Running DeepSeek-V4 Locally on Four $2,500 Consumer GPUs — Hitting 255 tok/s
The story of a Reddit user running DeepSeek-V4 — the MoE frontier model that has been making the rounds on Hugging Face's Daily Papers and various media outlets — locally on four previous-generation RTX 2080 Ti cards (roughly $2,500 total) has drawn huge attention in the LocalLLaMA community.
The key tricks were threefold. First, custom kernels were rewritten for the Turing architecture. Second, the setup uses W8A8 quantization (8-bit weights, 8-bit activations). Third, it takes advantage of MoE's expert-distribution mechanism to route efficiently across the four GPUs.
The result: a prefill speed of 255 tok/s. It's proof that a frontier model widely assumed to require an H100 cluster can run at a practical level on legacy GPUs sitting in your drawer. As a result that once again demonstrates the potential of local LLMs, discussion in the community is heating up.
Elon Musk Loses OpenAI Lawsuit — Jury Finds No Contract Breach in the Shift to For-Profit
Tesla CEO Elon Musk has officially lost, by jury verdict, the lawsuit he brought against OpenAI and its CEO Sam Altman.
At the heart of the case was whether OpenAI betrayed its founding nonprofit mission by transforming into a for-profit company. As one of the founding members, Musk claimed that OpenAI had broken its promise to prioritize the public good.
The jury, however, found that OpenAI's conversion to a for-profit entity did not violate any contractual obligations. The ruling could influence industry-wide debates over AI company governance and nonprofit-to-for-profit transitions. In effect, the legal legitimacy of OpenAI's current corporate structure has now been affirmed.
Simultaneous Headwinds for the AI Industry — Linus Torvalds, a Former Microsoft VP, and Schmidt Voice Criticism in the Same Week
In a span of just three days, May 15–17, 2026, three key figures in the AI industry one after another took skeptical stances on the AI push, racking up more than 30,000 combined Reddit upvotes and becoming a major topic of conversation.
Drawing the most attention was Linux creator Linus Torvalds, who criticized AI-powered automated bug hunters for flooding the Linux kernel security mailing list with a torrent of reports, leaving it "nearly impossible" to manage. He pointed to the reality that the poor quality of AI-generated security reports is squeezing out human review.
Next, former Microsoft VP Mat Velloso said that "Microsoft missed the AI wave just like it missed the internet and mobile," highlighting the structural dilemma that giant corporations face at technological inflection points.
The third was former Google CEO Eric Schmidt. Details are limited in the data gathered for this roundup, but all three drew attention as voices applying the brakes to the AI push.
What's telling is that the three didn't merely happen to speak out in the same week — each also struck a powerful chord with the community. Concerns over the speed and quality of AI development, and over the strategic judgment of giant corporations, are surfacing all at once across the industry.
The Tri-Polar AI Era — Shifting Shares of ChatGPT, Claude, and Gemini, and How to Choose Between Them
In May 2026, a major tectonic shift is underway in the AI assistant market. ChatGPT's share has plunged from 45.3% to 38.1%, while Claude has surged from 1.5% to 13.1%. The market is shaping into what could be called a "tri-polar" landscape.
Behind the shift are changes in each company's pricing and performance strategy. At Google I/O 2026, Gemini 3.5 Flash was billed as the cheapest option on the market and the fastest via API; Anthropic cut the price of Claude Opus 4.5 to $5/$25; and OpenAI made GPT-5.5 Instant the default model in ChatGPT.
A Zenn article introduces practical implementation examples for switching among AI APIs in Python depending on the task. As five principles for splitting up usage, it summarizes an approach for selecting the optimal model along the axes of cost, speed, accuracy, context length, and strengths.
Putting Local Coding Agents to the Test — Just How Usable Is Qwen3.6-27B?
An article evaluating how capable a coding agent powered by a local LLM — with no cloud dependency — really is has been published on Zenn.
The backdrop is the fact that frontier models' coding ability has been improving rapidly: Claude Opus 4.7, for one, has reached 87.6% on SWE-bench Verified. All of these, however, run through the cloud, and their sizes make them unrealistic to run on your own hardware.
That's where Qwen3.6-27B, released in April 2026, came into the spotlight. Using this Open Weight model, the author tested how practical it is on real development tasks. The verdict: it reaches a practical level for simple refactoring and test generation, but for complex architectural changes and work spanning multiple files, cloud models still have a clear edge.
Choosing between local and cloud is becoming a new decision axis in day-to-day development.
---
Running DeepSeek-V4 locally with 4x legacy RTX 2080 Ti / Reddit LocalLLaMA
イーロン・マスク、サム・アルトマンとOpenAIに対する訴訟で敗訴 / Zenn AI Quotidia
AI業界3名の批判発言が同じ週にRedditで同時バズ / Zenn amu_lab
AI三極化時代の使い分け実践ガイド / Zenn kairosai
Local Coding Agentが身近なタスクをどれくらいこなせるのかを検証 / Zenn aishift