Albie retweeted
What a difference 9 weeks makes in the Local AI world!
Same hardware (2X @NVIDIAAI DGX Sparks) - same awesome model(GLM 5.3 Flash with NVIDIA quants) - SoTA performance today (top) vs just 9 weeks ago (bottom). Crazy gains entirely due to the community burning frontier tokens to test ideas and implement kernel optimizations with Opus 5.5 and Astra.
This test shows one of the hardest scenarios for this model - 4 concurrent varied, staggered, mostly prose requests. It purposefully interleaves prefill and decode to tax the GPU scheduling. Before you knee-jerk that "but X is faster!" - make sure you test it. :D
We’re bringing back the original XBOX boot-up animation for all XBOX Series X|S and XBOX One players, along with two additional XBOX 360 boot-up animations!
This update will be rolling out throughout the day and thank you for being part of 25 years of XBOX.
Albie retweeted
DOWNLOAD THESE 24 OPEN MODELS BEFORE MORE REPOS GET GATED OR TAKEN DOWN.
You asked for the exact links, so I put together the full local AI pack: LLMs, uncensored quants, image generation, video, voice, music and speech to text. Put them on your own disk while you still can.
2 repos in this list already block anonymous downloads behind a login gate, and Hugging Face has disabled uncensored repos with zero warning before. If a provider bans your account, raises prices or rewrites its rules next month, nobody can take away files that are already on your drive.
For all 4 LLMs below I included 3 versions: original weights, standard GGUF and abliterated GGUF with refusals removed. For Qwen3.8 Flash-Next and GLM-5.3-Flash, the uncensored GGUF repos are my own quants.
[LLMs: original, GGUF and uncensored]
Qwen3.8-27B
Original: huggingface.co/Qwen/Qwen3.8-…
GGUF: huggingface.co/unsloth/Qwen3…
Abliterated GGUF: huggingface.co/huihui-ai/Hui…
Qwen3.8 Flash-Next
Original: huggingface.co/Qwen/Qwen3.8-…
GGUF: huggingface.co/unsloth/Qwen3…
Abliterated GGUF (my Strata quants): huggingface.co/alesha-pro/Qw…
GLM-5.3-Flash
Original: huggingface.co/zai-org/GLM-5…
GGUF: huggingface.co/unsloth/GLM-5…
Abliterated GGUF (my Maya-M quants): huggingface.co/alesha-pro/GL…
DeepSeek V4.1 Flash 1017
Original: huggingface.co/deepseek-ai/D…
GGUF: huggingface.co/unsloth/DeepS…
Abliterated GGUF: huggingface.co/huihui-ai/Hui…
[Image generation]
Qwen-Image 2.1: huggingface.co/Qwen/Qwen-Ima…
MOSS-Image Base (Apache 2.0): huggingface.co/OpenMOSS-Team…
FLUX.3 Klein for ComfyUI (gated, accept terms on HF first): huggingface.co/black-forest-…
[Video generation]
Wan 2.2 Text-to-Video A14B (Apache 2.0): huggingface.co/Wan-AI/Wan2.2…
Wan 2.2 Image-to-Video A14B (Apache 2.0): huggingface.co/Wan-AI/Wan2.2…
HunyuanVideo 1.5: huggingface.co/tencent/Hunyu…
LTX-2.5 (gated, accept terms on HF first): huggingface.co/Lightricks/LT…
[Voice and TTS]
Fish Audio S2 Pro (non-commercial research license): huggingface.co/fishaudio/s2-…
Qwen3-TTS 12Hz 1.7B Base (Apache 2.0): huggingface.co/Qwen/Qwen3-TT…
[Music generation]
ACE-Step 1.5 (MIT): huggingface.co/ACE-Step/Ace-…
YuE2 ComfyUI bundle (non-commercial CC-BY-NC-4.0): huggingface.co/Comfy-Org/YuE…
[Speech recognition]
Whisper Large v3 Turbo (MIT): huggingface.co/openai/whispe…
[How to download cleanly]
Use the official Hugging Face CLI:
pip install -U huggingface_hub
hf download OWNER/REPO --local-dir ./MODEL
Do not pull every quant inside a GGUF repo by mistake. Pick 1 quant (and all of its split shards if it is split) with a file filter:
hf download OWNER/REPO --include "*Q4_K_M*" --local-dir ./MODEL
For original weights and media models, save the configs, tokenizers, text encoders, VAEs, README and LICENSE files too. For the 2 gated repos, accept the terms on the model card and run hf auth login first.
An API account can get cut off in 5 seconds. Weights on your own disk stay yours even if the lab changes its mind tomorrow.
Save this post and go fill your drives while every link still works.
I know what you need to do right now.
Go download the weights of these 4 open models and put them on your disk. Right now, even if you cannot run the big ones yet. Nobody knows where any of this goes next.
