Will the highest score achieved by a Moonshot Kimi model on Humanity’s Last Exam in 2026 be 45% or higher?

- [2026-10-17] The agi.safe.ai page, which is the resolution source, shows no Kimi model in its visible static table. That table is headed "Dataset Updated: April 3rd, 2025" and lists Gemini 3 Pro at 38.3% as its top score. The page also has a JavaScript-loaded "Loading HLE results…" section that I could not read. I could not confirm whether the official leaderboard lists any Kimi model. https://agi.safe.ai/ - [2026-09-22] The agi.safe.ai news feed has a new entry: HLE-Diamond was released as a migration path for when frontier models hit the noise ceiling on the original HLE. This could change what the leaderboard displays. https://agi.safe.ai/ - [2026-09-22] Wikipedia's HLE results table cites Artificial Analysis, not the CAIS leaderboard. It gives Kimi K3 46.9% on the text-only subset, which is above 45%. Rivals on the same table: Claude Opus 5.5 61.4%, Gemini 4 Argon 57.1%, GPT-6 Astra 54.7%, Muse Spark 1.3 48.7%. https://en.wikipedia.org/wiki/Humanity's_Last_Exam - [2026-07-16] Moonshot released Kimi K3, a 2.8T-parameter model. Weights followed on 2026-07-27. It debuted at No. 3 on the Artificial Analysis leaderboard. https://en.wikipedia.org/wiki/Kimi_(AI) - [2026-10-16] Intelligentliving.co cites HLE of about 47% for Kimi K3, versus 44% for Muse Spark Open. This looks like the same Artificial Analysis figure. https://intelligentliving.co/muse-spark-open-vs-kimi-k3-five-points - [2026-10-16] BenchLM.ai lists Kimi K3 at 56% on HLE, with K2.6 at 34.7% and K2.5 at 30.1%. The 56% is probably a self-reported or tool-assisted figure. https://benchlm.ai/benchmarks/hle - [2026-04] Moonshot's own K2.6 model card reports HLE-Full with tools at 54.0%. Without tools it reports 34.7%. The tools figure is self-reported and above 45%. K2.5 scored 50.2% with tools and 30.1% without. https://huggingface.co/moonshotai/Kimi-K2.6 - [2025-11] K2 Thinking self-reported 44.9% on HLE text-only with tools, which is just under 45%. Artificial Analysis measured 22.3% without tools. https://www.deeplearning.ai/the-batch/kimi-k2-thinking-outperforms-proprietary-models-with-new-techniques-for-agentic-tool-use - [2026-08-16] An Intuition Labs article says the official public leaderboard's leading score was 46.44% as of that date, and that no model on it had reached 50%. This suggests the official board's top scores were in the mid-40s to about 50 by late summer. https://intuitionlabs.ai/articles/humanitys-last-exam-ai-benchmark - [2026-10] LLM-Stats, last updated October 2026, lists Kimi K2-Thinking-0905, K2.5 and K2.6, with no K3 visible in the snippet. https://llm-stats.com/benchmarks/humanity's-last-exam - [2026-09-07] Moonshot is consolidating on K3 and sunset kimi-k2.5 and moonshot-v1 on 2026-08-31. I found no announced Kimi release scheduled before 2026-12-31. https://www.morphllm.com/kimi-k3 BOTTOM LINE: Kimi K3 already scores above 45% on third-party and self-reported HLE numbers (46.9% text-only per Artificial Analysis, 54% or more with tools in Moonshot's own reports). The open question is whether the official agi.safe.ai leaderboard lists a Kimi entry at 45% or higher. Its static table shows no Kimi model, and I could not read the dynamic results. NOT FOUND: the contents of the dynamic agi.safe.ai leaderboard and any Kimi row on it; an official CAIS or Scale HLE score for any Kimi model; the Moonshot K3 model-card HLE figures (only secondary sources); any announced Kimi release or leaderboard update scheduled before 2027-01-01.

