At 27B parameters with a native 262,144-token context (extensible up to ~1,010,000), this release focuses on making dense open-weight models practical for agentic coding and real-world multimodal tasks. Rather than only increasing size, the update targets stability, agent workflows, and retaining useful intermediate reasoning so iterative developer workflows cost fewer tokens and require less orchestration.
Key Capabilities
- Agentic coding improvements: better repository-level reasoning and frontend workflows (noted gains on internal SWE-bench and LiveCodeBench metrics), plus built-in tooling support for tool-calling agents.
- Preserved reasoning traces: an option to retain historical "thinking" blocks to reduce redundant re-reasoning across turns and improve decision consistency in multi-step agent tasks.
- Multimodal & long-context support: vision encoder + causal LM fusion; native 262,144-token context and guidance for YaRN-based extensions to ~1M tokens for long-horizon tasks.
- Serving & efficiency features: examples and recommended stacks for production serving (vLLM, SGLang, KTransformers, Hugging Face Transformers) and MTP/speculative decoding options to boost throughput.
- Practical benchmarks: competitive internal results across coding (SWE-bench) and vision-language suites (MMBench / RealWorldQA / OCRBench), showing the model is tuned for applied tasks rather than synthetic scaling alone.
Who it's for and trade-offs
Great fit if you need a deployable, open-weight multimodal LLM that balances capability and operational cost—teams building coding assistants, multimodal QA systems, or agent pipelines that benefit from retained reasoning traces. Look elsewhere if you need the absolute top-tier few-shot performance of very large (100B+) models for tasks where extra scale outweighs deployment complexity, or if you need a permissive commercial license beyond Apache-2.0 nuances.
Where it fits
Practically positioned between smaller local models and the largest foundation models: it aims to deliver agent-style code reasoning and strong vision-language performance at a size that remains deployable on multi-GPU inference stacks. The release emphasizes real-world utility (tool use, long-context reasoning, deployment recipes) over raw parameter count.
(First published: 2026-04-21.)