Qwen3.8-27B
Qwen3.8 Flash-Next
GLM-5.3-Flash
DeepSeek V4.1 Flash 1017
Grab the abliterated quants too while every link still works.
RIGHT NOW!
Albie retweeted
DeepSeek just avenged Old Huang after Anthropic’s decapitation strike by releasing its entire App stack for everyone to use and adapt freely via DeepSeek API.
Dario and Sam are now dead men walking as closed models can’t compete.
“Now, all the capabilities behind our official products can be freely accessed and combined through the DeepSeek API. We hope developers can use this to skip repetitive foundational work and build more diverse, more polished, and more imaginative products. We welcome you to use the DeepSeek API and create freely.”
🚨DeepSeek居然把他们整个App也开源了。。。。
刚刚 @deepseek_ai 正式发布全家桶应用API,直接把DeepSeek App里几乎所有的功能都开放给开发者了。
包括联网搜索、深度思考搜索、文件解析、阅读链接、语音交互,以及OCR、ASR、TTS等底层能力。
甚至官方还直接开源了一套完整的聊天前端Demo,支持导入DeepSeek历史对话。
也就是说,现在开发者拿着DeepSeek的API,基本就能1:1复刻一个DeepSeek App。
DeepSeek在公告最后还专门表示:「DeepSeek现有的对话产品只是较为基础的一种形态。现在,官方产品背后的各项能力均已支持通过 DeepSeek API 随心调用、自由组合,我们期待开发者能够借此省去重复的基础工作,打造出更多样、更极致、更具想象力的产品。欢迎使用 DeepSeek API,尽情创造。」
现在各家AI公司都在拼命做自己的超级入口,想方设法把用户留在自家App里。
DeepSeek倒好,直接把自己App背后的能力全部开放出来,连复刻方案都给你准备好了。
你想做一个自己的DeepSeek?可以。
你想把DeepSeek的搜索、语音、文件处理能力拆出来,重新组合成一个新产品?也可以。
甚至可以在DeepSeek已有的能力之上,做出比官方App更复杂的产品。
这才是这次发布最值得关注的地方,DeepSeek对自己在AI生态里要扮演什么角色,似乎想得很清楚。
就是下次能不能别在北京时间的周五晚十点发布了。。。。
Albie retweeted
AI 圈已经有人开始为「AI 断供」做末日准备了,甚至囤出了 10.5TB 的 AI 战略储备😂
黄仁勋离开 NVIDIA 后,Ruby on Rails 创始人 DHH 突然发了一条意味深长的推文:
「你最好提前备份几份开放权重模型,再准备一些能运行它们的基本设备。没人知道 NVIDIA 到底发生了什么,但这个世界刚刚失去了自由开放 AI 最重要的支持者之一。」
Reddit上,已经有人认真讨论起了「如果有一天开放权重模型不再发布,我们该提前囤些什么?」
结果评论区直接变成了 AI 末日生存物资展示大会。
一位叫 Kahvana 的玩家晒出了自己的 2TB NVMe 硬盘。
里面不只是几个聊天模型,而是囤了 Qwen3.8、Gemma 4、DeepSeek V4 Flash、Nemotron 等多个系列,连不同精度的量化版本都准备好了。
而且考虑得相当周全:
不仅有负责聊天和写代码的大语言模型,还有语音识别、语音合成、图片生成、Embedding、Reranker,以及 ComfyUI、MCP 工具、离线维基百科。
甚至还专门把最重要的模型复制到 NAS,并备份 llama.cpp 和 SillyTavern,确保不只是有模型权重,还能真正运行起来。
但这还不是最夸张的。
另一位叫 TheBigHODL 的玩家直接晒出了一份 10.5TB 的 AI 储备清单。
中国的 Kimi、DeepSeek、Qwen、GLM、MiniMax,美国的 Llama、GPT-OSS、Gemma,以及 Mistral 等模型,几乎把主流开放权重阵营都囤了一遍。
还不止如此。
从 Whisper 语音模型、FLUX 图片模型,到训练数据集、评测数据、完整维基百科、推理框架、量化工具,甚至一些开发工具的源码,都在他的存档范围内。
这已经不是简单地下载几个 GGUF 文件了,而是在试图给自己保存一整套能独立运转的 AI 工具箱。
他们担心的也很现实:
万一未来中国模型受到更多出口或访问限制,美国 AI 实验室又逐渐停止开放权重,今天随手就能从 Hugging Face 下载的东西,明天可能就没那么容易拿到了。
当然,这只是社区对未来的担忧,并不代表相关禁令已经出台。
但这种变化确实挺有意思。
以前本地 AI 圈比的是谁的显卡多、显存大、Token 跑得快。
现在有人已经开始比谁的硬盘大、谁的模型囤得全、谁断网之后还能继续用 AI。
以前末日生存玩家囤粮食、柴油、发电机,现在 AI 玩家开始囤模型、算力和离线知识库了。
而且仔细想想,这事还真不完全是杞人忧天。
云端 AI 再强,使用权终究取决于服务商;本地模型即使落后几个版本,只要权重、软件和硬件都在,至少还能自己运行。
Albie retweeted
Look at the contempt in the eyes of the civil servants. They are stunned that they’re being questioned “why can’t a combat engineer change a lightbulb?”.