- [2026-10-17] The official leaderboard at agi.safe.ai, as fetched today, has a static table with a "Dataset Updated: April 3rd, 2025" label. Gemini 3 Pro (38.3%) is the top row, and the table lists no Moonshot/Kimi model. The page's news items run to 09/22/2026 (HLE-Diamond release). A dynamic leaderboard may be loaded by JavaScript, and my browser render did not show it. https://agi.safe.ai/ - [2025-11-06] Reference class, Kimi K2 Thinking: Moonshot's own HLE (text-only, with tools) figure was 44.9%. Heavy mode was 51.0%, and no-tools was 23.9%. Third-party no-tools runs gave 22.3% (Artificial Analysis). Vendor "with tools" numbers run about 20 points above independent no-tools numbers. https://moonshotai.github.io/Kimi-K2/thinking.html - [2026-07-16] Kimi K3 (2.8T MoE) was released. Moonshot's model card reports HLE-Full of 43.5 without tools and 56.0 with tools. Both were reported against the full-set metric. The with-tools figure already exceeds 45%. https://huggingface.co/moonshotai/Kimi-K3 - [2026-07-16] Artificial Analysis lists K3 (Max) as released July 16, 2026, with HLE as one of its Intelligence Index components. Its independently measured HLE score is not visible in the snippet I retrieved. https://artificialanalysis.ai/models/kimi-k3 - [2026-09-22] Wikipedia's HLE results table, sourced to Artificial Analysis and dated 22 Sep 2026, lists Kimi K3 at 46.9% on the text-only subset. That is above 45%, but it is not the agi.safe.ai leaderboard. It shows K3 ranked 6th, behind Claude Opus 5.5 (61.4), Gemini 4 Argon (57.1), GPT-6 Astra (54.7), MiMo-V2.6-Pro (49.4) and Muse Spark 1.3 (48.7). Treat this as a secondary summary. https://en.wikipedia.org/wiki/Humanity's_Last_Exam - [2026-09-22] The agi.safe.ai page's own results table is stale (April 2025 dataset, judge o3-mini). Its top entry was Gemini 3 Pro at 38.3%. CAIS adds models to its table slowly and selectively. It has also shown no Kimi entry in the snapshot I read, even though K2 Thinking and K3 had already been released. https://agi.safe.ai/ - [2026-10-15] Release cadence: K2 (Jul 2025), K2 Thinking (Nov 2025), K2.5 (Jan 2026), K2.6 (Apr 2026), K2.7 Code (Jun 2026), K3 (Jul 2026). A leaker says the next flagship, codenamed "x4", has finished pre-training but is still in post-training, and the report says not to expect it soon. This is a rumor from a single leaker. https://www.techcityauthority.com/2026/10/kimi-next-flagship-x4-post-training.html - [2026-10-03] Moonshot reportedly released Kimi K2.6 open-source on Oct 3, 2026, per a NewsBytes-derived report. Another source dates K2.6 to April 2026, so the dates conflict. https://aiunderstanding.org/news/moonshot-ai-launches-open-source-kimi-k2-6-with-up-to-1-000-mini-agent-helpers - [2026-09-25] K3 is still described as Moonshot's newest and strongest shipped model two months after launch. A secondary source puts the weights on Hugging Face on 27 Jul 2026 and says a technical report followed on 28 Jul. https://lu-labs.ai/blog/kimi-k3-explained-and-lu-labs-cloud - [2026-07-16] Comparators on the K3 model card, HLE-Full without tools / with tools: GPT-5.6 Sol 44.5/58.0, Claude Opus 4.8 49.8/57.9, Claude Fable 5 53.3/63.0. Frontier no-tools scores sit around 44–53%, so K3 at 43.5 no-tools is slightly below frontier. https://huggingface.co/moonshotai/Kimi-K3 - [2025-11 to 2026-10] Pattern: each Kimi flagship has beaten 45% only on vendor-reported or tool-augmented HLE. The independent text-only figure for K3 (46.9, per Wikipedia citing Artificial Analysis) is the first independent number above 45%. I found no external forecast, poll or expert estimate for this question. BOTTOM LINE: A Kimi score of at least 45% already exists on non-official sources (K3 at 56.0 with tools per Moonshot, 46.9 per the Wikipedia/Artificial Analysis table). The outside view therefore turns on resolution mechanics: whether agi.safe.ai or an accepted "official" result ever shows a Kimi entry at 45% or higher. CAIS's own table looked stale and omitted Kimi in my snapshot. A new flagship before the end of 2026 looks unlikely based on the leaker's report. NOT FOUND: a live rendered agi.safe.ai leaderboard that includes Kimi entries; Artificial Analysis's own page value for K3 HLE; any prior Polymarket/Kimi HLE strike history; any CAIS-stated K3 score.

