The Wasmtime 28 surge and the runtime selection divide

The Bytecode Alliance and Wasmtime standardization

The Bytecode Alliance drives WebAssembly standardization through the Wasmtime runtime. Fastly built its delivery platform on Varnish, which uses the Varnish Configuration Language (VCL). VCL lacks for-loops and recursion, which limits the applications developers build at the edge. VCL also cannot access the body of a request, making JSON manipulation difficult. To overcome these limitations, Fastly developed Lucet, a native WebAssembly compiler and runtime written in Rust. Lucet used the Cranelift code generator, a project from Mozilla. When Fastly formed the Bytecode Alliance with Mozilla, Intel, Google, Microsoft, and Amazon, they decided to merge Lucet and Wasmtime. Wasmtime now incorporates all the logic from Lucet. The Bytecode Alliance operates as a nonprofit organization dedicated to building secure software foundations. Wasmtime provides capability-based security through WASI. Containers represent another option for edge compute, but they require heavy infrastructure and suffer from Docker escape vulnerabilities. Wasmtime provides a much lighter alternative for executing code securely. V8 was another option for serverless compute, but V8 only supports JavaScript, which is not fit for intense cryptography or image processing. V8 also has a delay for new invocations, which requires reusing instances and risks state leakage. Recent releases like Wasmtime 27.0 include support for the WebAssembly garbage collection proposal. Wasmtime 25.0 also added support for WASI 0.2.1. The Bytecode Alliance members continue to work on the WASI Subgroup, which recently ratified WASI 0.3.0. This specification re-bases WASI onto the WebAssembly Component Model’s async primitives. The Bytecode Alliance members, including Microsoft and Amazon, work to implement the Component Model which allows Wasm modules to be composed like Lego bricks across various language ecosystems without needing to know the underlying implementation language.

Runtime performance and deployment choices

I examine the 2026 benchmark data to evaluate Wasmtime, Wasmer, and WasmEdge. Wasmtime shows an execution overhead of 5% to 10% compared to native code. In tests for an HTTP echo server handling 10,000 requests, Wasmtime achieved 8,200 requests per second. Wasmer reaches 8,500 requests per second in the same test. WasmEdge achieves 9,100 requests per second due to its optimized async I/O. Wasmer provides three different compiler backends to give users flexibility. You can choose Singlepass for light compilation or LLVM for maximum optimization. Wasmer Singlepass instantiation takes less than 1 millisecond. With the LLVM backend, Wasmer can achieve an overhead of 2% to 3% compared to native code. Wasmer also provides the WAPM package manager to install tools like Python, Bash, and FFmpeg in one command. WasmEdge targets cloud-native environments and includes native integrations for Kubernetes and Docker. It also provides extensions for database connectivity, such as MySQL, PostgreSQL, and Redis. WasmEdge supports Ahead-of-Time compilation to eliminate JIT overhead for production deployments.

If you already know the basics, you know that runtime selection dictates your deployment velocity. Wasmtime maintains a memory footprint of approximately 15MB. WasmEdge operates with a much smaller footprint of 8MB.

Implementation reality and the BBC case

The BBC uses the Wasmtime C API to embed WebAssembly into their media players. This allows them to run the same code in the browser and in edge services. They use Emscripten to port native Single Service Player components to the web. They aim to use WebAssembly as a plugin model for interactive content like branching narratives. This approach allows them to host experiences from third-party producers while maintaining security through sandboxing. Large-scale production examples like Figma and Adobe Photoshop already rely on WebAssembly for heavy computation. WasmEdge provides extensions for AI inference, such as TensorFlow, PyTorch, and OpenCV, which other runtimes lack. It also supports image processing and async networking. I find the WasmEdge memory footprint of 8MB better than Wasmtime’s 15MB for constrained edge devices. I skip Wasmer for strict sandboxing needs because Wasmtime follows WASI specifications more closely. The debugging experience remains a painful part of the process because source maps often fail to accurately reflect the original source code. Will the Wasm Component Model 1.0 release finally resolve the fragmentation between different runtime implementations?