James Aita (@jaita) header
James Aita (@jaita) profile picture

Building with Local AI | DGX Spark x2 Health-Tech | EdTech Need more VRAM. Father of 2. https://teapaw.org/t.co/9MPYYPGGFV 🎧 "Health AI Morning Briefing" 🍎 Podcasts & Spotify

Reddit- DGX Spark
Joined November 2008
Photo by James Aita (@jaita)Photo by James Aita (@jaita)Photo by James Aita (@jaita)Photo by James Aita (@jaita)Photo by James Aita (@jaita)Photo by James Aita (@jaita)Photo by James Aita (@jaita)Photo by James Aita (@jaita)Photo by James Aita (@jaita)Photo by James Aita (@jaita)
Pinned Tweet
Hy4 flash EXL3 3.0 bpw with engram offload. I have it serving in @vLLM and Tensorfold (recipes to come) but can’t seem to get it to do more than 16 t/s decode - would love some help from the masters @ashxhart @0xBakeer @sudoingX
4
1
6
337
GTA VI watch party - pot luck and bevs and absolute insanity. Tunes are 👌👌👌
4
124
Replying to @Teknium
@Teknium - [Feature REQUEST] my biggest issue with Hermes desktop these days - travelling. At home, I have my remote setup (LAN mdns address) - works great, then when I leave home to go anywhere, I have to manually re-enter my tailscaile address for the remote address. But you have "saved connections" - it would be great if it could ask me to automatically connected to the "secondary" remote address if the primary is unavailable (i.e. LAN-->TS when away from home). Each time I leave the house (and then return) I need to go through this dance. Seems like a simple tweak. @witcheer
2
2
200
The annoying thing is that when you open the app - it just says "can't connect" - and you can't choose to swap it to your other saved connection. There's no drop down or anything. (Would be a good fix for the mobile app as well) @fl_rn_st
1
36
My hy4-flash EXL3 3bpw quant is available here: huggingface.co/aitamedia/Hy4… @MiaAI_lab @ashxhart @vllm_project @ViC305 @0xBakeer
Hy4 flash is now open-weight, including Hy4 flash-FP8 and Hy4 flash-Base. 224B (12B active), 1M context, natively multimodal. Fast, efficient and built for everyday productivity. It's yours now. And stay tuned for Hy4 coming up next! 🤗: huggingface.co/tencent/Hy4-f… 🐙: github.com/Tencent-Hunyuan/H… 📄: hy.tencent.ai/research/hy4-f… Released under the Tencent Hy License. See full terms here: huggingface.co/tencent/Hy4-f…
3
2
19
1,422
🚨🚨🚨 IMPORTANT MODEL DROP - this has the potential to be as important as Qwen FN, DS4.1 or GLM5.3 flash - the engram architecture is similar to the above. Look out for quantizations in the next couple of days! I will work on this myself as well!
Hy4 flash is now open-weight, including Hy4 flash-FP8 and Hy4 flash-Base. 224B (12B active), 1M context, natively multimodal. Fast, efficient and built for everyday productivity. It's yours now. And stay tuned for Hy4 coming up next! 🤗: huggingface.co/tencent/Hy4-f… 🐙: github.com/Tencent-Hunyuan/H… 📄: hy.tencent.ai/research/hy4-f… Released under the Tencent Hy License. See full terms here: huggingface.co/tencent/Hy4-f…
3
165
Ok quantizers - Inkling small 2.0- 600b - people are saying it holds up at a 2q. I’m thinking an EXL3 1.95q with some offloading? Is that possible? How would one serve it? @0xSero @0xBakeer @ViC305 @ashxhart @turboderp_
1
114
This is not going to age well.
