AI Potluck
Model components / Fine-tuned / chat models

K2-V2

LLM360

LLM360 and MBZUAI's 70B dense model, released December 2025 as a reasoning-enhanced LLM the project calls 360-open: the weights ship alongside the complete TxT360 pretraining, mid-training and SFT data, the training code and intermediate checkpoints, so the run is reproducible. It supersedes the 65B K2 Chat.

The Hub repositories now render under the IFM organization rather than LLM360, though the declared LLM360 URLs still resolve. Verified 2026-08-13 via the model card and the arXiv abstract.

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)

360-open / fully reproducible: Apache-2.0 weights + complete TxT360 data + training code + checkpoints. A clean fully-open exemplar at 70B. The repo now renders under the org IFM rather than LLM360; the LLM360 URL still resolves.

  • https://huggingface.co/LLM360/K2-V2 recorded 2026-08-13

    `license:apache-2.0` and `"gated":false`; card links Tech Report, Training Code and Evaluation Code, and names Pretraining Data TxT360, Midtraining Data TxT360-Midas and SFT Data TxT360-3efforts; a "Using specific checkpoints" section; repo now shown under the org IFM

  • https://arxiv.org/abs/2512.06201 recorded 2026-08-13

    "K2-V2: A 360-Open, Reasoning-Enhanced LLM", submitted 5 Dec 2025, revised 26 Jan 2026

Adoption

1 high confidence
1.0

New as of Dec 2025 and still niche, like its K2 predecessor; the product is included as a clean fully open 70B exemplar rather than on popularity. The two declared SKUs sum to 2,970 downloads in the trailing 30 days (K2-V2 223; K2-V2-Instruct 2,747), which lands under 10K on the model adoption scale. The repositories now render under the org IFM, and the older LLM360 URLs still resolve.

Capability

4 medium confidence
4.0

The strongest fully open model by parameter scale (70B dense), with strong long-context and math and reasoning results, though OLMo 3 still leads on fully open evaluations. The figures come from the model card and technical report rather than an independent evaluation, hence medium confidence.

  • https://arxiv.org/abs/2512.06201 recorded 2026-08-13

    abstract - "It stands as the strongest fully open model, rivals open-weight leaders in its size class, outperforms Qwen2.5-72B and approaches the performance of Qwen3-235B"

Verified 2026-08-13