K2-V2 debuts as open 70B reasoning model
Artificial Analysis has listed K2-V2, a 70-billion-parameter reasoning model with a 512k-token context window and Apache 2.0 weights, adding a permissively licensed entry to the mid-size open-model field.
K2-V2 now has a release page on Artificial Analysis, which lists the model's specifications alongside its benchmarking framework. According to the page, the release includes three models, each with its own intelligence, performance and pricing profile. The page focuses on the reasoning version and notes that a non-reasoning variant may also exist.
Here are the specifications Artificial Analysis lists:
- Parameters: 70B total
- Context window: 512k tokens, which the site compares to about 768 A4 pages
- Modalities: text in, text out
- License: Apache 2.0, with weights on Hugging Face
At 70B parameters, K2-V2 is in Artificial Analysis's "medium" open-weights class (40B–150B). The site ranks open-weights models only against others in the same size class. Reasoning models are compared with both reasoning and non-reasoning models.
The page lists the evaluations that make up its capability indexes. They include AA-Omniscience v1.1, AA-Briefcase v1.1, HLE-Diamond, AutomationBench-AA, GDP.pdf, MLCR-AA, CritPt v2 (Hard Subset), Terminal-Bench 4.0, Terminal-Bench-Science 0.1 and FrontierCode-AA 2.0.
The page excerpt does not include K2-V2's scores, cost per task, speed or latency. It also doesn't name the developer or describe how the three models in the release differ. Without those figures, it's not yet possible to say how K2-V2 compares with other medium-sized open models.