Replying to @mattt
move to cloud agents. local isn't the future
69
This one is interesting for the science folks
Replying to @sageai_official
@sageai_official open releases Sage-S2.1-Flash, a 320B-A18B multimodal MoE that runs and verifies workflows in scientific software. MIT-licensed weights.🤖modelscope.ai/models/Shangha… 🔬 10+ scientific input types. Sage SciPerception unifies molecular graphs, crystal structures, DNA/RNA sequences, spectra, and EEG in a shared representation. 🛠️ Tool use with verification. In the CT demo, it selected 139 slices from 12,487 mixed images, corrected their order after a failed consistency check, and reconstructed the skeleton in 3D Slicer. Other demos cover satellite imagery, CAD assembly checks, and EEG artifact removal. ⚡ ~1/10 the inference cost of Xinghe on the same task, per Shanghai AI Lab. 🏆 Benchmark leads on MLE-bench Lite, OfficeQA Pro, Harvey LAB, and SpreadsheetBench 2 in the lab’s comparison. 🔄 RSI. Task trajectories, tool outputs, and human verification guide the next round of model improvements.
53
Model drop - this one is very interesting. Another one for the quantizers - wondering if they can optimize it for a pair of Sparks.
新加坡公司Sapiens AI发布Agnes 2.5 Pro Theta,多项基准测试成绩接近DeepSeek V4.1 1017 近日,全新模型检查点Agnes 2.5 Pro Theta正式发布。该模型的总参数量约187.8B。 在同步公开的评测快照中,官方展示了Agnes 2.5 Pro Theta在八项基准测试中的表现,并将其与Nemotron 3.5 Super VL 120B、Qwen3.8-27B、GLM-5.3-Flash以及DeepSeek-V4.1-Flash-1017四款同类模型进行了横向对比。数据显示,Agnes 2.5 Pro Theta在其中五项主流测试中均取得了最高分。 huggingface.co/Agnes-AI/Agne…
3
134
This is a pretty good looking group if I do say so myself. Feeling good about being part of this local AI community today.
4
10
197
I’ve finally made it ! Topped the local AI benchmarking @spark_arena leaderboard after many, many runs.
3
7
221
Hey @spark_arena - the leaderboard has been frozen for at least a week - what gives ?
2
3
240
If you’re “weighting” for Qwen 4 to drop- it’s almost certain it won’t be until Nov 17 - HK Qwen summit. I suspect they will post the countdown tomorrow.
🚨重磅预警!Qwen 4.0 Max或将于11.17香港峰会正式亮相! 阿里云香港峰会11月17日登场,Qwen4成为绝对主角: 🔹更深知识储备 🔹更丰富世界理解 🔹更强编程 + Agent能力 🔹1M超长上下文 + 多模态(图/视频) 🔹专攻真实场景:复杂编码、前端开发、知识工作、专业软件工作流 目前Qwen4-Max Preview已开放(至11.25),Token Plan可直接上手体验。 通义实验室大佬将分享“从模型到业务影响”路线图,这波落地节奏很猛! 你们觉得Qwen4能直接卷翻Claude/GPT吗?评论区见! #Qwen4 #Qwen #AlibabaCloud #阿里云 #AI #大模型 #刘大一恒 #香港峰会 #通义千问
4
227
Prediction: @nvidia hires Jensen back within a year just like Apple did with Jobs.
39
My God the internet is fast. This is so well done
Whoever made this is a genius 😂😂
71
Yes
Maybe Qwen 4 was never a model release, but the friends we made along the way 🥰
1
103
Qwen always does a countdown of at least a few days before their open weights come out. So we are - I would guess - a week out at least based on their track record.
Qwen4 is coming. Deeper knowledge. Richer world understanding. Stronger coding and agentic capabilities. Built for real-world tasks — from complex coding and knowledge work to front-end builds and professional software workflows. Qwen4-Max is now in preview on Alibaba’s Token Plan. You can also get early access to the new model on QwenWork, Qoder, Qwen APP, and QwenStudio. Put it to the test on a real problem. Show us what you build. 👀 Token Plan: qwencloud.com/pricing/token-…
1
87