On 2 September 2026, Qwen Developers promoted zg (zvec-grep), a local search tool from Alibaba's Zvec team. The project had entered public preview on 27 August (v0.2.0 on npm); the September push adds a long-form X article, benchmark tables, and official Qwen amplification — not a net-new repo drop.
zg is local-first retrieval infrastructure: it indexes code and documentation on your machine, then exposes semantic search, BM25, hybrid fusion, and ripgrep exact matching through one CLI and an MCP server. The pitch is that coding agents no longer need to guess keywords, run repeated rg passes, and manually stitch context from scattered hits.
What's new
- Project: zvec-ai/zvec-grep — Apache 2.0, npm @zvec/zvec-grep (v0.2.1 published 1 September)
- Install path: npm install -g @zvec/zvec-grep → zg install (auto-discovers Codex, Claude Code, Cursor, OpenCode) → zg index in a repo
- Search modes: vector (intent discovery), BM25 (ranked keywords), hybrid (RRF fusion), and zg query --rg for managed ripgrep without an index
- Agent integrations: MCP tools for Codex, Claude Code, Qwen Code, Qoder, Cursor, and OpenCode — indexing stays CLI-only; agents search a shared local index
- Default embedding: on-device local/potion-code-16m-v2 (16M params, ~32 MiB cache, no GPU required); optional remote models including qwen/text-embedding-v4 with explicit opt-in
- Built on: Alibaba Zvec vector + BM25 engine plus ripgrep for exhaustive exact match
Early benchmark claims
In paired A/B runs published with the launch (same agent, model, and prompts; baseline used native tools only), Zvec reports:
- SWE-QA-Bench (20 real-repo Q&A tasks): tool calls down more than half, input tokens down nearly half, judge score +1.50 points
- BrowseComp-Plus (80 deep-research questions): accuracy 98.67% → 99.00%, input tokens −37.56%, tool calls −43.52%, agent time −38.58%
These are vendor-run early evaluations on small samples — useful as a directional signal on context efficiency, not as independent proof. Index build is a one-time cost (Zvec cites Django's 3,457 files indexing in under 30 seconds on an Apple M4 Pro with the default local model).
Why it matters
Agent harnesses already ship file search, but the split is awkward: ripgrep wins on exact symbols and error strings; vector search wins when the user describes behavior in natural language ("restore theme preferences on startup") while the code names it hydratePreferences. Most agents today bounce between tools, burning tokens on irrelevant file reads.
zg's design bets on one retrieval layer with mode selection per stage — explore semantically, narrow with BM25, verify with rg — and token-efficient MCP output (compact previews by default, full text on demand). That maps directly onto the cost problem In-Terminal Coding agents face as cache pricing and long-horizon tasks expand context windows.
Our take
This is not a Qwen product launch — it is Alibaba Zvec infrastructure that Qwen Developers is endorsing for the Qwen Code ecosystem. Treat v0.2.x as an early preview: ~289 GitHub stars and modest npm uptake mean production hardening is still ahead (roadmap items include PDF/Office extractors, graph search, and mobile targets).
Still worth watching. If hybrid local retrieval becomes a default MCP plugin alongside shell and browser tools, the competitive question shifts from "which model codes best" to "which harness wastes fewest tokens finding the right file." The September Qwen push gives zg distribution; the next test is whether teams outside the Alibaba orbit adopt it over rolling their own rg + embeddings stack.
