DeepSeek

Moonshot AI open-sourced its Kimi K3 model this week. The company says it still trails Claude Fable 5 and GPT 5.6 Sol, but independent evals from Arena.ai and Vals AI place it near flagship closed models. The release coincided with Xi Jinping's speech at the World AI Conference in Shanghai; the Nasdaq dropped about 1% on Friday as chip stocks like Nvidia sold off. The discourse echoes the DeepSeek R1 moment from early 2025, now amplified by the Trump administration's tariff war with China, Anthropic's national-security scrutiny, and major AI firms preparing to go public. The post does not disclose K3's parameter count, training cost, or open-source license details.

Mozilla's first State of Open Source AI report shows open-weight models now route the majority of tokens on OpenRouter, with DeepSeek V4 Flash at #1. Inference cost for GPT-4-class models dropped 50× in 36 months to $0.40 per 1M tokens. The capability gap to closed models is 3.3%, concentrated in reasoning and multimodality; coding is at parity. 79% of developers use open models vs. 71% for closed, but only 51% reach production with open (63% for closed). The bottleneck is operational tooling—integration, maintenance, deployment—not model quality. The report highlights real-world cases: a Māori speech model, PwC running a fine-tuned finance model on its own hardware, and a Red Cross medical model headed for clinical trials.

Mira Murati's startup Thinking Machines released its first model. The FT reports the model borrows training techniques from Chinese firms like DeepSeek, achieving near-frontier performance with less compute. The article does not disclose the model name, parameter count, benchmark scores, or whether it is open-weight.

FT reports that DeepSeek, a month after closing its first-ever funding round in June, is already talking to investors about a new round. The post doesn't disclose the first round's valuation, amount, or backers, and the size and purpose of the new round are still undecided. The Chinese AI company had previously been self-funded by founder Liang Wenfeng; back-to-back fundraising signals a rapid shift into expansion mode.

Reasoning models can get stuck in doom loops, repeating useless tokens until the context window fills up. Liquid AI open-sourced Antidoom, which uses Final Token Preference Optimization (FTPO) to fix this. The method trains the model on 1,040 preference pairs to learn when to stop at the end of reasoning. On DeepSeek V4 Pro, the doom-loop rate dropped from 3.2% to 0.3% without hurting math or coding scores. The post doesn't disclose training cost or how well it transfers to non-DeepSeek models.

Anthropic told senators in June that Alibaba used tens of thousands of unauthorized accounts to distill its Claude model at industrial scale. Distillation itself is a decade-old Google invention, and Elon Musk admitted xAI does it too. The post doesn't include Alibaba's response or a clear legal ruling. I'd discount the alarm a bit: export controls and new laws have been slow to materialize, and distillation matters less for the coming wave of AI agents anyway.

This piece reframes AI answer scraping from an SEO tool into a model access market. NetNut's takedown matters not for web scraping but because its scraper catalog openly sold access to ChatGPT, Perplexity, and Google AI Mode. Anthropic's Feb report flagged 24,000 fraudulent accounts making over 16 million Claude interactions from DeepSeek, Moonshot, and MiniMax; Reuters later reported an Alibaba-related allegation of nearly 25,000 accounts and 28.8 million interactions. The core argument: tokens aren't the only cost—stable, programmable access to model product surfaces is the real scarce resource. Brand monitoring is the legitimate buyer; distillation attacks are the risky one. Both rely on the same infrastructure layer.

A 30-month panel study of 26,000 secondary students in central China found that AI use raised homework scores by 18% and cut completion time from 64 to 45 minutes, but closed-book exam scores dropped 20%. The full 18–24% decline on high-stakes entrance exams took about two years to appear. Roughly 81% of long-term users showed an outsourcing pattern—fast homework, high grades, poor exams. Students who spent similar time as non-users saw no exam penalty. Social sciences took the biggest hit at 27%. The post does not name the county or the lead institution.

The author gave 11 LLMs a 1,500-line LangGraph god node to refactor. Fable-5's proposal scored highest in peer review, followed by GPT-5.5 and DeepSeek-4-pro. GPT-5.4 and Opus-4.7 ranked near the bottom. Each model produced full code and architecture docs, then other models cross-evaluated them. Raw data and the ranking matrix are public. Caveat: this is one refactoring task, not a general coding benchmark, but it reveals clear differences in engineering taste across models.

Meituan open-sourced LongCat-2.0, a 1.6-trillion-parameter MoE model with ~48B active parameters per token. It was pretrained on over 35 trillion tokens using 50K+ in-house AI ASICs with no rollbacks or irrecoverable loss spikes, showing frontier-scale training is viable on non-GPU hardware. The model targets long-context and agentic workloads: it introduces LongCat Sparse Attention to speed up 1M-token processing and was trained on hundreds of billions of 1M-context tokens. Official charts place it alongside Gemini 3.1 Pro, GPT-5.5, and Opus 4.8 on Terminal-Bench 2.1, SWE-bench Pro, and other coding/agent benchmarks, though the post does not provide exact numeric comparisons. An N-gram Embedding module with 135B parameters expands the embedding space roughly 100×, which the team claims outperforms scaling standard MoE experts by the same amount. The model is integrated with Claude Code, OpenClaw, and Hermes; code and weights are available on GitHub and HuggingFace.

