Qualiteg Inc.@Qualiteg_global

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.

 Tokyo, Japan
qualiteg.com
Joined August 2026
19 Photos and videos
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.
1
1
302
vLLM refused the model: minimum capability 89, current 86. - The card was an NVIDIA RTX A6000, capability 86 - FP8 arithmetic arrives with Hopper, so 89 is the floor - Blackwell is capability 100 with native FP4 Quantization is a hardware question. Link in reply. #vLLM #AI
1
10
Why vLLM 0.5.1 rejected fp8 quantization on an NVIDIA RTX A6000, what GPU capability levels mean, which generations support FP8 natively, and what native FP4 on Blackwell will change: journal.qualiteg.com/gpu_cap…

Native FP8 and FP4 Support, and "fp8" Quantization with vLLM

Hello, this is the Product Development Department at Qualiteg. When a new model is released, we often go hunting for the right recipe to speed up inference — quantizing with various techniques and...

journal.qualiteg.com
5
Gartner named the AI that watches AI: guardian agents. - Three types, reviewer, monitor and protector - Gartner sees 10 to 15% of the agentic AI market by 2030 - It also sees 70% of AI apps multi-agent by 2028 No product yet, risk now. Link in reply. #AISecurity #AI
2
22
What Gartner means by guardian agents, the reviewer, monitor and protector split, how they automate AI TRiSM that people do by hand today, and what you can put in place while the products catch up: journal.qualiteg.com/llm-sec…

What Are "Guardian Agents," the New Sentinels of the AI Agent Era?

Hello! Today we take a closer look at Guardian Agents, a concept Gartner recently announced. Gartner's Official Definition In June 2025, Gartner—well known for its Hype Cycle—published its perspect...

journal.qualiteg.com
6
Cross-entropy, part 5: one sample becomes a batch. - Score all four samples, then add the errors up - t and y take a second subscript i for the sample - Divide the sum by N so batch sizes stay comparable That is categorical cross-entropy. Link in reply. #DeepLearning #AI
1
15
Part 5 of our cross-entropy series: summing the error over four training samples, the double subscript notation, and the mean over N that gives you categorical cross-entropy: journal.qualiteg.com/cross-e…

Cross-Entropy Explained, Part 5: Cross-Entropy for Multi-Class Classification (N Samples)

Good morning from the Qualiteg Research Team. In this installment, we extend the previous discussion and derive the cross-entropy for multi-class classification with N samples, working through actual...

journal.qualiteg.com
4
Ollama 0.40.0-rc0 makes MLX the default on Apple Silicon. - Supported architectures now pick MLX on their own - MLX first landed in 0.19, back in March - No speedups and no documented off switch in the notes - The supported list still grows during the RC Link in reply. #LocalLLM
2
55
Release notes: github.com/ollama/ollama/rel…

Release v0.40.0 · ollama/ollama

What's Changed Models run on MLX on Apple Silicon by default In this release, model architectures supported by the MLX runner will run by default on Apple Silicon devices. ollama pull qwen3.8 o...

github.com
9
GPU PC build, part 13: the wiring, and the machine is done. - The 24-pin ATX bundle goes in until the latch clicks - CPU power is a 6-pin plus 2-pin, one end at each side - This card wants three PCIe feeds, so a 3-in-1 splitter Panels back on. Link in reply. #GPU #AI
1
39
Part 13 of the GPU PC build: the 24-pin ATX and CPU power cables, a three way splitter for the graphics card, SATA power for the removable drive bay, and the end of the assembly series: journal.qualiteg.com/jisaku_…

Building a GPU PC, Part 13: Connecting the Power Cables

We are now in the final stretch of the build. Let's keep going! 1. Connect the motherboard power cable The 24-pin ATX cable is the one with the largest bundle of wires. First, plug it into the...

journal.qualiteg.com
28
LLM inference, part 5: how many users fit on one GPU. - Llama 8B at 16-bit takes 16 GB before any user - 2,000 tokens of context costs about 1 GB per request - So 24 GB serves 6, 48 GB serves 28, 80 GB serves 56 Those are ceilings, so load test. Link in reply. #LLM #AI
1
12
Part 5 of our LLM inference infrastructure course: sizing GPU nodes from KV cache math, quantization and parallelization strategies, staged load testing, and the final cost versus performance trade-off: journal.qualiteg.com/how_to_…

LLM Inference Infrastructure Provisioning Course, Part 5: A Practical Process from GPU Node...

Hello! In the previous installments of our LLM Inference Infrastructure Provisioning Course, we covered defining inference speed, estimating request volume, calculating memory consumption, and...

journal.qualiteg.com
5
The newest ESP32 runs real Linux. Note it is the S31, not the P4. - Dual-core RISC-V with a real MMU, so virtual memory works - Espressif ships a Linux BSP preview on a 6.18 kernel fork - Buildroot 2025.02, 16MB PSRAM, 1G Ethernet Link in reply. #ESP32 #Linux
1
1
38
Espressif's BSP repo. ESP32-S31 only, and the branch is marked a developer preview that is not recommended for production yet: github.com/espressif/esp-lin…

GitHub - espressif/esp-linux-bsp

Contribute to espressif/esp-linux-bsp development by creating an account on GitHub.

github.com
1
21
Is a bigger tokenizer vocabulary always better? No. - Mistral's Tekken has about 130,000 tokens on tiktoken BPE - Reported 1.5 to 2 times more efficient on Japanese - Yet on one text Tekken used 74 tokens, a rinna model 42 Fit it to your language. Link in reply. #NLP #AI
1
32
What the Tekken tokenizer is, why Mistral adopted it for 128k contexts, and our own measurement where a Japanese specialized SentencePiece model still used fewer tokens on the same Japanese text: journal.qualiteg.com/about-t…

[Explainer] What Is the Tekken Tokenizer? Inside the New-Generation Tokenizer Adopted by Mistral

Hello! Today we are going to talk about Tekken! When you hear "Tekken," what comes to mind? The fighting game Tekken (鉄拳, "iron fist"), perhaps? Personally, I thought of the Long Swordsman (鉄剣戦士,...

journal.qualiteg.com
1
1
45
Why did enterprise security get so complex? - One proxy once held the only way out, and it knew who - Cloud and mobile broke that path in the mid 2010s - SWG, CASB, ZTNA and SASE all try to win the view back Know who sent what to the AI. Link in reply. #CyberSecurity #AI
1
18
How enterprise security went from one proxy that saw everything to SWG, CASB, ZTNA and SASE, and why consolidated observation points still matter now that staff send company data to AI: journal.qualiteg.com/ai-era-…

Why Did Enterprise Security Become So Complex? From the AD + Proxy Era to the Modern Cloud

Hello! Now that generative AI services like ChatGPT and Claude are making their way into everyday work, "the risk of sending confidential information to AI" has emerged as a new security challenge....

journal.qualiteg.com
7
"Cheaper and smarter" is the pitch. What the docs and Artificial Analysis say: Claude Opus 5.5 is 20% below Opus 5, cuts cache reads by 60%, and is #1 on the Artificial Analysis Intelligence Index. The catch is that turning thinking off now returns 400. #Claude #Anthropic
1
22
The full 18-chapter guide, from pricing to Claude Code settings: journal.qualiteg.com/claude-…

Claude Opus 5.5 Complete Guide: Model Specifications, API Notes, and Claude Code Operations

A complete guide to Claude Opus 5.5: prices 20% below Opus 5 with 60% cheaper cache reads, API migration notes such as the medium default effort and thinking that can no longer be disabled, and the...

journal.qualiteg.com
8
Load more