The README that ships with them carries a comparison table: K3 against Claude Fable 5, GPT-5.6 Sol, Claude Opus 4.8, GPT-5.5 and GLM-5.2, every model at its maximum setting, every score Moonshot’s own. Forty-five benchmarks sit in four sections — reasoning and knowledge, coding, agentic work, vision — and forty of them report a score for all three of K3, Fable 5 and Sol.

Benchmark scores

Each of the three finishes first somewhere, and by very different amounts. Fable 5 takes HLE-Full by nearly ten points and BabyVision by close to five. Sol takes CritPt at 32.3, where K3 scores 23.4. K3 takes MCPMark-Verified at 94.5 and SWE-Marathon at 42.0. FrontierSWE spreads the three over fifteen points, from Fable 5 at 86.6 down to Sol at 71.3.

Other rows barely move at all. SpreadsheetBench 2 has K3 at 34.8 and Fable 5 at 34.7. OfficeQA Pro has K3 at 63.3 and Sol at 63.2. Terminal-Bench 2.1 puts all three within 0.8 of each other, at 88.3, 88.0 and 88.8.

Counted as wins and losses, the answer moves with the comparator. Against Sol, K3 is ahead on 27 of the forty comparable rows. Against Fable 5 it is ahead on 15, behind on 24, and level on one — ZeroBench (pass@5), where both score 23.0. Section by section against Fable 5, that runs two-two in reasoning, three-six in coding, seven-twelve in agentic and three-four in vision, with the tie falling in vision.

The size of a lead moves with the comparator too. MCPMark-Verified gives K3 7.1 points on Fable 5 and 1.6 on Sol. BrowseComp gives it 3.2 and 0.8. AA-LCR gives it 4.7 and 1.0, and AA-LCR is the only row in the reasoning section where K3 finishes above Sol. SWE-Marathon puts K3 seven points above Fable 5’s 35.0, with Claude Opus 4.8, also in the table, at 40.0.

Twenty-two of the forty-five rows are agentic work, so most of any win-and-loss count is decided there: browser search, MCP and tool calls, OS control, spreadsheets and office documents, and a run of domain agents — Harvey Lab-AA, CorpFin v2, Finance Agent v2, Legal Research Bench. Two of those rows report Elo ratings instead of percentages. The 61 points between K3’s 1686 and Fable 5’s 1747 on GDPval-AA v2 are not the same unit as the 61 points that would separate two percentage scores, and nothing in the table converts between them.

Price and speed

K3 lists at $3 per million input tokens and $15 per million output, with cached input at $0.30. Fable 5 is $10 and $50 — the same rate Anthropic kept for Pro and Team Standard in July while folding Fable 5 back into the top plans. Sol is $5 and $30. Against those two, K3 is the cheap one.

A comment on the Hacker News launch thread reads:

pricing is $3/$15 for 1M tokens (cache $0.3), which is extremely high for a Chinese open-weight model

The thread under it drifted into whether per-token prices can be compared at all. Tokenizers differ, so the same page of text becomes a different number of tokens depending on which model is counting it, and more than one person wanted a price per page or per byte instead.

Artificial Analysis measures K3’s output at 34.9 tokens per second, 57th among the 101 models on its list, running on hosted infrastructure. One Hacker News post is titled Run Kimi K3 using 29 GB of RAM at 0.50 tok/s — thirty tokens a minute, which is what the downloadable weights come to on a machine that size.

The model is under three weeks old and a version bump can move any of these numbers. If you have run K3 next to Fable 5 or Sol on the same job, which rows held up? And if something here is wrong, tell me and I will fix it.

