Moonshot releases Kimi K3, a 2.8T-parameter open-weights multimodal model
Kimi K3 is a 2.8T mixture-of-experts model with 104B active parameters and 1M-token context, under the Kimi K3 License.
Tagged open-llms
StepFun releases Step 3.7 Flash, a 198B vision-language MoE with 11B active parameters
Step 3.7 Flash is a 198B sparse MoE vision-language model with about 11B active parameters and 256k context, under Apache 2.0.
IBM releases Granite 4.1 8B, an Apache 2.0 instruct model for enterprise tasks
Granite 4.1 8B is a long-context instruct model with better tool calling and instruction following, under Apache 2.0.
Mistral releases Medium 3.5, a dense 128B model with a modified MIT licence
Mistral Medium 3.5 is a dense 128B instruct, reasoning and coding model with 256k context. The licence excludes large-revenue companies.
Google DeepMind releases Gemma 4, with a 31B dense model under Apache 2.0
Gemma 4 comes in five sizes from E2B to 31B, with multimodal input, 256K context on the larger models and an Apache 2.0 licence.
NVIDIA releases Nemotron 3 Nano 30B-A3B, a hybrid Mamba-MoE reasoning model
Nemotron 3 Nano is a 30B hybrid Mamba-2 and MoE model with 3.5B active parameters, under the NVIDIA Nemotron Open Model License.
Ai2 releases Olmo 3, with a fully documented 32B Think model under Apache 2.0
Olmo 3 has 7B and 32B Instruct and Think variants. Ai2 releases code, checkpoints and training details alongside the weights.
Moonshot releases Kimi K2 Thinking, a 1T MoE thinking agent with native INT4
Kimi K2 Thinking is a 1T-parameter MoE with 32B active, 256K context and native INT4 quantisation, under a modified MIT licence.
Mistral releases Ministral 3, edge models in 3B, 8B and 14B under Apache 2.0
Ministral 3 is a family of 3B, 8B and 14B edge models with vision, in base, instruct and reasoning versions, under Apache 2.0.