YES means: a Moonshot Kimi model is shown at 45% or higher HLE Accuracy on the official HLE leaderboard (agi.safe.ai), or, if that source is unavailable or lacks Kimi, on an accepted official HLE result elsewhere, by 2026-12-31 11:59 PM ET. - [2026-10-17] The official CAIS dataset page on Hugging Face (cais/hle) carries an "Official Benchmark" leaderboard with a Kimi-K3 entry at 56 (rank 5 of 50). Kimi-K2.6 is also listed at 54 and is marked with an asterisk. The top entries are DeepSeek-V4.1-Flash 63.9, GLM-5.3 62.5 and Muse-Spark-Open 61. Kimi K3 therefore already has a ≥45% number on a CAIS-hosted "official benchmark" board. The rows are model-card self-reports, not CAIS-run evaluations, which is a caveat. https://huggingface.co/datasets/cais/hle - [2026-07-16] The Moonshot Kimi-K3 GitHub README gives HLE-Full as 43.5 without tools and 56.0 with tools. Both are labelled as the HLE-Full metric. This is the source of the 56 on the HF leaderboard. If a "clear equivalent metric" is read loosely, 56.0 clears 45%. https://github.com/MoonshotAI/Kimi-K3 - [2026-10-15] Scale Labs runs its own public HLE leaderboard, mirrored by BenchLM in an Oct 15 snapshot with 13 models. Earlier BenchLM mirror (Sep 30, 12 rows) lists Kimi K3 at 22.2%, well below 45% and ranked 11th. That table appears to be a harder HLE-Diamond-style view (leader GPT-6 Astra 60.6%, runner-up Claude Opus 5.5 55%). So a Scale/CAIS-run Kimi number exists but is under 45% on that variant. It is a risk for YES if the resolver uses it, though the earlier original-HLE numbers differ. https://benchlm.ai/benchmarks/scale-hle-diamond - [2026-10-17] The agi.safe.ai render today shows only the contributor and affiliation text. It does not show a visible Kimi row. The earlier static table of 38.3% for Gemini 3 Pro is stale. Resolution likely turns on how the resolver reads "official leaderboard", which has been broadened across CAIS, Scale and Hugging Face surfaces. https://agi.safe.ai/ - [2026-10-17] BenchLM lists Kimi K3 HLE as 56% (provider-reported, from the Moonshot launch blog) and 43.5% without tools. Its profile carries a "Supported" evidence label. https://benchlm.ai/models/kimi-k3 - [2026-10-13] BenchmarkList shows Kimi K3 at 46.9% on HLE (rank 20 of 478), matching the Artificial Analysis figure that Wikipedia cites. This is a second aggregator reproducing an independent measurement above 45%, with no tools. https://benchmarklist.com/models/moonshotai-kimi-k3/ - [2026-10-17] Artificial Analysis's HLE page links directly to agi.safe.ai and cais/hle as the benchmark's official references. This supports reading third-party runs as "official HLE results published elsewhere" if the primary site is unavailable. https://artificialanalysis.ai/evaluations/humanitys-last-exam - [2026-09-03] Precedent: Polymarket's Astra-on-HLE market, which uses agi.safe.ai, priced a 50%+ strike at 84%. The market creators evidently treat the site as showing frontier scores of 50–60%, and Kimi K3's 56 would sit inside that range. Related earlier Gemini markets resolved Yes on scale.com/leaderboard figures (Jan 2026). Past resolutions have followed whatever the leaderboard shows, including vendor-reported numbers. https://polymarket.com/event/openais-astra-model-humanitys-last-exam-debut - [2026-10-17] A same-batch sibling market exists: "Will the highest score achieved by a SpaceXAI Grok model on HLE in 2026 be 45% or higher?" It shows the question family is built around the 45% strike. Moonshot's K3 clears it on several sources, so a 45% strike for Kimi looks set low relative to existing figures. https://polymarket.com/breaking Most likely path to YES: the resolver reads the CAIS Hugging Face "Official Benchmark" leaderboard, or the live agi.safe.ai results, where K3 shows 56 (or 46.9 from Artificial Analysis if third-party results are accepted). No new Kimi model is needed. The prerequisites are that CAIS or the resolver accepts the self-reported with-tools figure and that the K3 entry stays on the board through year-end. The same path would also work with K2.6 at 54. BOTTOM LINE: K3 already sits above 45% on the CAIS-hosted HF official-benchmark board (56), and on third-party aggregators (46.9 independent, no tools). So YES needs only a lenient resolver reading, and no new model. The main counter-risk is a resolver who uses the agi.safe.ai table or the Scale-run view, where Kimi is absent or 22.2%. NOT FOUND: a Kimi K3 row on the live agi.safe.ai JavaScript leaderboard; the full Scale-run leaderboard rows for the original (non-Diamond) HLE; any CAIS statement on how the HF "Official Benchmark" scores are sourced.

