The giants: open models you (probably) can't run at home.

The largest model Alibaba has published, and the top trending release on the Hub at launch. Same generation as the Apache-licensed 27B, under a different licence.

Size

2.4T MoE (95B active)

Context

262k tokens, native multimodal

What running it yourself actually takes

Roughly 1.2 TB at 4-bit. Multi-node territory. Community GGUF conversions exist and are mostly of academic interest unless you have a cluster.

Realistic access

Alibaba Cloud's API and the usual inference providers. FP8 weights are published alongside the BF16 ones if you are renting datacenter GPUs.

Runnable sibling

Qwen3.8-27B — dense, Apache 2.0, multimodal, and in the picker.

A frontier-scale open-weight MoE whose sheer weight footprint makes it infrastructure rather than homelab software.

Size

2.8T MoE

Context

1M tokens, native multimodal

What running it yourself actually takes

Even aggressive quantization requires well over a terabyte of memory. Practical self-hosting means a multi-node datacenter GPU cluster.

Realistic access

Moonshot API or a third-party inference provider.

Runnable sibling

Kimi K2.x family for smaller agentic workloads.

Same base model as GLM-5.2, with every gain coming from post-training. Z.ai reports the strongest open-weight coding scores to date; the benchmark table on the model card is vendor-run.

Size

753B MoE

Context

1M tokens

What running it yourself actually takes

The smallest unsloth GGUF (UD-IQ1_S) is about 217 GB; UD-Q2_K_XL is 254 GB and UD-Q4_K_XL 467 GB. Even the 1-bit build needs a 256 GB machine, and anything worth running needs more.

Realistic access

Z.ai API and most large inference providers (Together, Fireworks, DeepInfra, Novita, Baseten).

Runnable sibling

GLM-5.3-Flash — 321B MoE with 18B active, MIT, around 109 GB at UD-Q2_K_XL. For a single 24 GB card, GLM-4.7-Flash is still the one in the picker.

The flagship of Xiaomi's V2.6 series, trained with one mixed reinforcement-learning run across coding, agents, vision and security. Xiaomi's own table puts it close to the closed frontier on agentic coding.

Size

1.02T MoE

Context

1M tokens, text, image, video and audio

What running it yourself actually takes

About a terabyte in the FP8 weights Xiaomi publishes, and still well over 500 GB at 4-bit. Community GGUF conversions exist; running one means a server, not a workstation.

Realistic access

Xiaomi's MiMo API platform, or a rented multi-GPU node.

Runnable sibling

MiMo-V2.6-Flash — 309B with 15B active, MIT, and in the picker for machines with 192 GB or more.

Meta's frontier-tier open-weight MoE: relatively light compute per token, but very heavy memory requirements.

Size

400B MoE (~17B active)

Context

1M tokens

What running it yourself actually takes

Roughly hundreds of gigabytes at practical quantizations; a single consumer GPU is not enough.

Realistic access

Hosted inference or Meta-supported API access.

Runnable sibling

Llama 4 Scout is smaller but still workstation-class.

Frontier-scale DeepSeek model for code, math and agents under a permissive license.

Size

1.6T MoE (49B active)

Context

1M tokens

What running it yourself actually takes

The full checkpoint needs several hundred gigabytes to more than a terabyte depending on precision and runtime overhead.

Realistic access

DeepSeek API or third-party hosting.

Runnable sibling

DeepSeek V4 Flash — 304B-A13B, MIT, and in the picker for machines with 128 GB or more.