Back to Gap Map Model components / Fine-tuned / chat models

K2

open source / Overall score: 3.1

LLM360 and MBZUAI's K2 model line. K2-V2, a 70B dense model released in December 2025 as a reasoning-enhanced LLM the project calls 360-open, ships its weights alongside the complete TxT360 pretraining, mid-training and SFT data, the training code and intermediate checkpoints, so the run is reproducible; it superseded the 65B K2 Chat. K2 Horizon, released in September 2026 under Apache-2.0 in sizes up to a 375B-parameter mixture of experts with 23B active, is the current flagship.

Openness

5 high confidence

5.0

weights: open(Apache-2.0)
data: open(TxT360 pretraining + midtraining + SFT)
code: open(k2v2_train)
checkpoints: open
license: Apache-2.0(OSI)

K2-V2 ships Apache-2.0 weights, the complete TxT360 pretraining, midtraining and SFT data, training code and checkpoints, a fully reproducible open release at 70B.

https://huggingface.co/api/models/LLM360/K2-V2 recorded 2026-09-29
Hub API lists the weight files as "pytorch_model-00001-of-00006.bin" through "pytorch_model-00006-of-00006.bin" and tags the repo Apache-2.0 and ungated; the repo is served as IFM/K2-V2.
https://huggingface.co/LLM360/K2-V2 recorded 2026-09-29
Card states "Apache 2.0", links Training Code ("github.com/llm360/k2v2_train"), names datasets "Pretraining Data: TxT360", "Midtraining Data: TxT360-Midas" and "SFT Data: TxT360-3efforts", and has a "Using specific checkpoints" section.

Adoption

1 high confidence

1.0

Adoption is measured as Hugging Face downloads across the two declared checkpoints, the base and instruct models. Both see light use.

https://huggingface.co/LLM360/K2-V2 recorded 2026-09-29
Model page (now served as IFM/K2-V2) whose card reads "K2-V2 is our most capable fully open model to date" shows a "Downloads last month" count of 388.
https://huggingface.co/LLM360/K2-V2-Instruct recorded 2026-09-29
Model page (now served as IFM/K2-V2-Instruct) whose card reads "fully open model to date, and one of the strongest open-weight models" shows a "Downloads last month" count of 340.

Capability

4 medium confidence

4.0

This score reads K2 Horizon, the line's current flagship, which ships under Apache-2.0 with its pretraining and mid-training data. It is among the stronger open-weights models on the Artificial Analysis index and level there with Gemini 3.1 Pro, but well behind the closed leading group, so it scores level with Gemini Pro.

https://artificialanalysis.ai/leaderboards/models recorded 2026-10-08
The leaderboard's embedded model records, ranked by intelligenceIndex, put Claude Opus 5.5 (Max) first at 57.62, then Claude Sonnet 5.5 (Max) 56.0, Claude Fable 5.1 (Max) 53.35, GPT-6 Astra (Max) 52.67, Gemini 4 Argon (High) 52.56, GPT-6.1 Sol (Max) 51.83 and Muse Spark 1.3 (Max) 48.09 among models not marked deprecated; Grok 4.7 (Xhigh) follows at 46.45, Gemini 3.8 Flash (High) at 40.93, K2 Horizon 375B A23B at 30.5 and Gemini 3.1 Pro Preview at 29.72. The digest drifts between fetches because the page carries live pricing and throughput figures.
https://artificialanalysis.ai/models/k2-horizon-375b-a23b recorded 2026-10-08
"K2 Horizon 375B A23B scores 31 on the Artificial Analysis Intelligence Index, placing it well above average among comparable models (median: 18)"; Intelligence rank "# 14 / 117" among open-weights models.
https://huggingface.co/IFM/K2-Horizon-375B-A23B recorded 2026-10-08
Model card tagged "license:apache-2.0", "dataset:IFM/K2-Horizon-Pretrain-Data" and "dataset:IFM/K2-Horizon-Midtrain-Data"; its first benchmark column (# Params 375B, activated 23B) reports "Terminal-Bench 2.1 Agentic terminal use 70.2", "SWE Bench Pro (strict) Software engineering 42.6" and "GPQA Diamond Graduate-level science QA 87.3".

Verified 2026-09-29