YES means: a Moonshot Kimi model is shown at 45% or higher on the official HLE leaderboard (agi.safe.ai, or an official CAIS/Scale result if that source is unavailable) by 2026-12-31. - [2026-10-17] **Official HLE results now sit on HLE-Diamond, where Kimi K3 is far below 45%.** I fetched the Scale Labs HLE-Diamond leaderboard, which is built with CAIS. It ranks Kimi K3 14th of 17 at 22.20 ± 2.6, with calibration error 54.4. That is the lowest-ranked of the models that clear 20%, and only GLM-5.3, DeepSeek V4 Pro and Inkling are lower. For comparison, GPT-6 Astra leads at 60.60 and Claude Opus 5.5 is at 55.00. The board states that its score is overall accuracy across all 1,000 questions. https://labs.scale.com/leaderboard/hle-diamond - [2026-10-17] **The frozen question set means no re-scoring drift.** Scale says HLE-Diamond is fixed with stable question IDs, so a score from today is comparable to one from next quarter. Scale also says it expects the board to carry signal for another 6 to 12 months. Kimi K3's official number therefore won't change unless Moonshot ships a new model that is tested. https://labs.scale.com/leaderboard/hle-diamond - [2026-10-17] **The agi.safe.ai page I rendered does not show Kimi on its readable part.** The render returned mostly the contributor list, and 10,708 characters were omitted. I could not see a Kimi row on that page. The earlier static table (April 2025 dataset label, top row Gemini 3 Pro at 38.3%) also lists no Kimi. A literal reading of "official leaderboard at agi.safe.ai" therefore finds no Kimi entry at all. https://agi.safe.ai/ - [2026-09-22] **CAIS announced HLE-Diamond on its own site, listing only US frontier models.** The lastexam.ai launch post lists GPT-6 Astra 59.9%, Gemini 4 Argon 56.9%, Claude Opus 5.5 54.6%, GPT-6.1 Sol 53.2%, Claude Fable 5.1 51.6%, Muse Wonder 34.7%, Claude Sonnet 5.5 33.1% and Grok 4.7 22.8%. Kimi is not in the official launch list. CAIS's own presentation therefore contains no Kimi score near 45%. https://lastexam.ai/blog/hle-diamond - [2026-10-06] **Klu says "there is no single HLE leaderboard in 2026," which creates a resolution gap.** It splits the sources: Anthropic runs its own harness, Artificial Analysis runs a text-only subset, Scale runs the full set, and CAIS/Scale publish HLE-Diamond. It lists Kimi K3 only at 22.2 on the Scale Diamond board. Only the CAIS/Scale sources count as "official." The 46.9% (Artificial Analysis), 56% (Moonshot, with tools) and 43.5% (no tools) figures all come from non-official sources. https://klu.ai/benchmarks/humanitys-last-exam - [2026-10-05] **BenchLM mirrors the CAIS/Scale Diamond release and has Kimi K3 at 21.7% (no tools, high reasoning).** It also carries a web+code table, where GPT-6 Astra scores 82.9%. That table has 8 rows, and I saw no Kimi row on it. Even if the official board adds a tools view, Kimi isn't listed on it so far. https://benchlm.ai/benchmarks/hle-diamond and https://benchlm.ai/benchmarks/hle-diamond-web-code - [2026-10-17] **The 46.9% figure is not the official leaderboard value.