Hymba 1.5B
Hybrid Mamba-attention small model with a tiny KV cache and fast inference.
About
NVIDIA's Hymba-1.5B (November 2024) runs attention heads and Mamba state-space heads in parallel inside each layer, prepends learnable 'meta tokens' to every prompt, and shares KV cache across layers — cutting cache size roughly 10x versus pure transformers. Team-reported: it beats Llama-3.2-3B on commonsense reasoning with an 11.67x smaller cache and 3.49x faster inference. Trained on 1.5T tokens with an 8K context window. It ships under the NVIDIA Open Model License, which NVIDIA describes as permitting commercial use — read the actual license text before shipping a product on it. Honest caveat: this is a research-first release, so quantizations and serving integrations are thinner than for Llama or Qwen.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported: +1.32% average commonsense accuracy over Llama-3.2-3B with 11.67x smaller cache.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: November 2024 release verified via NVIDIA developer blog and press; trained Sep-Nov 2024 per HF card.; NVIDIA Open Model License with commercial-use claim; advised reading license text.; Team-reported +1.32% over Llama-3.2-3B with 11.67x smaller cache — labeled as team-reported.; Thin-ecosystem caveat included..