> top 8 local LLMs to run at home for December 2025 – update…

> top 8 local LLMs to run at home for December 2025 – updated

> Devstral-2-123B at 1st place

> senior engineer energy

> not flashy, just relentlessly effective

> this is the model i trust when i need results

> just “fix the thing”

> agentic coding is the real superpower

> clean multi-file edits + sane tool use

> massive 256k context actually matters

> big repos, tickets, logs, zero disorientation

> understands the project, not just the file

> dump a repo now, ask follow-ups later

> it remembers why you’re here

> absurd swe-bench for its size (and it’s not 400B+)

> generalizes well, survives messy, undocumented codebases

> closest thing to claude code

> fits on 2x rtx pro 6000

> or 8x 3090s if you’re in your home-lab era

> with the full 256k context

> the exact model i use is cyankiwi/Devstral-2-123B-Instruct-2512-AWQ-4bit

> MiniMax-M2 at 2nd place

> great for agentic workflows thanks to interleaved thinking

> very solid at coding & good at UI/design as well

> fits on 2x RTX PRO 6000

> or 8x RTX 3090s w/ full context

> the exact model i use is cyankiwi/MiniMax-M2-AWQ-4bit

> GLM-4.5-Air at 3rd place

> all-around MVP & daily driver for most tasks that aren’t pure coding

> also, when running multiple agents in parallel, this is my go-to

> uses fewer GPUs & less memory for KVCache

> which allows me to spin up more agents

> fits on a single RTX PRO 6000 or 4x RTX 3090s w/ full context

> nearly as good as big brother, however, not as power hungry

> the exact model i use is cyankiwi/GLM-4.5-Air-Derestricted-AWQ-4bit

> Qwen3-VL-235B-A22B at 4th place

> my main multimodal assistant & visual agent

> very strong with wild dense detection

> massive native context window (256k expandable to 1M)

> we are talking huge documents and long videos with full recall

> sometimes casually out-reasons text-only models

> also great teacher for auto-labeling

> comes in Thinking & Instruct variations

> fits on 2x RTX PRO 6000

> or 8x RTX 3090s w/ full context (256K)

> the exact model i use is cyankiwi/Qwen3-VL-235B-A22B-{Thinking,Instruct}-AWQ-4bit

> GLM-4.6 (REAPed) at 5th place

> the “ok fine, bring out the big brain” model

> very capabable agentic model & coder

> if M2 can’t handle something, i try this one

> fits on 2x RTX PRO 6000

> or 8x RTX 3090s w/ full context

> the exact model i use is 0xSero/GLM-4.6-REAP-218B-A32B-W4A16-AutoRound

> NVIDIA-Nemotron-3-Nano-30B-A3B at 6th place

> this one is sneaky

> looks small on paper, behaves way bigger than it is

> best-in-class SWE / agentic for its size

> consistently punches above its weight on agentic work

> strong, reliable tool use

> clean, predictable structured outputs

> comfortably beats Qwen3-30B-A3B

> absolute trivia sponge

> dumps knowledge on command

> very strong at code + math

> long-context with solid multilingual support

> reasoning-on actually reasons: pause → think → answer

> “spin up five agents without guilt” energy

> low active params = predictable memory

> great throughput

> extremely local-AI friendly

> GPT-OSS-120B at 7th place

> feels like a GPT-5 at home

> big brain, smart and consistent

> agentic, coding, but feels dry in writing

> GPT-OSS-20B at 8th place

> absolute speed monster

> holds up well at tool calling (short context preferred)

> pretty good at instruction following, quick bug fixes, and general use

> Qwen3-Coder-30B as a bonus model

> very capable at coding

> will run on 16GB VRAM quantized

> perfect when you just need something solid for autocomplete

— @TheAhmadOsman