@marc.kaz · Marc Kaz
Saved 2026-06-15 · Posted 2026-06-06 · Status: New
TurboVec is a new open-source vector index built on Google Research’s TurboQuant that delivers extreme compression without sacrificing much performance.
Ships with:
• 16x memory reduction (31 GB → 4 GB on 10M vectors)
• 2–4 bit quantization with strong recall
• Outperforms or matches FAISS on speed
• Pure Rust + Python bindings, no training phase
• Search-time filtering with SIMD optimizations
👉 https://github.com/RyanCodrai/turbovec
Content ideas (0)
No ideas generated yet. Run /instagram-sync ideate from Claude Code to create some.
Comments (15)
How ? Quantized ? Pruning? Knowledge distillation?
Keep your camera still ... hippie!
🚨 GOOGLE JUST SHRANK 31GB OF AI MEMORY DOWN TO 4GB
TurboVec is a new open-source vector index built on Google Research’s TurboQuant that delivers extreme compression without sacrificing much performance.
Ships with:
• 16x memory reduction (31 GB → 4 GB on 10M vectors)
• 2–4 bit quantization with strong recall
• Outperforms or matches FAISS on speed
• Pure Rust + Python bindings, no training phase
• Search-time filtering with SIMD optimizations
Massive efficiency win for local and production vector search.
👉 https://github.com/RyanCodrai/turbovec
Who’s running vector search on a laptop today? Drop a 🔥
I genuinely did not understand. They shrink the vectors so that you save ram space? Like, do we even need it to be in the ram? I run elastic for fun, and it's search engine is pretty decent for my case (around 30k records, soon to be 1M). I don't care how much ram it uses.
Should come to Ollama soon. Or probably is already integrated. Need to check.
@r18hu815w45 nah not possible made some research
What it does: It compresses text embedding vectors (like those used in Retrieval-Augmented Generation / RAG) from 32-bit floats down to 2 or 4 bits. It reduces the memory footprint of your search database (e.g., shrinking 10 million search documents from 31 GB to 4 GB).
What it does NOT do: It does not compress LLM neural network weights (the matrices that actually generate tokens and understand language), nor does it have an inference engine to execute a model.
With quant and vec open source just became viable
Oh they used winrar? Hope they pay their license ;)
Anyone else building optimizations and days and weeks later some else is doing it too.
The AI cholo has spoken.
I haven't had time to fully check, try, or use this repo yet. However, I have a few concerns regarding its architecture and implementation (and I'm not entirely sure if the author @codrairyan - hope it's you - considered them):
1) What happens if just a single piece of a single document changes?
2) Is the command idempotent?
3) How is processing optimization handled?
I have some ideas and proposals to address this, mainly based on a combination of chunking, hashing, and multithreading. If anyone wants to take a coding agent, verify these points, and - if valid - open a pull request or an issue after generating a detailed implementation spec, that would be great. If these optimization issues are present, the repo can't really be considered production-ready.
Finally, if any of you reading this have time to verify it, please check if my proposal makes sense and feel free to open an issue or a PR. I'd love to get some feedback, and why not... credit/cite me in the implementation if you use the idea!
Google day by day !! , but still conflicts with gemma 4
It means it compresses the context window / knowledge, not the LLM itself. CMIIW
Could this not just be compounded then for more GBs?
I have created sped-cli npm package, do check that out it helps reduce token usage on basic terminal commands that claude code and other agents use and its free and opensource
Here’s the link
———————————
https://github.com/electroo69/sped-cli