開源模型發表的時候通常會附一張跟閉源模型的比較表。K3 這張比的對象是 Claude Fable 5 跟 GPT-5.6 Sol，下面用到的數字都來自 Moonshot README 裡的對照表，表上所有模型都跑 max 設定。

對照表上的位置

表分成四塊：推理與知識、寫程式、agent 任務、視覺。

整體的方向一致。對上 Sol，K3 多數列領先；對上 Fable 5，多數列落後。位置大致就在兩者之間，比較靠近 Sol。

有意思的是 Moonshot 的講法比這張表保守。官方部落格上只說 K3 的整體表現仍然落在 Fable 5 跟 Sol 後面，達到的是 frontier level，把兩家並列放在前面。但表上 K3 對 Sol 是多數列領先的。

這份文件裡還有一列容易被跳過的 Kimi Code Bench 2.0。Kimi Code Bench 2.0 是 Moonshot 自家使用的 benchmark，K3 拿 72.9，Fable 5 是 76.9。同樣也可以拿來參考，但是評分細節我們可能就不知道了。

不過整體排名我們大概能看出 K3 落在哪一區，可是還不夠拿來直接判定工具好壞。比較值得關注的是它贏在哪幾列、又輸在哪幾項。

輸贏的比較

下面這幾項是差距比較明顯的，分數都取自同一張對照表：

先看 SWE-Marathon。K3 是 42.0，Fable 5 拿 35.0，Sol 是 39.0。

這波是長時間、多回合的任務。題目不會跑一輪就結束，模型得要根據前一輪的結果繼續接著往下改，而且任務拖得愈長，前面留下來的資訊就愈重要。

換成平常會碰到的工作，大概是把整個 repo 的 lint 修過一遍，或者讀完幾十份文件再整理成一張表。每一個步驟和任務單獨處理都不難，難的是好幾個步驟加在一起後，是否還能完成任務不走歪。

同個方向上還有三個分數，括號裡是 Fable 5 跟 Sol。長上下文的 AA-LCR 是 74.7（70.0、73.7），工具操作的 MCPMark 94.5（87.4、92.9），BrowseComp 91.2（88.0、90.4）。

這四項來自不同的 benchmark 家族，題目長相也不一樣，可是排出來的順序一樣。共通的地方是任務都要撐得久、中間要記得前面發生過什麼，而且多半得接外部工具。K3 比較好的地方大多在這個類別。

相較之下，輸的地方集中在推理與知識。

CritPt 是 K3 23.4、Fable 5 28.6、Sol 32.3；HLE-Full 是 K3 43.5、Fable 5 53.3、Sol 44.5。兩列的對手還不一樣，CritPt 對兩邊都輸，HLE-Full 主要輸給 Fable 5，跟 Sol 只差一分。

這個差別呈現的資訊不只是分數高低。如果是較長流程的題目跑不好、跑不出來還能多跑幾輪、多給一點上下文、多接不同工具慢慢補回來；但推理補不了。上下文再長、工具再多，模型想不出來就是想不出來，能力到哪就是在哪。

價格

價格則是個重要的關注點。

K3 的數字取自 Artificial Analysis 的模型頁。換算下來大約是 Fable 5 的三分之一、Sol 的六成。同樣一筆預算，改用 K3 可以跑到 1.7 到 3.3 倍的量，這就讓開發有不同的使用和想像空間。

cache 的 0.3 美元只有一般 input 價的十分之一。如果流程會反覆讀同一批文件，第二次之後就是按這個價算，跑得愈久差距愈明顯。

輸出速度也有提供。K3 是 34.9 tok/s，在 101 個模型裡排第 57。這個速度吐一千個 token 大概要半分鐘，單次問答就感覺得到了，流程變長更明顯，任務跑上幾十輪，等待時間會顯著的變長。

可以交給它的任務

照這張表，我們可以看出 K3 適合的任務有幾個共通點：流程可以長、中間要回頭看前面做過什麼、還有去接外部工具的任務。

長流程的程式修改、資料整理，還有串 MCP 或別的工具的自動化，大致都在這個範圍。這類工作本來就要跑很多輪，三倍的價差會一路累積下去。

推理上的題目就是另一回事。如果需要推論到答案、而且要有效率的答對，Fable 5 跟 Sol 都領先不少。便宜換到的很可能就是弱一點的答案。

所以工作類型較多元的話，也不用只挑一個模型。長流程交給 K3，難題留給 Fable 5 或 Sol，可以有很多不同的搭配和嘗試。

寫在最後，模型的發展很快，說不定改天 Gemini 又彎道超車，幾個月後又是不同的世界，我們該做的就是保持彈性和學習的心態，有甚麼想討論的歡迎聯絡～