Code1T-A32B official

kimi-k2

by Moonshot AI

Kimi K2: a 1-trillion-parameter MoE (32B active) tuned for agentic work. The top of the range: big servers, or very aggressive quants.

#moe #agentic #code #frontier

arcflare pull unsloth/Kimi-K2-Instruct-GGUF:UD-TQ1_0

[ 01 ] // quickstart

┌─ arcflare — zsh // tty1

[ 02 ] // readme

kimi-k2

Kimi K2: a 1-trillion-parameter MoE (32B active) tuned for agentic work. The top of the range: big servers, or very aggressive quants.

Get it

arcflare pull unsloth/Kimi-K2-Instruct-GGUF:UD-TQ1_0

arcflare run kimi-k2-instruct

ArcFlare reads the GGUF header and loads it at the largest context your free VRAM can actually hold. Use arcflare fit to see that before downloading anything else, or arcflare use opencode to code with it.

Highlights

1T total parameters, 32B active per token
Strong at tool use and long multi-step tasks
Pull the smallest dynamic quant first. Even that needs a very large-memory machine

Sizes

1t-a32b · 245GB+

Other sizes live in sibling repos: swap the size in the repo name when pulling.

[ 03 ] // relatedmore →

glm4.6

Z.ai

Code

GLM-4.6: Z.ai's flagship ~357B MoE for coding and agents, with a 200K context. Server- or big-workstation class.

size: 357B
vram: ~220 GB
license: MIT

9.2× a 24 GB card

#code #agentic #moe

glm4.7-flash

Z.ai

Code

GLM-4.7-Flash: a 30B-A3B MoE built for agentic coding. Only ~3B parameters active per token, so it runs at small-model speed.

size: 30B-A3B
vram: ~20 GB
license: MIT

83% of a 24 GB card

#code #agentic #moe

qwen3-coder-next

Alibaba Qwen

Code

Qwen3-Coder-Next: an 80B MoE with only 3B active, a 256K context, and training aimed at agent harnesses like Claude Code and Cline.

size: 80B-A3B
vram: ~52 GB
license: Apache-2.0

2.2× a 24 GB card

#code #agentic #moe