---
id: "nr-nunchux-ai-s-vc-attention-speeds-up-minimax-h3-7df29c"
slug: "nunchux-ai-s-vc-attention-speeds-up-minimax-h3-7df29c"
title: "Nunchux AI's VC-Attention speeds up MiniMax-H3 attention without retraining"
observed_at: "2026-09-18T13:48:49+00:00"
record_date: "2026-09-18"
date_kind: "observed_at"
novelty: "new"
source_platform: "x.com"
topics: ["inference-optimization","quantization","video-generation"]
entities: ["MiniMax","Nunchux AI","VC-Attention"]
related_patterns: []
source_url: "https://x.com/MiniMax_AI/status/2100792780494889194"
source_urls: ["https://x.com/MiniMax_AI/status/2100792780494889194"]
basket: {"id":"0325e6b8-f527-4f9b-bfa3-3f30e57f380b","revision":1,"cluster_id":"7df29c4c-8643-4784-b220-8f5b818339fc"}
type: "post"
description: "Nunchux AI's training-free VC-Attention reports 1.6x and 1.5x attention speedups over FlashAttention-4 on MiniMax-H3, with better fidelity than SageAttention2."
why_it_matters: "The posts do not publish end-to-end generation latency, quality metrics for generated video, or results on any model other than MiniMax-H3. An independent reproduction on MiniMax-H3 or another open model would show whether the reported fidelity advantage holds outside the authors' evaluation. A published combination of VC-Attention with a sparse method, with measured speed and fidelity, would test the stacking question MiniMax raised."
verification_level: "source-inspected"
signal_type: "field_note"
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:nunchux-ai-s-vc-attention-speeds-up-minimax-h3-7df29c"
retrieval_nugget: "Nunchux AI's training-free VC-Attention reports 1.6x and 1.5x attention speedups over FlashAttention-4 on MiniMax-H3, with better fidelity than SageAttention2. Nunchux AI and collaborators introduced VC-Attention, a low-bit attention method that works without retraining, and MiniMax highlighted it in a post about faster video attention for its MiniMax-H3 model. According to Nunchux AI, VC-Attention speeds up attention on MiniMax-H3 by 1.6x"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:b448671c2a376f33c5fc0a409fd6b3fc4461e2d2f2929dbc34fd058f04b1b070","reviewed_at":"2026-09-24T13:28:46Z","source_evidence_count":1,"verified_claim_count":3,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":3,"implication_count":1,"watch_condition_count":1,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"Training-free low-bit attention is turning into a deployment lever that model owners can adopt after release rather than a decision fixed at training time.","observed_facts":[{"text":"Nunchux AI reports VC-Attention speeds up attention on MiniMax-H3 by 1.6x on B200 over FlashAttention-4.","source_urls":["https://x.com/MiniMax_AI/status/2100792780494889194"]},{"text":"The reported speedup on B300 over FlashAttention-4 is 1.5x.","source_urls":["https://x.com/MiniMax_AI/status/2100792780494889194"]},{"text":"V-Smooth reduces value quantization error, while ExpCast-FP8 makes softmax faster through approximation.","source_urls":["https://x.com/MiniMax_AI/status/2100792780494889194"]}],"mechanism":"The method pairs two components, with V-Smooth reducing value quantization error and ExpCast-FP8 making the softmax step faster through approximation, so speed and fidelity are traded explicitly.","why_now":"The speedups are reported against FlashAttention-4 on B200 and B300, so the comparison targets a current attention baseline rather than an older kernel.","implications":["Teams serving attention-heavy video models can test low-bit attention kernels as a post-training step before considering retraining or architectural changes."],"evidence_boundary":"The posts do not publish end-to-end generation latency, quality metrics for generated video, or results on any model other than MiniMax-H3.","watch_conditions":["An independent reproduction on MiniMax-H3 or another open model would show whether the reported fidelity advantage holds outside the authors' evaluation."],"related_records":[],"new_branch_reason":"This is a concrete, hardware-specific result on post-training attention quantization for a named production model, distinct from general kernel news."}
routes: {"html":"https://newruntime.com/posts/nunchux-ai-s-vc-attention-speeds-up-minimax-h3-7df29c/","markdown":"https://newruntime.com/posts/nunchux-ai-s-vc-attention-speeds-up-minimax-h3-7df29c.md","json":"https://newruntime.com/posts/nunchux-ai-s-vc-attention-speeds-up-minimax-h3-7df29c.json"}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
---

# Nunchux AI's VC-Attention speeds up MiniMax-H3 attention without retraining

## Retrieval answer

Nunchux AI's training-free VC-Attention reports 1.6x and 1.5x attention speedups over FlashAttention-4 on MiniMax-H3, with better fidelity than SageAttention2. Nunchux AI and collaborators introduced VC-Attention, a low-bit attention method that works without retraining, and MiniMax highlighted it in a post about faster video attention for its MiniMax-H3 model. According to Nunchux AI, VC-Attention speeds up attention on MiniMax-H3 by 1.6x

Nunchux AI and collaborators introduced VC-Attention, a low-bit attention method that works without retraining, and MiniMax highlighted it in a post about faster video attention for its MiniMax-H3 model. According to Nunchux AI, VC-Attention speeds up attention on MiniMax-H3 by 1.6x on B200 and 1.5x on B300 over FlashAttention-4, with better fidelity than SageAttention2.

Training-free low-bit attention is turning into a deployment lever that model owners can adopt after release rather than a decision fixed at training time. The method pairs two components, with V-Smooth reducing value quantization error and ExpCast-FP8 making the softmax step faster through approximation, so speed and fidelity are traded explicitly.

The speedups are reported against FlashAttention-4 on B200 and B300, so the comparison targets a current attention baseline rather than an older kernel. Teams serving attention-heavy video models can test low-bit attention kernels as a post-training step before considering retraining or architectural changes. Nunchux AI also states that the method works with existing sparse attention, and MiniMax openly asks whether combining low-bit computation with sparse methods such as Sol-Attn could push efficiency further.

The posts do not publish end-to-end generation latency, quality metrics for generated video, or results on any model other than MiniMax-H3. An independent reproduction on MiniMax-H3 or another open model would show whether the reported fidelity advantage holds outside the authors' evaluation. A published combination of VC-Attention with a sparse method, with measured speed and fidelity, would test the stacking question MiniMax raised.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Inference Optimization](https://newruntime.com/topics/inference-optimization.md): Explore the inference-optimization topic hub.
- topic hub: [Quantization](https://newruntime.com/topics/quantization.md): Explore the quantization topic hub.
- topic hub: [Video Generation](https://newruntime.com/topics/video-generation.md): Explore the video-generation topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)