🚀 @ruvector/ruvllm-wasm v2.0.0 — Browser-native LLM inference is here

[ruvnet]

@ruvector/ruvllm-wasm v2.0.0 — Browser-Native AI Runtime

The first functional release of @ruvector/ruvllm-wasm is live on npm. This replaces the deprecated v0.1.0 placeholder with a fully compiled 435KB WebAssembly binary — semantic routing, adaptive learning, KV cache management, and chat template formatting running entirely client-side.

No server. No API keys. No network latency.

Full documentation: crates/ruvllm-wasm/README.md

---

Install

npm install @ruvector/ruvllm-wasm

Quick start

import init, { ChatTemplateWasm, ChatMessageWasm, HnswRouterWasm, healthCheck } from '@ruvector/ruvllm-wasm';

await init();
console.log(healthCheck()); // true

// Format conversations for any model family
const template = ChatTemplateWasm.detectFromModelId("meta-llama/Llama-3-8B");
const prompt = template.format([
  ChatMessageWasm.system("You are a helpful assistant."),
  ChatMessageWasm.user("What can you do?"),
]);

// Route queries to the best agent in <1ms using HNSW semantic search
const router = new HnswRouterWasm(384, 1000);
router.addPattern(embedding, "code-agent", "handles code generation");
router.addPattern(embedding2, "research-agent", "handles research queries");
const result = router.route(queryEmbedding);
console.log(result.name, result.score); // "code-agent" 0.95

// Persist and restore router state
const snapshot = router.toJson();
const restored = HnswRouterWasm.fromJson(snapshot);

---

What ships in this release

Feature	Type	Description
HNSW Semantic Router	HnswRouterWasm	Bidirectional graph with cosine similarity — 150x faster than linear scan. Add patterns, route queries, serialize to JSON.
KV Cache	KvCacheWasm	Two-tier token cache (FP32 tail for recent tokens + u8 quantized store for history). Configurable tail length, head count, head dimension.
Chat Templates	ChatTemplateWasm	Format prompts for 7 model families: Llama3, Llama2, Mistral, Qwen, ChatML, Phi, Gemma. Auto-detect from model ID.
MicroLoRA	MicroLoraWasm	Per-request LoRA adaptation (rank 1-4). Forward pass, gradient accumulation, weight updates — all in <1ms. Serialize/restore adapter state.
SONA Instant	SonaInstantWasm	EMA quality tracking, adaptive rank adjustment, EWC-lite weight importance, pattern buffer with cosine-similarity suggestion.
Memory Pool	InferenceArenaWasm	O(1) bump allocator for inference temporaries. Pre-size for model dimensions. Tracks high-water mark.
Buffer Pool	BufferPoolWasm	Pre-allocated buffers in size classes (1KB-256KB). Configurable max per class. Pool hit rate tracking.
Web Workers	ParallelInference	Async parallel matmul, attention, layerNorm across multiple workers. SharedArrayBuffer detection, capability-level reporting.
Feature Detection	Free functions	feature_summary(), optimal_worker_count(), supports_parallel_inference(), is_simd_available(), cross_origin_isolated()
TypeScript	.d.ts	Complete type definitions for all 20+ exported types and free functions

Package stats

Metric	Value
WASM binary	435 KB (178 KB gzipped)
JS glue	128 KB
TypeScript defs	45 KB
Exported types	20+
Browser support	Chrome 57+, Firefox 52+, Safari 11+, Edge 79+

Architecture

JavaScript/TypeScript App
        │
        ▼
  ┌─────────────┐
  │ ruvllm_wasm │  ← 435KB compiled WASM
  │   .js glue  │  ← wasm-bindgen generated
  └──────┬──────┘
         │
    ┌────┴────────────────────────┐
    │                             │
    ▼                             ▼
┌──────────┐  ┌──────────────────────┐
│ Core     │  │ Intelligent Features │
│          │  │                      │
│ KvCache  │  │ HnswRouter (150x)   │
│ Arena    │  │ MicroLoRA (<1ms)     │
│ BufPool  │  │ SONA Instant         │
│ Chat TPL │  │ Web Workers          │
└──────────┘  └──────────────────────┘

Build from source

# Requires: rustup target add wasm32-unknown-unknown && cargo install wasm-pack

# Release build (workaround for Rust 1.91 codegen bug)
CARGO_PROFILE_RELEASE_CODEGEN_UNITS=256 CARGO_PROFILE_RELEASE_LTO=off \
  wasm-pack build crates/ruvllm-wasm --target web --scope ruvector --release

# Dev build (no workaround needed)
wasm-pack build crates/ruvllm-wasm --target web --scope ruvector --dev

See ADR-084 for build details and known limitations.

What's next

• IntelligentLLM — Combined router + LoRA + SONA in one unified API
• WebGPU attention — GPU-accelerated attention (matmul GPU path already works)
• GGUF streaming loader — Load quantized models directly in the browser
• Worker completion signals — Replace setTimeout polling with message-based coordination

Related packages

Package	Version	Description
@ruvector/ruvllm	2.5.2	Node.js LLM orchestration with SONA learning
ruvector	0.2.11	Full CLI with 48 commands + 91 MCP tools
@ruvector/rvf	—	Cognitive container runtime

---

PR: #241 | ADR: ADR-084 | Docs: crates/ruvllm-wasm/README.md

[ruvnet]

Moved to Discussions > Announcements.