Browser / WASM LLM harness

Comparative Browser LLM Benchmarks

This page benchmarks browser-first LLM runtimes now, while leaving an explicit integration lane for Qualia once the independent WASM work is ready. No Qualia engine wiring is done here yet.

Scope

Live adapters are wired where the browser path is stable enough to exercise today. Engines without a pinned browser bootstrap stay marked as adapter shells so the page remains truthful instead of inventing benchmark data.

Benchmark Controls

Shared prompt and run settings

Runs focus on cold-start load time, time-to-first-token, completion latency, estimated throughput, and browser heap movement where the platform exposes it.

Keep this short so browser runs stay practical.
Low randomness helps make timings more comparable.
Execution Environment
Device summary
Loading…
Topology
Loading…
Cross-origin isolation
Loading…
WebGPU adapter
Loading…
Heap API / SIMD
Loading…
Engines
Results

Latest benchmark matrix

`ttft` is exact for streaming adapters and falls back to full completion time for non-streaming browser APIs.
Engine Load TTFT Completion Est. tok/s Heap Δ Summary
Benchmark Log

Live harness log


            


            

                
Qualia Lane

                
Reserved integration contract

                

                    This slot is intentionally a placeholder. The page is ready to accept a Qualia adapter later, but it does not import or mutate the current Qualia WASM work.
                

                

                    

                        
                        Waiting on independent engine handoff
                    

                    

interface BrowserLlmAdapter {
  prepare(config, hooks): Promise<void>;
  run(prompt, config, hooks): Promise<{
    loadMs: number;
    ttftMs: number;
    generationMs: number;
    outputText: string;
    approxTokens: number;
    heapDeltaMb: number | null;
  }>;
}
                    

                

                

                    Planned handoff points: `.q42`/GGUF selection, OPFS cache policy, worker isolation, provenance export, and any Qualia-specific semantic overhead metrics once the runtime is stable.