Zamba2 7B
Hybrid Mamba2-attention 7B with strong instruction following and low-latency inference.
About
Zyphra's Zamba2-7B (November 2024, Apache 2.0) interleaves Mamba2 layers with a single shared-weight attention block — plus per-layer LoRA adapters so each shared block can specialize — giving near-linear-time prefill and a fixed-size recurrent state instead of a growing KV cache. Team-reported: the instruct version hits 69.95 on IFEval and 67.2 on MMLU, ahead of several 7B transformers, with roughly 4x the throughput of standard transformer blocks. A context-extended mode is available via a load flag. Honest caveats: Zyphra is a small team and the community ecosystem is thin; the strongest numbers are on instruction-following rather than broad knowledge.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported: Zamba2-7B-Instruct — IFEval 69.95, MMLU 67.2, MMLU-Pro 32.43.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (metrics tier, metrics-v1). Evidence: November 2024 (arXiv 2411.15242) verified; Apache 2.0 per HF card.; Team-reported IFEval 69.95 / MMLU 67.2 labeled as team-reported.; Small-team / thin-ecosystem caveats included; metrics tier due to lighter source depth..