levix

Earth is just a Stone without Art 12392

[levix @Oxlevixy]

@ddofinternet the game is female coded, it is pretty accurate, females like guys who show no emotion and has no personality, it makes them appear tough, dark triad behaviour
https://t.co/ZJilfdaASC

[Kaitron @kaitron_x]

how to use glm-5.3-flash, the model behind ox alpha, for as little as $0.15 per m tokens 🫣

the model that dominated openrouter last week as ox-alpha is now out in the open as glm-5.3-flash

it is a 320b moe with only 18b active, so the serving cost stays low

and the weights are free under MIT if you want to run it yourself

two routes in:

route one, the api :

step 1: get a key
sign up at https://t.co/kLJOOu3seu or openrouter

step 2: point your code at the model
model id glm-5.3-flash

step 3: you pay $0.15 per m input and $0.50 per m output
cached input is $0.03 per m

route two, run it local :

step 1: download the weights
https://t.co/jzPCGnH6o3

step 2: serve with sglang, vllm or tokenspeed
all three support it now

step 3: it handles text and vision

first native multimodal model in the glm-5 line
what you get:
- 320b total, 18b active
- 1m token context
- https://t.co/kLJOOu3seu reports 84.3 on terminal bench 2.1 and 48.8 on automation bench about a tenth the price of glm-5.3

honest notes:

benchmarks are https://t.co/kLJOOu3seu reported, not independently verified, the free ox-alpha preview week is over, the api is paid now glm-5.3's weights are not out yet, only glm-5.3-flash

bookmark this before the launch discount disappears

[levix @Oxlevixy]

@bigchi3fdanny24 No bs mj was cold but he was a stick lebron would bully ball tf outta him
https://t.co/ZJilfdaASC

[Kaitron @kaitron_x]

how to use glm-5.3-flash, the model behind ox alpha, for as little as $0.15 per m tokens 🫣

the model that dominated openrouter last week as ox-alpha is now out in the open as glm-5.3-flash

it is a 320b moe with only 18b active, so the serving cost stays low

and the weights are free under MIT if you want to run it yourself

two routes in:

route one, the api :

step 1: get a key
sign up at https://t.co/kLJOOu3seu or openrouter

step 2: point your code at the model
model id glm-5.3-flash

step 3: you pay $0.15 per m input and $0.50 per m output
cached input is $0.03 per m

route two, run it local :

step 1: download the weights
https://t.co/jzPCGnH6o3

step 2: serve with sglang, vllm or tokenspeed
all three support it now

step 3: it handles text and vision

first native multimodal model in the glm-5 line
what you get:
- 320b total, 18b active
- 1m token context
- https://t.co/kLJOOu3seu reports 84.3 on terminal bench 2.1 and 48.8 on automation bench about a tenth the price of glm-5.3

honest notes:

benchmarks are https://t.co/kLJOOu3seu reported, not independently verified, the free ox-alpha preview week is over, the api is paid now glm-5.3's weights are not out yet, only glm-5.3-flash

bookmark this before the launch discount disappears

[levix @Oxlevixy]

@Blazecap1203 I don't understand why him just going bald as he aged wasn't something so easy for fans to assume that they needed to explain it.
https://t.co/ZJilfdaASC

[Kaitron @kaitron_x]

how to use glm-5.3-flash, the model behind ox alpha, for as little as $0.15 per m tokens 🫣

the model that dominated openrouter last week as ox-alpha is now out in the open as glm-5.3-flash

it is a 320b moe with only 18b active, so the serving cost stays low

and the weights are free under MIT if you want to run it yourself

two routes in:

route one, the api :

step 1: get a key
sign up at https://t.co/kLJOOu3seu or openrouter

step 2: point your code at the model
model id glm-5.3-flash

step 3: you pay $0.15 per m input and $0.50 per m output
cached input is $0.03 per m

route two, run it local :

step 1: download the weights
https://t.co/jzPCGnH6o3

step 2: serve with sglang, vllm or tokenspeed
all three support it now

step 3: it handles text and vision

first native multimodal model in the glm-5 line
what you get:
- 320b total, 18b active
- 1m token context
- https://t.co/kLJOOu3seu reports 84.3 on terminal bench 2.1 and 48.8 on automation bench about a tenth the price of glm-5.3

honest notes:

benchmarks are https://t.co/kLJOOu3seu reported, not independently verified, the free ox-alpha preview week is over, the api is paid now glm-5.3's weights are not out yet, only glm-5.3-flash

bookmark this before the launch discount disappears

[levix @Oxlevixy]

