Zamba2 2.7B

Zamba2-2.7B is Zyphra's small hybrid model that replaces most of the transformer stack with Mamba2 state-space blocks, keeping just two shared attention layers threaded through the network in an ABAB pattern. LoRA-style projectors on the shared MLPs let each invocation specialize slightly without paying for full separate layers. The result is a 2.7B model trained on 2 trillion tokens that Zyphra positions squarely at on-device and edge deployment, where the missing full KV cache keeps memory use far below a comparable transformer.

On the provider's own numbers, the instruct variant scores 72.4 aggregate MT-Bench and 48.0 on IFEval, ahead of Gemma2-2B-Instruct and even Mistral-7B-Instruct on those two suites. Inference is the real pitch: lower time to first token and faster generation with a smaller memory footprint than transformer baselines. The tradeoff is ecosystem friction. Mamba kernels want specific builds, community fine-tunes are thin next to Llama-family models, and every headline number comes from Zyphra itself.

• Aggregate MT-Bench: 72.40; IFEval: 48.02 (Zamba2-2.7B-Instruct, provider-reported)
• Mistral-7B-Instruct: 66.4 / 45.3; Gemma2-2B-Instruct: 51.69 / 42.20 on the same two suites (provider-reported)
• Latency/memory advantage over transformer baselines claimed, no absolute figures published