This is an example of why we know government spending is the most wasteful and unproductive use of funds in the economy.
Hard working people pay a fortune in tax so these bureaucrats can cook up insane spending programs.
If you’re a tax payer, watch this video and consider how little they care about the money they take from you.
Albie retweeted
Australia's "eSafety" commissioner has sent several of my U.S. clients demands which, if sent by the U.S. government, would be unconstitutional.
Australians: everything that "eSafety" does, from an American POV, is censorship.
We don't tolerate it. Neither should you.
Australia’s "eSafety" Commissioner calls fears of over-censorship “a stretch.” She is also the official who would write the guidance under the Digital Duty of Care bill. The bill doesn’t define harmful content, but it does carry fines of up to A$109 million.
TikTok told the same committee that platforms would over-moderate to avoid those penalties.
She wants her rules binding and stronger than the industry’s, and told senators “We do want to have our cake and eat it too.”
reclaimthenet.org/australia-…
Albie retweeted
Blackfold is my Uncensored DeepSeek v4.1 1017 TP2 recipe and daily driver.
It’s not just your standard DS It was built for me. I wanted reliable, I wanted fast, I wanted capable, I wanted uncensored.
But I also didn’t want to waste $40,000 worth of hardware when you can get similar performance out of two Sparks.
So I put it in a neat little box packaged and gift wrapped for you. Enjoy.
github.com/Blackwellboy/Deep…
We're giving away a Hogwarts Legacy custom wrapped NVIDIA GeForce RTX 5090 to celebrate its official release with DLSS 5 🪄
To enter:
🟢 Share this post
🟢 Comment #RTXON
Terms and Conditions: nvidia.com/en-us/geforce/con…
26
Albie retweeted
Local AI will set you free.
No usage limits. No price hikes.
No model getting quietly worse overnight.
You pick the model, run it & it stays the same.
Your prompts never leave your machine.
Own your intelligence. Fully private.
Albie retweeted
RIP Patrick Stewart. What an actor. I still can’t believe he agreed to do this scene. And he didn’t even mind when I ruined every take. Legend.
Albie retweeted
Overnight autoresearch run on Strobe - making progress on inference speed, but still a very long way to go to useable inference (glm 5.3 flash official weights) on rtx6000 pro (with only 50gb free) and ~190gb ddr4. (Model weights are 360gb)
Albie retweeted
GLM-5.3-Flash EXL3 for DGX Sparks is now Abliterated 👀
- A simple flag to enable/disable it.
- Uses different weights.
- Gated - you must agree to the terms.
- Only slight loss in performance, around 6.5%
Also in this release:
- Next turn in chat is ~1s, down from ~16s
- Parallel agents no longer overwrite memory.
- Other improvements and fixes
Get it here:
mia-ai.net/models/GLM-5.3-Fl…
i don't care how you get there, but the day you secure 2x DGX Spark is the endgame of local ai. you feel frontier intelligence right in front of your desk.
if a single RTX 3090 shows you the way, 2x DGX Spark will enlighten you. secure the boxes and you'll rarely miss closed models like claude or chatgpt.
it's been about 3 months with my 2x DGX Spark and i'm not making assumptions, i've run Qwen 3.8 Flash Next, DeepSeek V4 Flash and now GLM 5.3 Flash on them, and i'm still sitting here surprised by what these two nodes can do.
you can do some crazy shit with these boxes and they sip power, the two GPUs pull about 73W together while GLM decodes.
i'm really happy with these two boxes, and i hope i can add a 3rd one if nvidia shows grace.
loading GLM-5.3 Flash on my 2x DGX Spark, 256GB of unified memory across two boxes, NVIDIA's NVFP4 weights split over the ConnectX cable on stock vLLM
numbers so far:
> 320B total, 18B active, 204GB weights
> decode, MTP off: 15 tok/s, flat to 128K
> decode, MTP on (k=3): 19.5-31.7 tok/s
> code + math, MTP on: 26.7-31.7 tok/s
> 120K context, MTP on: 30.4 tok/s
> prefill, MTP off: 1,350-1,580 tok/s
> 128K prompt: first token in 84s
> thinking on, temp 1.0, live server
working on improving it now, testing quality and all, i've been waiting to try this model, moved home, got my wifi just yesterday and today it's loaded.



































