Kimi-Linear 48B-A3B: cutting-edge long-context attention engineering, not a frontier all-rounder

Asked:

“how advanced is Kimi-Linear 48B-A3B”

19 benchmark and efficiency claims about Moonshot AI's Kimi-Linear 48B-A3B — a mixture-of-experts model activating ~3B of 48B parameters with a hybrid 3:1 Kimi Delta Attention–to–global-MLA design and a 1M-token context window — collected from the arXiv paper, the Hugging Face model card, and one independent analysis (AcingAI). Scores come from architecture-matched research comparisons, not head-to-head tests against today's proprietary frontier models, and most originate with Moonshot itself.

Where Kimi-Linear wins its matched comparisons

Each benchmark has its own 0–100 panel — scores from different benchmarks are not directly comparable, and no synthetic overall score is drawn. Hover or tap a bar for scope and caveats.

Kimi-Linear 48B-A3BMatched full-attention (MLA) baseline, same size and trainingNo grey bar = baseline score not published for that row

Reported efficiency gains at long context

These are relative claims against a full-attention MLA stack — reported ratios, not measured benchmark scores — and they apply at contexts up to 1M tokens. AcingAI notes that at an ordinary ~8K context the speed win may be a rounding error.

What the numbers say

The core trade pays off: with 75% less KV cache, Kimi-Linear still beats its matched full-attention twin on MMLU (73.8 vs 71.6), RULER at 128K (84.3 vs 81.3), and the long-context suite average (54.5 vs 52.2) — arxiv.org

At a full 1M-token context, the 5.7T-token instruct checkpoint scores 94.8 on RULER, a range where the 8K-context Moonlight-Instruct baseline could not even be evaluated — arxiv.org

It did not win everything: the paper reports exceptions on LongBench V2 and Frames, and RULER-style retrieval is the friendliest possible test for a memory-constrained design — no shortcut-free retrieval curve at 512K–1M exists yet — acingai.com

The 6.3× faster time-per-output-token is a 1M-context result with no matched full-attention twin at full training scale; the independent analysis validates the architecture while flagging exactly that gap — acingai.com

Every claim, with scope and caveats

Category	Metric	Kimi-Linear	Comparator	Scope & caveat	Source

Method: 19 normalized benchmark and efficiency claims about Kimi-Linear 48B-A3B, one row per metric and source host, drawn from the arXiv paper (8 extracted rows), the Hugging Face model card (16 rows across huggingface.co and its hf.co mirror), and the independent AcingAI analysis (4 rows); collected November 2025. Benchmark values are scores on each test's own scale; efficiency values are reported ratios versus a full-attention MLA baseline at up to 1M-token context. Duplicate statements of the same number across hosts are merged in the charts and kept in the table.

This report was generated automatically by Keenable SELECT at a user's request, from publicly available web sources linked herein. Keenable does not review, verify, or endorse its contents and makes no representation as to accuracy, completeness, or timeliness; AI-based extraction may contain errors. Nothing in this report is investment, legal, financial, or other professional advice. All trademarks and referenced content remain the property of their respective owners; no affiliation or endorsement is implied. To report an error, rights concern, or request removal: legal@keenable.ai.