** Wikipedia's results table is Artificial Analysis's text-only subset, which is a different subset and harness from the CAIS/Scale boards. Benchmarklist.com, dated about 2026-10-04, shows the same 46.9%, which it files under "Intelligence" benchmarks. The same 46.9% also sits next to Artificial Analysis's own 43.5% for K3 in the Emergent summary, which labels 43.5% as no tools. Artificial Analysis therefore doesn't give a single clean K3 figure. A resolver who strictly reads the "HLE Accuracy" label on agi.safe.ai would not use any of these. https://benchmarklist.com/models/moonshotai-kimi-k3/ and https://emergent.sh/learn/kimi-k3-benchmark - [2026-10-17] **Even on vendor numbers, K3's no-tools score is below 45%.** Moonshot reports K3 at 43.5 with no tools, and 56.0 only with tools (Together AI's K3 page). The "HLE Accuracy" metric on CAIS/Scale boards is no-tools by default, and Scale only publishes tools settings as recommendations. The 56% figure is a self-reported tools number and wouldn't be tied to the official board. https://www.together.ai/models/kimi-k3 - [2026-10-17] **The Diamond change cuts Kimi's scores sharply.** Kimi K3's Diamond score of 22.2 is about 21 points below its own 43.5 no-tools full-set figure. GPT-6 Sol falls from 44.5 on the full set (K3 model card) to 33.8, and Muse Spark 1.3 falls from 48.7 (Artificial Analysis) to 25.4. If CAIS's agi.safe.ai front page switches to Diamond, Kimi would need roughly a 23-point jump to reach 45%. https://labs.scale.com/leaderboard/hle-diamond - [2026-10-15] **No new Kimi model looks likely before year-end.** A single leaker says the next flagship, codenamed "x4", is still in post-training. The Polymarket "Kimi K4 released by December 31, 2026?" market was at about 38% when last indexed (about 75 days ago, so stale). The Polymarket "Next Kimi K model released by…?" page lists Kimi K4 as a related market. Even a release would need CAIS or Scale to test it and post a result before 2026-12-31. https://www.techcityauthority.com/2026/10/kimi-next-flagship-x4-post-training.html and https://polymarket.com/event/next-kimi-k-model-released-by-20260722183157724 - [2026-09-23] **Scale takes time to add models.** K3 was released on 2026-07-16, and the Diamond board launched on 2026-09-22 with Kimi missing from the launch table. Kimi appears only in the later list of 17 models. Scale marked 6 entries on the board "NEW" in its latest update. Kimi K2.6 and K2 Thinking were never listed in the CAIS table I read. Resolution mechanics therefore favour omission over a high score. https://labs.scale.com/leaderboard/hle-diamond - [2026-09-22] **Contrast with sibling Polymarket strikes: the 45% threshold was set against self-reported numbers.** Polymarket's HUM listing showed the "Highest Kimi score" 45%+ market at 93% (stale, with $41.7K volume). That price assumes traders count non-official sources such as the Wikipedia/Artificial Analysis 46.9 or the Moonshot-reported 56.0. If the resolver reads strictly from agi.safe.ai or Scale, the market would resolve No. https://polymarket.com/predictions/hum BOTTOM LINE: The only CAIS/Scale-run HLE scoreboard I could read scores Kimi K3 at 22.2% on Diamond, and the readable part of agi.safe.ai shows no Kimi row. A No resolution needs the resolver to accept only official CAIS/Scale values. The 46.9%, 43.5% and 56% figures come from non-official sources, and no new Kimi model looks likely before the deadline. NOT FOUND: the dynamic results section of agi.safe.ai, because the render omitted most of the page; any official CAIS/Scale Kimi score of 45% or higher; any Kimi row on the Scale web+code (tools) board; any resolution guidance from Polymarket on which source or subset counts; any confirmed Kimi release date before 2027-01-01.