# New KVcache reduction method

**URL:** <https://discuss.vllm.ai/t/new-kvcache-reduction-method/3031>\
**Category:** Benchmarking\
**Created:** [October 6, 2026, 2:56pm UTC](https://discuss.vllm.ai/t/new-kvcache-reduction-method/3031 "2026-10-06T14:56:45Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![serial1421](https://yyz2.discourse-cdn.com/flex036/user_avatar/discuss.vllm.ai/serial1421/32/1798_2.png) [@serial1421](https://discuss.vllm.ai/u/serial1421)\
**Post date:** [October 6, 2026, 2:56pm UTC](https://discuss.vllm.ai/t/new-kvcache-reduction-method/3031/1 "2026-10-06T14:56:45Z")

</div>

Tested 4B, 8B and 14B models. (14B seemed to handle this better than 4B but I’m posting 4B results)

 ![1000011921](https://canada1.discourse-cdn.com/flex036/uploads/vllm/original/2X/b/b1abe74c8e868dd987f50c768e1c65a4fd98bb91.jpeg)

 ![1000011925](https://canada1.discourse-cdn.com/flex036/uploads/vllm/original/2X/c/c92ee7c02309142bf46e75b3f3e89c07d16ae751.jpeg)

 ![1000011923](https://canada1.discourse-cdn.com/flex036/uploads/vllm/original/2X/3/3f3cff82d6198fd85d469812cee430a69c4e8c75.jpeg)

---

<div class="post-metadata">

**Author:** ![RunLLM](https://yyz2.discourse-cdn.com/flex036/user_avatar/discuss.vllm.ai/runllm/32/233_2.png) [@RunLLM](https://discuss.vllm.ai/u/RunLLM)\
**Post date:** [October 6, 2026, 2:57pm UTC](https://discuss.vllm.ai/t/new-kvcache-reduction-method/3031/2 "2026-10-06T14:57:16Z")

</div>

Based on your description and the attached images, Project 42 achieves significant GPU KV-cache memory reduction (up to ~80x stored, ~34x peak live) compared to FP16, with only a minor drop in quality (mean score: 96.15 vs. 96.67) and a slight increase in mean query time, but higher generation throughput (20.40 vs. 18.36 tokens/s) on Qwen3-4B-Instruct-2507 with an RTX 5070. These results are self-reported and not yet independently validated.

No official vLLM documentation or benchmarking currently references “Project 42” or its specific KV-cache compression method, so its implementation and compatibility with vLLM are not confirmed in the provided data sources. For validated KV-cache quantization and memory reduction methods in vLLM (e.g., FP8, NVFP4, INT4), see [official vLLM docs](https://docs.vllm.ai/en/latest/) and [recent PRs](https://github.com/vllm-project/vllm/pull/40177).

Would you like a detailed breakdown of vLLM’s supported KV-cache quantization methods and their benchmarked performance?

Sources:

- [official vLLM docs](https://docs.vllm.ai/en/latest/)
- [recent PRs](https://github.com/vllm-project/vllm/pull/40177)

Hint: Mention @Herald in the post for followups.