Zyphra releases Zamba2-VL: vision-language models with 10x faster response

Zyphra has released Zamba2-VL, an open family of vision-language models at 1.2B, 2.7B and 7B parameters. They are built on a hybrid architecture that combines Mamba2 and Transformer blocks. The key result: time to first token drops by roughly 10x compared with pure Transformer VLMs of similar size.

Three sizes, one license

The family includes three variants: 1.2B, 2.7B and 7B parameters. All three are released under the Apache 2.0 license, which means free commercial use with no restrictions on embedding, modifying and redistributing them. That gives commercial and research projects full freedom.

Zamba2-VL models are full vision-language models. They process images and text together, which opens up a wide range of tasks: captioning pictures and photos, visual question answering, analyzing documents with illustrations, parsing screenshots of user interfaces, and working with medical images. Unlike text-only LLMs, VLMs can answer questions about what a picture shows and combine visual and textual context in a single request.

On standard benchmarks, Zamba2-VL holds its own against pure Transformer VLMs of similar size. Moving to a hybrid architecture does not require sacrificing accuracy for speed: both stay at a competitive level.

How the hybrid backbone works

Most modern language and multimodal models are built on a pure Transformer architecture. In it, every newly generated token "looks at" the entire preceding sequence through the attention mechanism. It is a powerful approach but computationally expensive: with long contexts, the workload grows quadratically. This is where the speed bottleneck appears, including high time to first token.

Mamba2 is an architecture based on a state space model (SSM). Instead of scanning the full history, it compresses the preceding context into a compact "state" that updates linearly as new tokens are processed. Zamba2-VL alternates Mamba2 blocks with regular Transformer layers: the SSM blocks provide speed and efficiency, while the Transformer layers add flexibility for handling complex dependencies.

The result:

• Time to first token drops by roughly 10x
• Quality stays competitive with pure Transformer VLMs
• A smaller compute footprint at inference
• Better scaling on long contexts
• Deployment on less powerful hardware without losing responsiveness

Why TTFT matters

Time to first token (TTFT) is the gap between sending a request and the first character of the response appearing. It defines how "alive" interactive systems feel: chatbots, voice assistants, API services where reaction speed matters. While the model thinks, the user waits. High TTFT feels like a freeze, even if the final answer is good.

Cutting TTFT 10x is a multiplicative practical gain. On the same hardware, it means either a much more responsive service or the ability to handle significantly more requests at once. For companies paying for GPU time, either option directly affects the unit economics of the product.

Open models with this kind of response speed make it possible to build

products where response latency previously made a whole class of solutions unviable.

What it means

Hybrid SSM + Transformer architectures keep moving from academic papers into practical products. Releasing Zamba2-VL as a family of three models, from the compact 1.2B to the full-size 7B, covers different deployment scenarios, from resource-constrained devices to server farms.

Openness under Apache 2.0 lowers the barrier to entry: teams can take a ready-made, fast multimodal model without depending on commercial APIs, with all their prices, limits and the risk of sudden changes in terms.