News
EveryLLM

Hybrid designs

Jamba and others that mix attention with state-space layers. 31 models on EveryLLM so far.

Hybrid models combine a few attention layers with many cheaper layers of another kind, usually state-space (Mamba) layers or forms of linear attention. Attention gives exact recall over the context, while the cheaper layers carry most of the load, cutting the memory needed for long inputs and speeding up generation. Since 2024 hybrids have moved from experiments into production models from AI21, Nvidia, IBM and Alibaba.

Jamba and after

AI21 Labs in Tel Aviv released Jamba in March 2024, mixing transformer and Mamba layers with a mixture of experts: 52 billion parameters in total, 12 billion active, and a context window of 256,000 tokens. Nvidia's Nemotron Nano 2 (August 2025) has 62 layers, of which only 6 use self-attention, and Nvidia reported up to six times the throughput of a similar-sized transformer on reasoning tasks.

Mainstream releases

IBM's Granite 4.0 family (October 2025) pairs Mamba-2 and transformer layers at roughly nine to one, and IBM reported over 70% less memory for long inputs. Alibaba's Qwen3-Next (September 2025) alternates three Gated DeltaNet layers, a form of linear attention, with one gated attention layer, in an 80-billion-parameter mixture of experts with 3 billion active. Such designs aim at long documents and agents that run for many steps.

Famous hybrid models

Further reading

fnSolar Open 2 馃嚢馃嚪 South KoreaSolar Open 2 250B parameter open-weight model for agentic tasksGeneral 路 Open weights 路 2026 MiMo-V2-Flash 馃嚚馃嚦 ChinaMiMo-V2-Flash High-speed MoE language model for reasoning and agentsGeneral 路 Open weights fnNemotron Nano 2 馃嚭馃嚫 United StatesNemotron Nano 2 A 9B hybrid Mamba-Transformer model for reasoning and chat, from NVIDIA.General 路 Open weights 路 2025 fnNemotron 3 Super 馃嚭馃嚫 United StatesNemotron 3 Super NVIDIA's 120B hybrid MoE language model with 12B active parameters and 1M-token contextGeneral 路 Conditional licence 路 2026 fnEXAONE 4.0 馃嚢馃嚪 South KoreaEXAONE 4.0 Unified language model with reasoning and non-reasoning modesGeneral 路 Conditional licence 路 2025 Qwen3.8-2.4T-A95B 馃嚚馃嚦 ChinaQwen3.8-2.4T-A95B Open-weight reasoning model with 2.4 trillion parametersReasoning 路 Open weights 路 2026 Falcon-H1R-7B 馃嚘馃嚜 United Arab EmiratesFalcon-H1R-7B A 7-billion parameter reasoning model with hybrid architectureReasoning 路 Open weights 路 2026 Falcon-H1 馃嚘馃嚜 United Arab EmiratesFalcon-H1 Hybrid transformer and state-space model series by TIIGeneral 路 Open weights 路 2025 fnMiniMax-M1 馃嚚馃嚦 ChinaMiniMax-M1 Open-weight reasoning model with hybrid attention and long contextReasoning 路 Open weights 路 2025 Hunyuan-TurboS 馃嚚馃嚦 ChinaHunyuan-TurboS Hybrid Mamba-Transformer MoE model by TencentGeneral 路 API only 路 2025 imgvidNemotron 3.5 Super VL 馃嚭馃嚫 United StatesNemotron 3.5 Super VL NVIDIA's open-weight vision-language model with 120B parameters and 1M-token contextMultimodal 路 Conditional licence 路 2026 fnNemotron 3 Nano 馃嚭馃嚫 United StatesNemotron 3 Nano NVIDIA's 30B hybrid Mamba-Transformer MoE model for reasoning and agentsGeneral 路 Conditional licence 路 2025

The weekly digest

Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.