Most large models trade context length or modality for simplicity; this release instead merges instruction-following, reasoning, coding, and vision into a single dense 128B weight set with an exceptionally long 256k context. That design means you can run short chat replies or long agentic reasoning traces from the same checkpoint and tune compute per request via a reasoning_effort knob.
Key Capabilities
- Multimodal input with a vision encoder trained from scratch: allows the model to incorporate variable-size images alongside text and produce text-only outputs for downstream agents or tools. This makes it practical for workflows that mix screenshots, diagrams, or UI images with prompts.
- Large-context dense architecture (128B / 256k): keeps long agentic traces, tool-call histories, or documents in context without stitching or external memory for many use cases; useful for long-form reasoning, multi-turn tool orchestration, and retrieval-augmented generation at large window sizes.
- Configurable reasoning effort and native function-calling: lets you choose fast, low-cost replies or scale up compute for high-quality reasoning or agent runs; built-in JSON/function outputs simplify tool integration for autonomous agents.
- Practical deployment options: first-class support for vLLM and SGLang servers and compatibility with Transformers, plus recommended inference paths (vLLM for production-like performance). An EAGLE variant is offered for faster local inference.
- Multilingual and agentic-oriented: supports dozens of languages and benchmarks well on instruction, reasoning, and coding tasks per the model card's reported metrics.
Who it's for and trade-offs
Great fit if you need a single checkpoint that handles conversation, vision reasoning, and coding with very long context windows — for example, building agentic pipelines that keep extensive history, multimodal assistants, or coding agents that benefit from large local context. The model is published on Hugging Face and released under a Modified MIT license with exceptions for large-revenue companies, so evaluate the license against your commercial plans.
Look elsewhere if you need an explicitly tiny or quantized footprint for edge devices (this is a dense 128B model and demands substantial inference hardware), or if your project requires a permissive license without revenue-based restrictions. Also consider specialized smaller models or heavily quantized variants/EAGLE builds when latency and cost are the main constraints.
