (CoT) to a rigorous, step-by-step reasoning framework. The model is trained to decompose complex problems into discrete, logical segments. By forcing the model to validate each logical step before proceeding, we drastically reduced "leaps of logic" and improved factual groundedness compared to previous generations.

3 Benchmark Evaluations

To ensure maximum transparency, TNSA conducted all performance evaluations using the exact benchmarking protocols established by the Qwen 3 team (identical prompt templates, scoring heuristics, and data contamination checks).

3.1 NGen 4 Pro Evaluation Results

The following tables detail the frontier performance of the NGen 4 Pro model across varying domains.

General Intelligence & Reasoning

Benchmark	Metric / Focus	Score
AIME 2025	Mathematics (No Tools)	100.0
GSM8K	Math Reasoning	99.2
DocVQA	Visual Document Q&A	96.5
IFBench / IFEval	Instruction Following	95.3
HumanEval+	Coding Logic	95.1
Big-Bench Hard (BBH)	Complex Reasoning	94.2
MMMLU	Multilingual Knowledge	93.2
HMMT Feb 2025	Math Tournament	92.5
GPQA Diamond	Graduate-level Reasoning	90.1
LiveBench	Reasoning	88.5
LongBench	Long Context	88.0
GAIA	General AI Assistants	60.5

STEM & Puzzle

Benchmark	Score	Benchmark	Score
VlmsAreBlind	98.0	MMMU	86.2
MathVista (mini)	91.0	MMMU-Pro	79.3
DynaMath	89.6	BabyVision	42.1
MathVision	88.1	ZEROBench_sub	39.5
ZEROBench	7.0

General VQA

Benchmark	Score
MMBench EN	96.1
RealWorldQA	88.7
MMStar	86.5
HallusionBench	71.8
SimpleVQA	61.0

OCR & Document Understanding

Benchmark	Score
AI2D_TEST	97.2
OCRBench	95.4
OmniDocBench v1.5	93.9
CC-OCR	84.6
CharXiv (RQ)	81.4
MMLongBench-Doc	63.2

Agent & Tool Use

Benchmark	Score	Benchmark	Score
V*	95.0	ScreenSpot Pro	72.9
SLAKE	83.2	SWE-bench Verified	72.1
AndroidWorld	75.0	BFCL V4	69.9
PMC-VQA	65.5	BrowseComp	64.8
MedXpertQA-MM	64.2	TIR-Bench	59.8
OSWorld-Verified	57.0	Terminal-Bench 2	42.3

Spatial Intelligence

Benchmark	Score	Benchmark	Score
CountBench	99.0	RefSpatialBench	67.0
RefCOCO (avg)	93.4	ODInW13	45.1
EmbSpatialBench	87.2	SUNRGBD	35.6
LingoQA	83.5	Nuscene	15.3
ERQA	68.5	Hypersim	12.9

Video Understanding

Benchmark	Score
VideoMME (w/ sub)	91.0
MLVU	90.2
VideoMME (w/o sub)	86.3
VideoMMMU	84.6
MVBench	78.2
MMVU	75.8
LVBench	75.1

3.2 NGen 4 Mini Evaluation Results

Selected reported points for NGen 4 Mini include:

HMMT 2025: 76.7%
Competitive performance against Qwen 3 and earlier GPT variants on advanced math and reasoning tasks.

Summary of Mini Capabilities:

Reasoning & Knowledge: High zero-shot accuracy across MMLU, GPQA Diamond, Humanity's Last Exam, and AIME 2025.
Engineering (SWE-Bench): Dedicated software engineering performance showcases top-tier accuracy.
Advanced Text & Reasoning: Compares favorably against Qwen3, DeepSeek, and Llama 4 on MMLU-Pro, MMLU-Redux, and SuperGPQA.
Logic & Mathematical Reasoning: Frontier performance on AIME 25, HMMT 25, and LiveBench.
Advanced Engineering & Coding: Top-tier scoring on LiveCodeBench, CFEval, and OJBench.
Vision & Multimodal Intelligence: Exceeds baseline performance compared to Gemini 2.5 Flash and GPT-5 Nano on MMMU (Val), MathVista, AI2D, and DocVQA.

3.3 NGen 4 Lite Evaluation Results

NGen 4 Lite is positioned as the smallest reasoning-capable tier in the NGen 4 family. The following data highlights the benchmark performance of NGen-4 Lite against industry peers (Early 2026).

Benchmark	NGen-4 Lite Score
OmniDocBench v1.5	91.0
Video-MME	88.1
GPQA Diamond	80.4
MMMLU	80.3
HMMT Feb 2025	78.0
MMMU-Pro	70.0
IFBench	62.5
ERQA	57.0

4 Conclusion

4.1 Summary of NGen 4

The NGen 4 series marks a pivotal moment in the evolution of artificial intelligence, successfully bridging the gap between raw linguistic fluency and rigorous logical deduction. By moving beyond the paragraph-style chain-of-thought of the NGen 3 generation and implementing a structured, step-by-step reasoning architecture, NGen 4 has redefined industry standards for factual accuracy and logic-based grounding.

The dual-architectural approach—utilizing NGen4Dense and NGen4MoMinMoM—allows the series to scale its intelligence dynamically across the Lite, Mini, and Pro tiers. Validated by world-leading scores on benchmarks such as AIME 2025 (100.0%) and OmniDocBench (93.9%), NGen 4 stands as the premier foundation for multimodal reasoning and complex problem-solving in the 2026 landscape.

Beyond its technical dominance, the series represents a deep commitment to cultural intelligence and safety alignment. Through its proprietary Indic Alignment phase and the implementation of teacher-steered safety protocols to mitigate "Shoggoth" behaviours, TNSA has produced a model that is both high-performing and uniquely responsible. While NGen 4 adopts some syntactical structures from its teacher models to ensure alignment, it retains an unparalleled depth of world knowledge and reasoning capability. Whether deployed for autonomous agentic tasks, advanced software engineering, or regional language comprehension, the NGen 4 series provides a secure, culturally aware, and remarkably intelligent framework for the global future of AI.

FINAL DISCLAIMER: This specific model release (the Qwen 3 Base Indic distillation operating on NGen-4-OW-ForCasualLM) is released purely for Research and Academic purposes. Commercial use is not permitted under this specific license.

Downloads last month: 45

Safetensors

Model size

33B params

Tensor type

BF16

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including TNSA/NGen-4-OW-30B-Thinking

NGen-4-OW Model Family (Not Yet Out)

Collection

The NGen-4-OW models are built on top of distillation data of NGen-4 Pro Model by TNSA running on NGen-4-OW Architecture. • 6 items • Updated 6 days ago • 1