February 2026
Open Weights Model
NGen-4-OW-120B-Thinking
A research-oriented NGen 4 open-weight checkpoint for direct evaluation, adaptation, and community experimentation.

Release Notes
This model is a fine-tuned derivative of the NGen-4 Pro model. We have distilled deep Indic Knowledge into the Qwen 3.5 120B Base model.
Please note that this model may not perform as expected with Qwen 3.5'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.
Architecture
The NGen 4 open-weight releases are distilled from the NGen 4 frontier family into public checkpoints intended for research, experimentation, and community evaluation.
Each model keeps the NGen 4 emphasis on structured reasoning, Indic intelligence, and safety-aware instruction following while adapting those traits into a smaller or more accessible base architecture.
For best results on the open-weight variants, use the recommended NGen-4-OW-ForCasualLM architecture when it is available for the checkpoint. Some releases remain compatible with their original base architecture, but the custom path is preferred for stronger Indic behavior.
Training & Alignment
The OW series uses a curated blend of open and synthetic data, with special attention on Indic language coverage, reasoning traces, and practical instruction-following tasks.
Distillation draws from NGen 4 Pro and related teacher models, then aligns outputs toward a more reliable research checkpoint that can be evaluated outside the hosted API platform.
Safety filtering, deduplication, and quality controls are applied before release so the checkpoints remain useful for academic and applied model evaluation.
Evaluation Focus
The open-weight models are meant to make NGen 4 behavior easier to inspect and compare. Their value is in transparent evaluation, controlled deployments, and research workflows where teams need direct access to weights.
- Reasoning and instruction-following behavior across math, planning, and multi-step analysis tasks.
- Indic-language understanding, translation robustness, and culturally aware response quality.
- Compatibility with community inference stacks and reproducible research workflows.
- Practical limitations of each base architecture compared with the hosted NGen 4 API models.
Responsible Release
This specific model release (the Qwen 3.5 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.

Build with TNSA models
Explore model families, snapshots, and developer tools for production AI systems.