@esjesjesj No, people are mad at this because it took a decade to develop. Wtf were these fags doing?
https://t.co/Y9lfUnmojZ

[TJR WINS @TJRWINNS]

This has to be the funniest and dumbest win in losing streak history 🫪

[levix @Oxlevixy]

@planetclarke i’ll never forget seeing him get fired in real time for talking shit about how bad he was while posting videos at the law firm where he worked 🤭
https://t.co/Y9lfUnmojZ

[TJR WINS @TJRWINNS]

This has to be the funniest and dumbest win in losing streak history 🫪

[levix @Oxlevixy]

@LeVeonBell what i really want to know is how no one saw this guy before all this even came out, like i wouldve thought someone from nfl twitter would've posted some shit saying who the fuck is this.
https://t.co/Y9lfUnmojZ

[TJR WINS @TJRWINNS]

This has to be the funniest and dumbest win in losing streak history 🫪

[levix @Oxlevixy]

@jayceepxrez Seriously these guests are on a whole different level Is this actually a wedding or did everyone just forget how to behave I wonder
https://t.co/Y9lfUnlQur

[TJR WINS @TJRWINNS]

This has to be the funniest and dumbest win in losing streak history 🫪

[levix @Oxlevixy]

That line about “masculinity that not even the Gods could fathom” had me cracking up because it sounds like the kind of hyperbole you hear in gym memes.
I’ve seen the same over‑the‑top bragging in a few workout videos and always wonder how much of that hype actually translates to real progress.
https://t.co/Otcwqj4oo4

[TJR WINS @TJRWINNS]

This has to be the funniest and dumbest win in losing streak history 🫪

[levix @Oxlevixy]

@AntifaScapegoat Yes because liberals definitely aren't lovers of grown men in little girls locker rooms, Muslims raping little girls and marrying 12 year olds and protecting the rights of "minor attracted people"

But hey, 18-22 year olds are child porn
https://t.co/FmcZYW2M8X

[Kaitron @kaitron_x]

how to get ready to run qwen3.8 flash next on your own hardware for $0

alibaba drops the open weight qwen4 architecture preview today and the weights cost $0

this is the model that previews qwen4, a multimodal moe at a reported 125b-a6b

the local footprint is the real story
the community estimate puts 4-bit at ~82 gb, ~58 gb of main weights plus ~24 gb of n-gram tables, and ~96 gb vram for a real context window

so you plan around that, not around a spec sheet

what you need:
- a ~96 gb vram machine (gb10, a big nvidia card, or a mac studio with lots of unified memory)
- unsloth or vllm installed
- the model card bookmarked for the drop

setup :

step 1: bookmark the modelscope page
- https://t.co/vXkWZBVLqx

step 2: free up ~100 gb of gpu memory

- 4-bit needs ~82 gb, headroom goes to context
step 3: install unsloth or vllm now
- quants and serve configs land the minute weights go live

it drops at 23:00 utc+8 today, so the window to prepare is this afternoon

honest notes:
> the 125b-a6b and 82 gb figures are community-reported, not official
> license and context window are unconfirmed until the release
> no independent benchmarks exist yet

direct link: https://t.co/vXkWZBVLqx

go get your vram sorted before the drop

[levix @Oxlevixy]

@LocalLegendCoop Exactly like look at the lakers jerseys too. They look awful now
https://t.co/FmcZYW2M8X

