Benchmark Almanac

LEXam-hard

Inkling 40.8% leads since September 2026
Not in model cards

Agents with legal research tools answer the 518 LEXam open questions that strong open models score lowest on. Questions answerable from memory or unanswerable from their text are removed.

A filtered copy of the LEXam open_question test split, joined back to LEXam by id.

Released
September 2026
Kind
Law
Level
Research frontier
Scored
Graded by a model
Answers
Open-ended
Tasks
All public
Strongest model tested
Kimi K3 #24 on the capability index
Last model added
July 2026
Scores refreshed
November 12th, 2026
16
Entries on its own leaderboard, 15 models we track

Inkling leads with 40.8% among models released by July 2026

Best score per entry from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders.

The newest model with a score here came out in July 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

16 entries. Scroll sideways for the rest.

  1. 40.8 Inkling
  2. 38.9 DeepSeek-V4-Pro
  3. 38.2 DeepSeek-V4-Flash
  4. 37.6 gpt-oss-120b
  5. 37.5 GLM-5.2
  6. 37.4 Gemma 4 31B IT
  7. 37.1 MiniMax-M3
  8. 36.7 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4
  9. 36.2 Kimi K2.6
  10. 31.9 Mistral Medium 3.5
  11. 29.5 Kimi K3
  12. 26.8 Llama 3.3 70B Instruct
  13. 24.4 Apertus 1.5 70B
  14. 23.9 Qwen3.5-35B-A3B
  15. 18.4 Olmo 3.1 32B Think
  16. 15.5 LFM2.5-8B-A1B
  1. 1 Inkling DeepSeek-R1-0528 judge; mean of German and English means 40.8%
  2. 2 DeepSeek-V4-Pro DeepSeek-R1-0528 judge; mean of German and English means 38.9%
  3. 3 DeepSeek-V4-Flash DeepSeek-R1-0528 judge; mean of German and English means 38.2%
  4. 4 gpt-oss-120b DeepSeek-R1-0528 judge; mean of German and English means 37.6%
  5. 5 GLM-5.2 DeepSeek-R1-0528 judge; mean of German and English means 37.5%
  6. 6 Gemma 4 31B IT DeepSeek-R1-0528 judge; mean of German and English means 37.4%
  7. 7 MiniMax-M3 DeepSeek-R1-0528 judge; mean of German and English means 37.1%
  8. 8 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 DeepSeek-R1-0528 judge; mean of German and English means 36.7%
  9. 9 Kimi K2.6 DeepSeek-R1-0528 judge; mean of German and English means 36.2%
  10. 10 Mistral Medium 3.5 DeepSeek-R1-0528 judge; mean of German and English means 31.9%
  11. 11 Kimi K3 DeepSeek-R1-0528 judge; mean of German and English means 29.5%
  12. 12 Llama 3.3 70B Instruct DeepSeek-R1-0528 judge; mean of German and English means 26.8%
  13. 13 Apertus 1.5 70B DeepSeek-R1-0528 judge; mean of German and English means 24.4%
  14. 14 Qwen3.5-35B-A3B DeepSeek-R1-0528 judge; mean of German and English means 23.9%
  15. 15 Olmo 3.1 32B Think DeepSeek-R1-0528 judge; mean of German and English means 18.4%
  16. 16 LFM2.5-8B-A1B DeepSeek-R1-0528 judge; mean of German and English means 15.5%

Source: Official LEXam-hard leaderboard

#EntryScoreSettingDateSource
1 Inkling 40.8% DeepSeek-R1-0528 judge; mean of German and English means Jul 2026 Official leaderboard
2 DeepSeek-V4-Pro 38.9% DeepSeek-R1-0528 judge; mean of German and English means Apr 2026 Official leaderboard
3 DeepSeek-V4-Flash 38.2% DeepSeek-R1-0528 judge; mean of German and English means Apr 2026 Official leaderboard
4 gpt-oss-120b 37.6% DeepSeek-R1-0528 judge; mean of German and English means Aug 2025 Official leaderboard
5 GLM-5.2 37.5% DeepSeek-R1-0528 judge; mean of German and English means Jun 2026 Official leaderboard
6 Gemma 4 31B IT 37.4% DeepSeek-R1-0528 judge; mean of German and English means Apr 2026 Official leaderboard
7 MiniMax-M3 37.1% DeepSeek-R1-0528 judge; mean of German and English means Jun 2026 Official leaderboard
8 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 36.7% DeepSeek-R1-0528 judge; mean of German and English means Jun 2026 Official leaderboard
9 Kimi K2.6 36.2% DeepSeek-R1-0528 judge; mean of German and English means Apr 2026 Official leaderboard
10 Mistral Medium 3.5 31.9% DeepSeek-R1-0528 judge; mean of German and English means Apr 2026 Official leaderboard
11 Kimi K3 29.5% DeepSeek-R1-0528 judge; mean of German and English means Jul 2026 Official leaderboard
12 Llama 3.3 70B Instruct 26.8% DeepSeek-R1-0528 judge; mean of German and English means Dec 2024 Official leaderboard
13 Apertus 1.5 70B 24.4% DeepSeek-R1-0528 judge; mean of German and English means Jul 2026 Official leaderboard
14 Qwen3.5-35B-A3B 23.9% DeepSeek-R1-0528 judge; mean of German and English means Feb 2026 Official leaderboard
15 Olmo 3.1 32B Think 18.4% DeepSeek-R1-0528 judge; mean of German and English means Dec 2025 Official leaderboard
16 LFM2.5-8B-A1B 15.5% DeepSeek-R1-0528 judge; mean of German and English means May 2026 Official leaderboard

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

Thinking Machines
Own leaderboard Not tested since Sep 2026
Inkling 40.8%

Source: Official LEXam-hard leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Inkling gpt-oss-120b Llama 3.3 70B Instruct 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Inkling Llama 3.3 70B Instruct

Source: Official LEXam-hard leaderboard

No lab model card reports it yet

We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.

More in Law, finance and medicine

All 179 in Law, finance and medicine
  • MedXpertQA-MM In use Models answer expert-level medical exam questions that combine images with clinical records and test results. Scoring is accuracy. Qwen 3.8 Max 80.4%
  • HealthBench Professional In use 525 physician-authored tasks from real clinician chats (care consult, documentation, research) with physician-adjudicated rubrics. Claude Sonnet 5.5 69.2%
  • HealthBench In use 5,000 realistic health conversations graded against physician-written rubrics. A length-adjusted variant penalizes verbosity. GPT-5 67.2%
  • FinSearchComp In use Agents search the web to fetch live financial data, look up history, and run multi-step investigations, across 635 expert-written questions. Scored by answer accuracy. Ling-3.1-flash 81.7%
  • KBL In use Models answer Korean legal knowledge and reasoning tasks and Korean bar exam questions. It has 7 knowledge tasks, 4 reasoning tasks and 2,510 bar exam items. Solar Pro 4 77.5%
  • KorMedMCQA In use Models answer 7,469 multiple-choice questions from Korean licensing exams for doctors, nurses, pharmacists and dentists, held from 2012 to 2024. Solar Pro 4 93.2%