Kimi K3: The World's Largest Open-Weight Model Drops Tomorrow
Moonshot AI, the Beijing-based startup backed by Alibaba, is releasing the open weights of Kimi K3 — a 2.8-trillion-parameter model — on July 27 at 00:00 UTC. It's the largest open-weight model ever released, and it's already been making waves since its API debut at the World AI Conference in Shanghai on July 16.
K3 uses a sparse mixture-of-experts architecture: only 16 of its 896 experts fire per token, meaning roughly 50 billion parameters are active at any given time. Despite that efficient design, the model topped the Frontend Code Arena benchmark with a score of 1,679, edging out even Claude Fable 5 in blind developer testing.
There's a catch for anyone hoping to run it locally: even in MXFP4 four-bit precision, the weights clock in at roughly 1.4 terabytes. You'll need serious hardware — or a cloud cluster — to self-host. The weights will be available on Hugging Face under a modified MIT license. API pricing sits at $3 per million input tokens and $15 per million output tokens.
The release underscores China's growing competitiveness in frontier AI, particularly in open-weight models where Chinese labs have been aggressively pushing the envelope. (Sources: VentureBeat, Tom's Hardware, TechTimes)
Google Delays Gemini 3.5 Pro as Coding Performance Falls Short
Google has delayed the broader release of Gemini 3.5 Pro, its flagship frontier model, after internal testing revealed the system's coding abilities fell short of expectations. The model was originally targeted for a June launch, then pushed to July — and there's still no confirmed general availability date.
According to Bloomberg, Google updated the model's training data late last month to try to improve coding performance, but the results were disappointing. The delay comes at a particularly awkward time, with OpenAI's GPT-5.6 and Anthropic's Claude models continuing to dominate coding benchmarks.
The situation has reportedly caused friction within Google's AI teams, with internal frustration over the risk of losing ground to competitors. Alphabet shares fell nearly 3% following initial reports of the delay. Google's Gemini 3.5 Pro remains available only in limited Vertex AI enterprise preview.
(Sources: Search Engine Journal, 9to5Google, WinBuzzer)
Apple Sues OpenAI Over Alleged Hardware Trade Secret Theft
In one of the most dramatic legal moves in tech this year, Apple has filed a lawsuit against OpenAI in the US District Court for Northern California, accusing the AI lab of stealing hardware trade secrets through former Apple employees.
The suit centers on two individuals: Tang Tan, OpenAI's hardware chief and former Apple VP of product design, and Chang Liu, a former Apple senior systems electrical engineer who joined OpenAI in 2026. Apple alleges that Tan emailed himself confidential information about Apple's suppliers before leaving, and used internal Apple codenames when interviewing Apple employees for OpenAI roles to extract further information.
Perhaps most damaging: Apple claims Liu failed to return a company laptop after departing, used it to download confidential documents, and discovered a bug that let him access Apple's cloud file storage while already employed at OpenAI. The stolen materials allegedly include specs for unannounced products, engineering presentations, and proprietary project data.
OpenAI has not yet publicly commented on the specifics of the lawsuit. The case could have significant implications for talent mobility between tech companies and the handling of proprietary information in the AI hardware race.
(Sources: TechCrunch, CNBC, Fortune)
Meta Breaks from Open-Only Strategy with Muse Spark 1.1 Paid API
Meta has officially entered the paid AI API business. The company launched Muse Spark 1.1 on July 9 alongside the Meta Model API in public preview — marking a significant strategic shift from its historically open-only approach.
Muse Spark 1.1 is Meta's second model from Superintelligence Labs, built for agentic tasks with a 1-million-token context window. Unlike Meta's open Llama family, Muse Spark 1.1 is proprietary and closed-weight. CEO Mark Zuckerberg positioned the pricing — $1.25 per million input tokens and $4.25 per million output tokens — at roughly 25% of what Anthropic and OpenAI charge for comparable models.
New developers get $20 in free credits. The API is currently available in public preview for US developers, with international rollout expected later this year. The move puts Meta in direct competition with OpenAI, Anthropic, and Google for developer API revenue — a market that's become central to the AI business model.
(Sources: Quartz, TechTimes, DataCamp)
Anthropic Signs $19 Billion, 20-Year Data Center Deal with TeraWulf
Anthropic has locked in one of the largest AI infrastructure deals to date: a $19 billion, 20-year lease with TeraWulf for a purpose-built AI campus in Hawesville, Kentucky. The deal, announced July 6, will support approximately 401 megawatts of critical IT load.
Initial capacity is expected online in the second half of 2027, with full buildout by early 2028. TeraWulf plans to invest between $3 billion and $4 billion in the facility — less than one-fifth the lease's total value. The deal sent TeraWulf shares up more than 10% on the announcement.
The deal comes as Anthropic's annualized revenue crossed $47 billion in May 2026, and the company has filed a confidential S-1 with the SEC signaling an upcoming IPO. The scale of this infrastructure commitment signals Anthropic's confidence in sustained, long-term demand for its Claude models.
(Sources: SiliconANGLE, Data Center Dynamics, Yahoo Finance)