Glows.ai (@glowsai) header
Glows.ai (@glowsai) profile picture

https://teapawt.org/t.co/ciH0KKecAW is a cloud computing platform for artificial intelligence (AI) professionals. It offers a one-stop development environment for AI developers.

AI cloud
Joined October 2022
Photo by Glows.ai (@glowsai)Photo by Glows.ai (@glowsai)Photo by Glows.ai (@glowsai)Photo by Glows.ai (@glowsai)Photo by Glows.ai (@glowsai)Photo by Glows.ai (@glowsai)Photo by Glows.ai (@glowsai)Photo by Glows.ai (@glowsai)Photo by Glows.ai (@glowsai)Photo by Glows.ai (@glowsai)
What goes into running an AI product? James, CTO of Nox AI, joins Julia from Glows to break down compute, cost, and the road from demo to production. 🗓 Thu, Nov 19 · 20:00 (UTC+8) 📷 Live on X @glowsai
12
Local inference now has a stable major version. llama.cpp shipped v1.0.0, 3.5 years in. ggml shipped v1.0.0 with it. Breaking API changes will now bump the major number. Do you upgrade a local stack on day one, or wait for the first patch?
23
Releasing a GPU shouldn't mean reinstalling it all tomorrow. Before we log off: Checkpoints go to /datadrive. The environment goes into a snapshot. Tick auto-release. The instance stops billing. Next day: boot the snapshot, mount the same drive. What's your shutdown routine?
2
28
MiniMax H3-Image is open weights now. 4K images in 6 steps. Read the model card before you try it on a gaming card: DiT: 33.1B params, 66 GB in BF16 Text encoder: another 66.7 GB License: excludes EU, UK, Korea, US This is a big-memory GPU job.
3
42
Same RTX 3090. Same Qwen3.8-27B. Self-reported by one dev: Normal chat: ~133 tok/s Reusing the prompt: up to 381 Repeat question, prefix cache on: TTFT 22.4s → 0.56s Your stack matters as much as your GPU. "Which GPU fits my serving stack?" Ask us at glows.ai
1
28
One prompt to Fable 5.2: a 3D room in Blender, 13 animated shots, a 4K render on 10 GPUs in 52 minutes. The agent didn't want a server. It wanted 10 GPUs for under an hour. Agent jobs go wide and short. Running agent-driven renders? Ask us: glows.ai
1
32
Before buying a machine, test the model on your actual workload. Model quality, memory and speed all matter. Owning the box doesn't guarantee a better result. Try a rented GPU before you commit: glows.ai
1
29
Great AI starts with a run that works. Every workshop has a bench. Every builder needs one. For AI, the bench is compute: a GPU that is actually free, already set up, billed by the second. That is what Glows is. The AI workshop for builders. Build AI here. glows.ai
1
29
Mistral Large 4 is open-weight: 1T total, about 50B active. Your build picks the hardware. FP8: 1.06TB of weights. One 8×80GB node holds just 640GB. NVFP4: 610GB, made for Blackwell. Mistral’s SGLang example uses 4 GPUs. “Which build, which GPUs?” Get a quote at glows.ai
53
Money is flowing into image and video inference. Open model Wan2.2: a 5s 720p clip, under 9 min on a consumer GPU. Its A14B guide calls for 80GB on one card. Unlike chat, video is a VRAM problem first. "Which model, what resolution?" Get a quote at glows.ai
1
114
An hourly GPU rate is only half of your inference cost. The other half is how many tokens you get out of each hour. We ran four Qwen3.5 models on the same 48GB L40S, with the same settings: - 35B-A3B (MoE): 122 tok/s - 4B: 102 tok/s - 9B: 61 tok/s - 27B: 21 tok/s Same card, same hour. The fastest model put out about 6× as many tokens as the slowest, so each of its tokens cost about a sixth as much. The fastest one was not the smallest. The 35B-A3B is a Mixture-of-Experts model, so only a small slice of its weights runs for each token. It outpaced the 4B dense model and ran 5.8× faster than the 27B dense one. Parameter count alone did not predict speed. Two caveats. Each number is the best of three runs per model, and the run did not record concurrency. Batching and concurrent requests change throughput per GPU, and the gaps can shift. And speed is not quality: the bigger dense models produced better output than the smaller ones, and the MoE produced the best of all four. Our take: before you compare hourly rates, measure tok/s for your model on the card you plan to use. The rate divided by the throughput is the number that ends up on the bill. When you size an inference deployment, what do you check first: the hourly rate or the tok/s?
2
1
90
Test setup and source: Our original run, posted Oct 17, 2026: teapawt.org/glowsai/status/2111140… Setup, same thread: teapawt.org/glowsai/status/2111140… 1× NVIDIA L40S 48GB, Xeon Platinum 8558 (12 vCPU), 80GB RAM vLLM 0.26, FP8, thinking off temp 0.7, top_p 0.8, top_k 20, seeds 1/2/3; best of 3 runs per model tok/s as shown in the video: 35B-A3B (MoE) 122, 4B 102, 9B 61, 27B 21 Single stream, one request at a time. 122 ÷ 21 ≈ 5.8. Measuring tok/s for your own model? glows.ai
Replying to @glowsai
Setup for all four: 1× NVIDIA L40S 48GB, Xeon Platinum 8558 (12 vCPU), 80GB RAM vLLM 0.26, FP8, thinking off temp 0.7, top_p 0.8, top_k 20, seeds 1/2/3 Video shows the best of 3 runs per model. No edits.
30
AMD Data Center revenue: record $8.5B in Q3, up 97% y/y. Lisa Su on the call: "It is true that demand is higher than supply." When the chipmaker is short, the line forms downstream. Access is the spec. "What do you need to run in Q1?" Get a quote at glows.ai
1
51
Kimi K3.1 Pro weights are out: 2.8T total, 104B active, 1M context. Even in MXFP4, that's far more than any single GPU can hold. Open weights don't make it cheap to run. They let you choose where it runs. "How many nodes for K3.1 Pro?" Get a quote at glows.ai
1
57
The big clouds have GPUs. So why Glows? Glows.ai co-founder Sung joins Julia to answer it live. Bring your hardest questions on GPU access, pricing and what we’re building next. 🗓 Thu, Nov 5 · 8pm UTC+8 📍 Live on X 🗣 Hosted in Mandarin
1
4
74
New to Glows.ai? We turned the whole platform into 9 short videos. From launching your first GPU instance, to snapshots and persistent storage, to sharing a workspace with your team and calling everything from the SDK. About 47 minutes in total. Start at 01, or skip straight to the one you need.
1
2
60
Glows.ai Platform Tutorials 01 Overview and basic operations 02 Images 03 Datadrive 04 Snapshot 05 Bind IP 06 AutoDeploy 07 Team setup 08 Team storage sharing 09 SDK API youtube.com/playlist?list=PL… Mandarin versions are up on the channel too.
31
One H100 can prove a reply came from the claimed model. Anthropic's FlashProve prototype: ~110 tokens/s of proofs on a 400B MoE. CPU check: 17s. Cost: ~70–95x the GPU time of vLLM. Verification is GPU work now. "What would you prove?" Get a quote at glows.ai
1
4
165