ModelBrief · Kimi K3.1 Pro · Topic Page
Last updated Nov 20, 2026

Kimi K3.1 Pro vs Open Rivals: Best Open Weights?

ModelBrief · Published 2026-11-16 · Fact-check 2026-11-16 · Reference: Artificial Analysis, Moonshot tech blog

"Best open weights" is Kimi K3.1 Pro's loudest title, sourced from the Artificial Analysis Index v5.0 score of 46 and Agent Arena's No. 1 open ranking. But the title's weight varies by test: on the overall index it leads; on cyber it loses to DeepSeek-V4.1-Flash. This article compares it with DeepSeek-V4.1-Flash and Tencent Hy4 using only figures with a verified comparable basis — and lists the gaps honestly. Where no comparable figure exists, this site won't invent one.

Where the title comes from: AA Index v5.0 and Agent Arena

AA Intelligence Index v5.0 (~Nov 13): K3.1 Pro 46, the best open-weights model on the board; Hy4 35. On v4.3.2: K3.1 Pro 55, No. 4 overall — on the same version, Opus 5.5 scored 58, Sonnet 5.5 56 and GPT-6 Astra 53, leaving K3.1 Pro just 3 points behind the closed flagships. Agent Arena: No. 1 open weights, No. 13 overall, net improvement +5.1% (Max), on 17,500+ real long-horizon agent sessions. AA-measured Terminal-Bench 4.0: 67.2%, the highest of every model on its page. These four are the whole basis for "best open weights" — overall index, agent practice, terminal tasks, three corroborating sources.

Cyber: leading on the AA index, losing on xdev-bench to DeepSeek

AA Cyber Index (Nov 13): K3.1 Pro 64, Hy4 58, DeepSeek-V4.1-Flash-1017 57 — K3.1 Pro leads by 6–7 points. AA CyberGym end-to-end (AA-measured): K3.1 Pro 94.7%, No. 1 of 26, ahead of GPT-6 Sol at 93.1%. But xdev-bench-hard pass@3 (the vendor's reported basis): the released K3.1 Pro 81.0%, DeepSeek-V4.1-Flash 85.7% — DeepSeek leads by 4.7 points. The reason, as the benchmarks article explains: K3.1 Pro's 81.0% is the deliberately weakened figure, with the uncleaned checkpoint at 90.5%. On the uncleaned comparison K3.1 Pro leads 90.5% to 85.7%; on the released comparison it trails 81.0% to 85.7%. Users can only download the released version, so "No. 1 open weights on cyber" does not hold for K3.1 Pro.

AA Cyber Index comparison (Nov 13, 2026) Kimi K3.1 Pro Tencent Hy4 DeepSeek-V4.1-Flash Kimi K3.1 Pro 64 Tencent Hy4 58 DeepSeek-V4.1-Flash 57 AA Cyber Index, Nov 13, 2026; orange is Kimi K3.1 Pro
AA Cyber Index (points; Artificial Analysis, Nov 13, 2026; orange is Kimi K3.1 Pro)

Scale and price: K3.1 Pro is the biggest, and among the priciest

Scale: K3.1 Pro is 2.8T total / 104B active; Hy4 is 770B total (per this site's Hy4 topic). K3.1 Pro is about 3.6× Hy4's size, and its weights are correspondingly heavier: 1.56TB, against Hy4's downloadable weights (whose exact size this site has not verified, so no comparison is drawn). Price: K3.1 Pro $2/$10; Hy4 TokenHub ¥6/¥18, 50% off through year-end. DeepSeek-V4.1-Flash's API price was not verified in this research round and is left blank. Biggest model, not the cheapest price — K3.1 Pro's "best open weights" is bought with the largest model and the heaviest weights, as the self-hosting guide accounts in detail.

The gaps: why Qwen4-Max and GLM-5.3-Flash are not in the table

The two most worth-comparing rivals are missing from the table for lack of comparable independent scores: Qwen4-Max (this site's topic, released Nov 11) and GLM-5.3-Flash (K3.1 Pro's base) have no verified AA Index v5.0 same-version score, nor LMArena or LiveBench scores on a comparable basis. GLM-5.3-Flash has only one comparable figure: it is K3.1 Pro's base, but the base's own independent scores are outside this research round. DeepSeek-V4.1-Flash's and Hy4's same-version AA Index v5.0, LMArena and LiveBench scores are likewise unverified and left blank. That is not an oversight, it is honesty: a comparison table with gaps is more useful than one filled with invented numbers. Filling those gaps is the first watch item on the 8-hour tracking loop.

How to choose: advice for three kinds of tasks

General agent work: pick K3.1 Pro — AA Index 46 (best open weights), Agent Arena No. 1 open, AA-measured TB 4.0 67.2% (highest on the page), three corroborating sources. Cyber tasks: pick DeepSeek-V4.1-Flash — xdev-bench 85.7% beats the released K3.1 Pro's 81.0%, and DeepSeek did not deliberately weaken its model. Budget-sensitive Chinese-language work: Hy4's TokenHub ¥6/¥18 (50% off through year-end) is worth comparing, but Hy4's AA Index 35 against K3.1 Pro's 46 means the lower price costs capability. Note: this advice rests only on verified comparable figures; it may change once Qwen4-Max's and GLM-5.3-Flash's scores arrive.

FAQ

Is Kimi K3.1 Pro really the best open weights?

On the overall index, yes: AA Index v5.0 46 is the best open weights, and Agent Arena ranks it No. 1 open. On cyber, no: the released xdev-bench 81.0% trails DeepSeek-V4.1-Flash's 85.7%.

Why does DeepSeek score higher on xdev-bench?

Because K3.1 Pro's 81.0% is the deliberately weakened figure: Moonshot removed cyber-related training data, and the uncleaned checkpoint scored 90.5%. On the uncleaned comparison K3.1 Pro leads; but the downloadable release is 81.0%.

Why aren't Qwen4-Max and GLM-5.3-Flash compared?

They lack comparable independent scores: no verified AA Index v5.0 same-version score, nor LMArena or LiveBench scores on a comparable basis, in this research round. This site won't invent numbers; the comparison comes once the scores arrive.

ModelBrief's take: the "best open weights" title fits K3.1 Pro — but only on the overall index and agent practice. On cyber, DeepSeek-V4.1-Flash's 85.7% beats its released 81.0%, the price of the vendor's deliberate weakening. Don't pick a model by its title; pick by task: K3.1 Pro for general agents, DeepSeek for cyber, Hy4 for budget-sensitive work.

Reference: Artificial Analysis — Index v5.0, Cyber Index, CyberGym, Terminal-Bench 4.0 · Moonshot — kimi.com tech blog (xdev-bench 90.5% to 81.0%) (rewritten and organized by ModelBrief; facts per the boards and the official blog)