The Day in AI

• Synchopate says llama-modchip v2.4.0 runs the 284B-parameter DeepSeek-V4-Flash on a 64GB machine with no GPU, at 256K context.
• Dave Wehner, Meta's chief strategy officer and its CFO from 2014 to 2022, is leaving Meta after 14 years to become CFO of Fireworks AI.
• Thinking Machines released Inkling 2 Small on 5 November, a model with 594B total and 17B active parameters, pretrained in 68 hours.
• Nous Research released Arete-27B, a Qwen3.8-27B fine-tune that it says repeats China's party line in 1% of answers, down from 22%.
• Anthropic bought commitments for 1.7GW of gas-fired generators from OnePWR for $3.4 billion, according to the Financial Times.

DeepSeek-V4-Flash is a mixture-of-experts model, so each token passes through about 13.17B of its 284.33B parameters. Synchopate compresses the weights to 1.71 bits each with its own format, called Fractal-64, adds a compression scheme called SwiGLU-H2, and runs on AVX-512 kernels for the CPU. The file comes to 56.56 GiB. A standard UD-IQ1_S build is 76.87 GiB in the project's own comparison.

All the speeds are Synchopate's, measured on a 16-core Zen 5 processor. Prefill, the step that reads the prompt, runs at 44.36 tokens per second. Decode runs at 9.1 to 14.7 tokens per second, and the range covers runs with and without Stockfish-PVS speculative decoding. In speculative decoding a cheap draft proposes several tokens and the big model checks them in one pass. The project reports peak memory of 57.61 GiB at 256K context, and news coverage quotes up to 58.9 GiB.

The files use a custom tensor type, number 40. Stock llama.cpp, Ollama and LM Studio reject them, so they run only on the project's patched runtime.

Dave Wehner, who joined Facebook in 2012 and was Meta's CFO from 2014 to 2022, announced on X on 7 November that he will become CFO of Fireworks AI. Fireworks runs inference and customization for open models. Its CEO, Lin Qiao, said Wehner will scale the company's finances and operations. Business Insider reported that Fireworks is in early talks to raise at a valuation of $30 billion to $40 billion, after a $17.5 billion Series D in July.

Thinking Machines released Inkling 2 Small on 5 November through its API. It is a multimodal mixture-of-experts model with 594B total and 17B active parameters, pretrained on under 12T tokens. The main pretraining phase used 8.4T tokens in NVFP4, a 4-bit floating point format, and took 68 hours on GB300s, according to Soumith Chintala. Artificial Analysis gave it 31 on its Intelligence Index at launch. The Latent Space newsletter reports that it beats Muse Spark.

• Nous Research released Arete-27B, which it says cuts party-line answers from 22% to 1% on its own 803-prompt eval with no significant loss on Hermes Bench. The earliest coverage is dated 8 November.
• Anthropic bought 1.7GW of gas-fired generator commitments from OnePWR for $3.4 billion, the largest known such purchase by an AI lab, according to the Financial Times.
• Empero, a German lab, says its invite-only Manta model reads 100 million tokens in one prompt and found 18 of 20 hidden codes in its own test, with no independent check.
• Tencent added three video models to its WorkRally platform, and says its UpScale-rally upscaler cuts processing time by 80%.
• AI companies sold or priced more than $300 billion of stock in 2026, excluding Anthropic's IPO, according to The Information.
• Goldman Sachs is reportedly pulling Claude Code from its bankers and traders.
• Apple John Ternus is pushing Craig Federighi to move Siri toward agentic features, Bloomberg's Mark Gurman reports.
• DeepSeek has a model that reportedly completed 21 of 31 tasks in a Vals AI hacking test.
• Google DeepMind reportedly found two flaws in 1024-bit RSA, and 2048-bit keys are reported to stay safe.
• Apiiro reports that Claude Code's "Don't Ask Again" setting exposed passwords in 1,122 public GitHub files.