Qualiteg Inc.

Tokyo AI company. We measure local LLMs, GPU inference, Japan's LLM rankings and AI on Raspberry Pi, then post the numbers. WireCanal, MotionVox, Bestllam.

[Qualiteg Inc. @Qualiteg_global]

Who we are, in numbers. Qualiteg is a Tokyo AI company founded in 2023. This account posts what we measure: local LLMs, GPU inference, Japan's LLM rankings, and AI on a Raspberry Pi. A short thread on what to expect.

[Qualiteg Inc. @Qualiteg_global]

NVIDIA blog, Oct 22: https://t.co/LP1VFAN1nH Model: https://t.co/lb9CJ6PfLH

[Qualiteg Inc. @Qualiteg_global]

NVIDIA runs full GLM 5.3 (753B, NVFP4) on 8 DGX Spark nodes.
- 1 TB unified memory across the cluster
- 31 tok/s decode at concurrency 1
- Qwen3.8-27B decode 11.86 to 52.58 tok/s (4.43x) with ROCEnante, per NVIDIA
Desk scale clusters are getting real.
#DGXSpark #LocalLLM

[Qualiteg Inc. @Qualiteg_global]

OpenAI docs: https://t.co/Vku75c2l1c TechCrunch, Oct 22: https://t.co/gHi1vLoWzb

[Qualiteg Inc. @Qualiteg_global]

OpenAI's Decisions API is now generally available.
- Typed answers (probability, choice, rubric score) from text, images, JSON
- About 10x faster than the Responses API, per OpenAI
- Runs on gpt-6-luna
Fast, low-cost calls for routing and classification.
#OpenAI #AIAPI

[Qualiteg Inc. @Qualiteg_global]

Anthropic: https://t.co/sMRSTHcCc8 Reuters, Oct 21: https://t.co/2Yg1MkOagb

[Qualiteg Inc. @Qualiteg_global]

Anthropic launches Claude Duet, an agent that works while you are away.
- Research preview for Claude Pro and Max in the US
- Connects to 6,000+ apps, Stripe, Gmail, Slack, Notion
- Asks approval before messages or payments
Autonomy with a human checkpoint.
#Anthropic #AIAgents

[Qualiteg Inc. @Qualiteg_global]

OpenAI: https://t.co/XDv1pnTYqO Google Cloud Blog, Oct 21: https://t.co/soLGUvGa8G

[Qualiteg Inc. @Qualiteg_global]

OpenAI brings GPT-6 Astra, Sol, and Luna to Google Cloud.
- Private preview in Vertex AI model garden
- Billed on one Google Cloud invoice via Marketplace
- Codex CLI/Desktop can run on Gemini infra today
More model choice for enterprise AI stacks.
#OpenAI #GoogleCloud

[Qualiteg Inc. @Qualiteg_global]

Google Cloud Blog, Oct 21: https://t.co/DPHngsuexQ GitHub: https://t.co/ZitFHiDlBD

[Qualiteg Inc. @Qualiteg_global]

Google and Meta open-sourced TorchTPU, native PyTorch on TPUs.
- Up to 52% Model FLOPs Utilization on Qwen 3 8B
- Near-linear scaling up to 8,192 Ironwood TPUs
- Zenteiq case, 4x+ inference throughput per 1,000 TPU v6e
More hardware for LLM serving. Link in reply.
#PyTorch #TPU

[Qualiteg Inc. @Qualiteg_global]

Claude Opus 5.5 and GPT-6 Astra tie at 168.8 on Epoch AI's index, and Claude Fable 5.1 tops Arena's Japanese category. Our Oct 22 Japanese LLM Rankings now come in two parts from multiple ranking sources, plus the Nejumi TOP 50.
https://t.co/L9hGXzKG1r
#LLM #AIBenchmarks

[Qualiteg Inc. @Qualiteg_global]

Recursive, Oct 21: https://t.co/FHR1DbmsfG Wired, Oct 21: https://t.co/98DRVpHoJg

[Qualiteg Inc. @Qualiteg_global]

Recursive says its new coding model Flint was built almost entirely by AI, not human engineers (Wired).
- 39.8B total, 2.5B active params (MoE)
- SWE-bench Verified 74.2%, Terminal-Bench 2.1 54.9%
- Runs on 1 B300 or H200 GPU, Apache-2.0
Link in reply. #AI #Coding

[Qualiteg Inc. @Qualiteg_global]

Mistral, Oct 20: https://t.co/o1wdWHQ2Aq The Decoder, Oct 21: https://t.co/q50yryPNw2

[Qualiteg Inc. @Qualiteg_global]

Mistral launches Voxtral 3, voice agents with speech-to-text, speech synthesis, and tool calling.
- 50+ languages, incl. low-resource Asian/Indic
- WER 4.0% vs Whisper large-v3 14.9% (FLEURS, Artificial Analysis)
- Mini model, 4B params, $0.0035/min
Link in reply. #AI #Mistral

[Qualiteg Inc. @Qualiteg_global]

DeepSeek, Oct 20: https://t.co/sWsycH0oSK The Decoder, Oct 21: https://t.co/W2TWlMJCMD

[Qualiteg Inc. @Qualiteg_global]

DeepSeek open-sources DeepInfer, the engine behind V4.1 Flash.
- Serves 155 trillion tokens a day on one cluster, DeepSeek says
- Revenue $38 per GPU-hour at valley pricing, DeepSeek says
- Dynamic-length decoding lifts throughput 4-7 percent
Link in reply. #AI #DeepSeek

[Qualiteg Inc. @Qualiteg_global]

arXiv, Oct 15: https://t.co/cuc3S8ltMP The Decoder, Oct 18: https://t.co/Jq62JQAlUq

[Qualiteg Inc. @Qualiteg_global]

Four Gemini 4 Argon agents beat Minecraft, zero human help.
- Beat the Ender Dragon in 27h14m, Google DeepMind says
- Wrote 2,783 JS programs, died 605 times along the way
- Cost $1,040 to win, $1,620 for the full run, Google says
Link in reply. #AI #Minecraft

[Qualiteg Inc. @Qualiteg_global]

Anthropic, Oct 16: https://t.co/W0mzUOnNvc The Decoder, Oct 18: https://t.co/guNeZ6kf1F

[Qualiteg Inc. @Qualiteg_global]

Anthropic's new report: AI agents now outwork junior staff, by hours logged.
- Agent hours beat human hours 6 to 1 by August, Anthropic says
- Junior tech share fell 21% to 14% since early 2025
- Junior tech new hires fell 27% to 9% in the same span
Link in reply.
#AI #Anthropic