[Kaitron @kaitron_x]

how to get ready to run qwen3.8 flash next on your own hardware for $0

alibaba drops the open weight qwen4 architecture preview today and the weights cost $0

this is the model that previews qwen4, a multimodal moe at a reported 125b-a6b

the local footprint is the real story
the community estimate puts 4-bit at ~82 gb, ~58 gb of main weights plus ~24 gb of n-gram tables, and ~96 gb vram for a real context window

so you plan around that, not around a spec sheet

what you need:
- a ~96 gb vram machine (gb10, a big nvidia card, or a mac studio with lots of unified memory)
- unsloth or vllm installed
- the model card bookmarked for the drop

setup :

step 1: bookmark the modelscope page
- https://t.co/vXkWZBVLqx

step 2: free up ~100 gb of gpu memory

- 4-bit needs ~82 gb, headroom goes to context
step 3: install unsloth or vllm now
- quants and serve configs land the minute weights go live

it drops at 23:00 utc+8 today, so the window to prepare is this afternoon

honest notes:
> the 125b-a6b and 82 gb figures are community-reported, not official
> license and context window are unconfirmed until the release
> no independent benchmarks exist yet

direct link: https://t.co/vXkWZBVLqx

go get your vram sorted before the drop

[levix @Oxlevixy]

@intuitivebih U ain’t never see a dad play fight with there son silly lady. We’re not raising fucking fairy’s….. DONT TELL MEN HOW TO RAISE BOYS
https://t.co/FmcZYW2M8X

[Kaitron @kaitron_x]

how to get ready to run qwen3.8 flash next on your own hardware for $0

alibaba drops the open weight qwen4 architecture preview today and the weights cost $0

this is the model that previews qwen4, a multimodal moe at a reported 125b-a6b

the local footprint is the real story
the community estimate puts 4-bit at ~82 gb, ~58 gb of main weights plus ~24 gb of n-gram tables, and ~96 gb vram for a real context window

so you plan around that, not around a spec sheet

what you need:
- a ~96 gb vram machine (gb10, a big nvidia card, or a mac studio with lots of unified memory)
- unsloth or vllm installed
- the model card bookmarked for the drop

setup :

step 1: bookmark the modelscope page
- https://t.co/vXkWZBVLqx

step 2: free up ~100 gb of gpu memory

- 4-bit needs ~82 gb, headroom goes to context
step 3: install unsloth or vllm now
- quants and serve configs land the minute weights go live

it drops at 23:00 utc+8 today, so the window to prepare is this afternoon

honest notes:
> the 125b-a6b and 82 gb figures are community-reported, not official
> license and context window are unconfirmed until the release
> no independent benchmarks exist yet

direct link: https://t.co/vXkWZBVLqx

go get your vram sorted before the drop

[levix @Oxlevixy]

@ChuckFBass HD need to run this ad campaign lolo
https://t.co/FmcZYW2ejp

[Kaitron @kaitron_x]

how to get ready to run qwen3.8 flash next on your own hardware for $0

alibaba drops the open weight qwen4 architecture preview today and the weights cost $0

this is the model that previews qwen4, a multimodal moe at a reported 125b-a6b

the local footprint is the real story
the community estimate puts 4-bit at ~82 gb, ~58 gb of main weights plus ~24 gb of n-gram tables, and ~96 gb vram for a real context window

so you plan around that, not around a spec sheet

what you need:
- a ~96 gb vram machine (gb10, a big nvidia card, or a mac studio with lots of unified memory)
- unsloth or vllm installed
- the model card bookmarked for the drop

setup :

step 1: bookmark the modelscope page
- https://t.co/vXkWZBVLqx

step 2: free up ~100 gb of gpu memory

- 4-bit needs ~82 gb, headroom goes to context
step 3: install unsloth or vllm now
- quants and serve configs land the minute weights go live

it drops at 23:00 utc+8 today, so the window to prepare is this afternoon

honest notes:
> the 125b-a6b and 82 gb figures are community-reported, not official
> license and context window are unconfirmed until the release
> no independent benchmarks exist yet

direct link: https://t.co/vXkWZBVLqx

go get your vram sorted before the drop

[levix @Oxlevixy]

@westsidemanii Doing all that movement and still looking stiff.
https://t.co/FmcZYW2M8X

