Why this preview matters The open-weight preview gives practitioners direct access to a hybrid-attention architecture and a long-context inference workflow without relying on a hosted API. That makes it useful for experimenting with large-context reasoning, tool-calling agents, and local multimodal research where reproducing model internals or serving on custom hardware matters.
Key Capabilities
- Long-context multimodal inference: the preview checkpoint supports a 262,144-token context window, enabling sustained multi-step reasoning and large-document analysis that typical short-window models struggle with. This is achieved by a hybrid-attention decoder that keeps only a fraction of layers’ KV caches growing with context.
- Hybrid-attention architecture: 72 decoder layers (54 recurrent “delta-rule” layers + 18 global-attention layers) to reduce KV-cache growth while retaining global attention periodically for cross-sequence reasoning. Model config: hidden size 5120, vocab ~248k, vision tower (27 layers) for image/video inputs.
- Agent and developer-friendly features: configurable reasoning effort levels (high/medium/low), explicit tool-call tokens and a simple tool-call format, OpenAI-compatible chat/serving bindings, and a bundled processor for image/video inputs.
- Practical local deployment: bf16 checkpoint (≈66 GB weights on disk) with recommended GPUs (H200/H100 class) and transformers integration; the repo ships its own model code and requires loading with trust_remote_code=True.
Who should use it — trade-offs and fit
Great fit if you need a runnable, open-weight model to prototype long-horizon multimodal agents, reproduce research on hybrid-attention recurrent designs, or evaluate tool-calling and reasoning-effort mechanisms locally. It’s valuable for teams that need to inspect model internals, run custom inference stacks, or integrate multimodal reasoning into on-prem workflows.
Look elsewhere if you require the production/API checkpoint behavior (the preview is explicitly distinct from the production model), need a smaller, latency-optimized footprint for CPU or tiny GPU deployments, or require formal SLA-backed hosted inference and pricing plans. The preview is sizeable (33B params, significant memory and GPU requirements) and intended for research and engineering validation rather than light mobile/edge use.
Practical notes
The model card emphasizes this is a preview checkpoint different from any production/API snapshot (which may feature different context limits and metrics). The repo includes transformer-based quickstarts, image/video processor hooks, and recommended inference settings; it also documents hardware and disk requirements for running the bf16 checkpoint locally.