DeepSeek announced V4 will launch in mid-July with peak/off-peak pricing. Peak hours are 9:00–12:00 and 14:00–18:00 Beijing time, when API prices double. For deepseek-v4-pro, output goes from ¥6 to ¥12 per million tokens; for deepseek-v4-flash, from ¥2 to ¥4. Users get email alerts 24 hours before any price change. The post doesn't cover model capability changes or an exact release date.

CNBC reported on June 26 that Lindy, a ~25-person San Francisco company, switched 100% of its traffic from Anthropic Claude to DeepSeek this month. CEO Flo Crivello said the monthly AI bill had exceeded total employee payroll and the move will save millions. His former employer Uber now caps some AI tools at $1,500/month. Consultant Jeff Henry of Highspring said some clients paused AI spending until ROI is proven. Companies are adopting model routing instead of using the priciest frontier models for every task.

DeepSeek is going on a hiring spree, intensifying China's AI talent war. The FT reports it's poaching from ByteDance, Alibaba and others, with some pay packages reaching 2–3x what candidates currently earn. The post doesn't give a specific headcount target, but says the team will expand rapidly from the current few hundred. I'd discount the hype a bit—whether this pace holds depends on its next funding round and revenue catching up.

A Washington Post test of six major AI chatbots found most lean left on political questions. OpenAI's GPT-5.5 gave exclusively left-leaning answers 80% of the time, DeepSeek V4 Pro 70%. Even Gab's Arya, marketed as conservative, skewed left more often. Google Gemini 3.1 Pro was the outlier, presenting both sides in 93% of cases. xAI's Grok 4.3 took a fully right-leaning stance on trans rights, matching Elon Musk's public position and suggesting deliberate intervention.

The author noticed DeepSeek V4 costs $0.09 vs Anthropic's $5.00—a ~50x gap. He argues Anthropic and OpenAI are trapped by high cost structures and can't compete on price. The post speculates they'll lean on scarcity branding and China-fear lobbying instead of cutting prices. It also points to Allen AI's OLMo as the real open-source path, with training data included. No performance benchmarks are provided; the price delta is the only hard number.

Bloomberg reports Microsoft is testing DeepSeek-R1 and DeepSeek-V4, planning to offer these Chinese models to Western customers. Microsoft also sells ChatGPT to Chinese enterprises, building a two-way AI model trade network across the US and China. The post doesn't disclose pricing, revenue splits, or launch dates.

The author benchmarks GPT-5.5, DeepSeek V4 Pro, and GLM-5.2 on the AA-Omniscience hallucination metric and a Python async coding prompt. GPT-5.5 hits 86% hallucination, DeepSeek V4 Pro 94%, while GLM-5.2 scores 28%. DeepSeek V4 Pro spent nearly 4 minutes and 7.7k reasoning tokens producing a confidently wrong solution; GLM-5.2 needed 12 seconds and ~800 tokens to flag the prompt as technically impossible under single-threaded, no-polling constraints. GLM-5.2 trails GPT-5.5 by only 4 points on the AA Intelligence Index and Claude Fable 5 by 9 points—Fable 5 was restricted by the US government three days post-launch over a single jailbreak. The post argues that scaling parameters and data makes models worse at saying “I don’t know,” and frames an unsolved trilemma: raw capability, hallucination calibration, and compute efficiency. The article does not disclose GLM-5.2’s training data size or exact release date.

DeepSeek researcher Deli Chen open-sourced AutoResearch, a protocol where an AI agent independently ran a full RL research loop on a 285B model—designing experiments, writing code, submitting GPU jobs, debugging, and summarizing results with zero human intervention. The system used GRPO. The post doesn't disclose the specific task, training duration, or success rate, so I'd hold off on getting too excited until there are reproductions.

DeepSeek's image understanding mode went live today on its app and web client, announced by researcher Xiaokang Chen. The mode sits alongside Quick and Expert modes and lets the model interpret visual content beyond simple text extraction. ITHome testing found the app still shows an 'image understanding in beta' notice, while the web version does not. In April, DeepSeek disclosed the underlying multimodal framework 'Thinking with Visual Primitives,' but the post doesn't specify which model powers this mode, whether it's free, or any resolution/file-size limits.

Reasoning models didn't appear overnight in 2024. Chain-of-thought prompting, STaR self-training, process reward models, and test-time compute scaling laws all predate o1. What o1 actually changed was productization: turning reasoning into a billable, schedulable resource and opening a second axis for scaling. DeepSeek R1 made the know-how public, triggering industry-wide convergence within five months. But the most hyped part—pure RL spontaneously creating reasoning—is the weakest claim. Independent studies show base models already contain reasoning fragments; RL merely amplifies their frequency. The real lesson: distinguish the birth of a capability from its packaging.