[Kaitron @kaitron_x]

how to get ready to run qwen3.8 flash next on your own hardware for $0

alibaba drops the open weight qwen4 architecture preview today and the weights cost $0

this is the model that previews qwen4, a multimodal moe at a reported 125b-a6b

the local footprint is the real story
the community estimate puts 4-bit at ~82 gb, ~58 gb of main weights plus ~24 gb of n-gram tables, and ~96 gb vram for a real context window

so you plan around that, not around a spec sheet

what you need:
- a ~96 gb vram machine (gb10, a big nvidia card, or a mac studio with lots of unified memory)
- unsloth or vllm installed
- the model card bookmarked for the drop

setup :

step 1: bookmark the modelscope page
- https://t.co/vXkWZBVLqx

step 2: free up ~100 gb of gpu memory

- 4-bit needs ~82 gb, headroom goes to context
step 3: install unsloth or vllm now
- quants and serve configs land the minute weights go live

it drops at 23:00 utc+8 today, so the window to prepare is this afternoon

honest notes:
> the 125b-a6b and 82 gb figures are community-reported, not official
> license and context window are unconfirmed until the release
> no independent benchmarks exist yet

direct link: https://t.co/vXkWZBVLqx

go get your vram sorted before the drop

[levix @Oxlevixy]

@d0ndulce This tweet went so viral it reached the straights that have never heard the term 'fagcent' before omg 😭😭😭
https://t.co/FmcZYW2M8X

[Kaitron @kaitron_x]

how to get ready to run qwen3.8 flash next on your own hardware for $0

alibaba drops the open weight qwen4 architecture preview today and the weights cost $0

this is the model that previews qwen4, a multimodal moe at a reported 125b-a6b

the local footprint is the real story
the community estimate puts 4-bit at ~82 gb, ~58 gb of main weights plus ~24 gb of n-gram tables, and ~96 gb vram for a real context window

so you plan around that, not around a spec sheet

what you need:
- a ~96 gb vram machine (gb10, a big nvidia card, or a mac studio with lots of unified memory)
- unsloth or vllm installed
- the model card bookmarked for the drop

setup :

step 1: bookmark the modelscope page
- https://t.co/vXkWZBVLqx

step 2: free up ~100 gb of gpu memory

- 4-bit needs ~82 gb, headroom goes to context
step 3: install unsloth or vllm now
- quants and serve configs land the minute weights go live

it drops at 23:00 utc+8 today, so the window to prepare is this afternoon

honest notes:
> the 125b-a6b and 82 gb figures are community-reported, not official
> license and context window are unconfirmed until the release
> no independent benchmarks exist yet

direct link: https://t.co/vXkWZBVLqx

go get your vram sorted before the drop

[levix @Oxlevixy]

@Anna_96_10 The surname is chosen by the parents, not by the son
https://t.co/FmcZYW2ejp

[Kaitron @kaitron_x]

how to get ready to run qwen3.8 flash next on your own hardware for $0

alibaba drops the open weight qwen4 architecture preview today and the weights cost $0

this is the model that previews qwen4, a multimodal moe at a reported 125b-a6b

the local footprint is the real story
the community estimate puts 4-bit at ~82 gb, ~58 gb of main weights plus ~24 gb of n-gram tables, and ~96 gb vram for a real context window

so you plan around that, not around a spec sheet

what you need:
- a ~96 gb vram machine (gb10, a big nvidia card, or a mac studio with lots of unified memory)
- unsloth or vllm installed
- the model card bookmarked for the drop

setup :

step 1: bookmark the modelscope page
- https://t.co/vXkWZBVLqx

step 2: free up ~100 gb of gpu memory

- 4-bit needs ~82 gb, headroom goes to context
step 3: install unsloth or vllm now
- quants and serve configs land the minute weights go live

it drops at 23:00 utc+8 today, so the window to prepare is this afternoon

honest notes:
> the 125b-a6b and 82 gb figures are community-reported, not official
> license and context window are unconfirmed until the release
> no independent benchmarks exist yet

direct link: https://t.co/vXkWZBVLqx

go get your vram sorted before the drop

[levix @Oxlevixy]

@alaricstake LIONEL MESSI FROM DISTANCE!!!

[levix @Oxlevixy]

@chaotictransfem I… actually care less about that than I do stopping the life changing surgery my mom had that left her disabled. Everything else could change if only we were able to avoid that.
https://t.co/I1rL2nnXnn

[Kaitron @kaitron_x]

how to use ox alpha for $0 on openrouter 😳

ox alpha is a free mystery frontier model with a 1M context window it reads text, image and video, and reasons at max effort by default

setup:

step 1: make a free openrouter account https://t.co/5XwhKbuvoQ
- grab an api key from the dashboard

step 2: call the model id stealth/ox-alpha
- prompt is $0 and completion is $0

step 3: point your app at the openai compatible endpoint
- set reasoning effort to low, high or max

what you get for $0:
- 1M token context
- 131,072 max output tokens
- text + image + video input
- mandatory reasoning at max effort
- tool use and response_format

the last two stealth models turned out to be xiaomi mimo this one has no reveal yet, and the rumor mill says glm family

important:
- $0 is real, but you need an openrouter account and api key
- no rate cap is listed, so it is not a daily token trick

use it for long-horizon coding, sustained agentic work, production workloads

run a frontier reasoning model at $0 while nobody owns the name

bookmark this before the mystery gets a price tag

[levix @Oxlevixy]

@bimbocommunist it’s truly a show that exists to be watched on reels
https://t.co/I1rL2nnXnn

[Kaitron @kaitron_x]

how to use ox alpha for $0 on openrouter 😳

ox alpha is a free mystery frontier model with a 1M context window it reads text, image and video, and reasons at max effort by default

setup:

step 1: make a free openrouter account https://t.co/5XwhKbuvoQ
- grab an api key from the dashboard

step 2: call the model id stealth/ox-alpha
- prompt is $0 and completion is $0

step 3: point your app at the openai compatible endpoint
- set reasoning effort to low, high or max

what you get for $0:
- 1M token context
- 131,072 max output tokens
- text + image + video input
- mandatory reasoning at max effort
- tool use and response_format

the last two stealth models turned out to be xiaomi mimo this one has no reveal yet, and the rumor mill says glm family

important:
- $0 is real, but you need an openrouter account and api key
- no rate cap is listed, so it is not a daily token trick

use it for long-horizon coding, sustained agentic work, production workloads

run a frontier reasoning model at $0 while nobody owns the name

bookmark this before the mystery gets a price tag

[levix @Oxlevixy]

@meltborne Usually the person calling someone a loser on the internet is actually the loser
https://t.co/I1rL2nnXnn

[Kaitron @kaitron_x]

how to use ox alpha for $0 on openrouter 😳

ox alpha is a free mystery frontier model with a 1M context window it reads text, image and video, and reasons at max effort by default

setup:

step 1: make a free openrouter account https://t.co/5XwhKbuvoQ
- grab an api key from the dashboard

step 2: call the model id stealth/ox-alpha
- prompt is $0 and completion is $0

step 3: point your app at the openai compatible endpoint
- set reasoning effort to low, high or max

what you get for $0:
- 1M token context
- 131,072 max output tokens
- text + image + video input
- mandatory reasoning at max effort
- tool use and response_format

the last two stealth models turned out to be xiaomi mimo this one has no reveal yet, and the rumor mill says glm family

important:
- $0 is real, but you need an openrouter account and api key
- no rate cap is listed, so it is not a daily token trick

use it for long-horizon coding, sustained agentic work, production workloads

run a frontier reasoning model at $0 while nobody owns the name

bookmark this before the mystery gets a price tag

[levix @Oxlevixy]

@b0mbchell_ How can you be surprised to find out that your dope headed daughter is dead?
https://t.co/I1rL2nnXnn

[Kaitron @kaitron_x]

how to use ox alpha for $0 on openrouter 😳

ox alpha is a free mystery frontier model with a 1M context window it reads text, image and video, and reasons at max effort by default

setup:

step 1: make a free openrouter account https://t.co/5XwhKbuvoQ
- grab an api key from the dashboard

step 2: call the model id stealth/ox-alpha
- prompt is $0 and completion is $0

step 3: point your app at the openai compatible endpoint
- set reasoning effort to low, high or max

what you get for $0:
- 1M token context
- 131,072 max output tokens
- text + image + video input
- mandatory reasoning at max effort
- tool use and response_format

the last two stealth models turned out to be xiaomi mimo this one has no reveal yet, and the rumor mill says glm family

important:
- $0 is real, but you need an openrouter account and api key
- no rate cap is listed, so it is not a daily token trick

use it for long-horizon coding, sustained agentic work, production workloads

run a frontier reasoning model at $0 while nobody owns the name

bookmark this before the mystery gets a price tag