GoKawiil
Tech News
Cerebras published a technical explanation of how it deploys very large AI models, such as the open-source ~3 trillion-parameter Kimi K3, across multiple WSE-3T wafer-scale chips. The company says it splits models to fit entirely in on-chip memory, since a single WSE-3T wafer holds only 44 GB while frontier models need terabytes for weights alone, enabling inference speeds of 1000-2000+ tokens per second.
Kimi K3.1 Pro has been benchmarked at 55 on the Artificial Analysis Intelligence Index, well above the median score of 27 for comparable models. The model handles text and image input with a 1M token context window, priced at $2.00 per 1M input tokens and $10.00 per 1M output tokens, matching median market pricing.
Moonshot AI released Kimi K3.1 Pro, an open-weight 3T-class model built on Kimi K3 with expanded post-training compute, offering a 1-million-token context window and gains in coding, 3D creation, financial research and agentic tasks. A faster variant, K3.1 Pro Highspeed, runs 2-3 times quicker and both are live now on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, with full model weights due by November 4.
A test of 12 AI models from ten different labs asked each for its most controversial opinion of humans, three times per model, producing 35 answers. Researchers found 29 of those answers independently converged on a single idea—that people prioritize social belonging and group loyalty over objective truth—with phrasing varying but the core claim repeating across models like GPT-6.1 Sol, Grok 4.7, DeepSeek V4 Pro, Kimi K3 and Gemini 4 Argon.
Interpretability startup Goodfire has released AI safety monitors that inspect a model's internal signals in real time, rather than reviewing its text output after the fact. The tool is now available to customers of Baseten, which hosts AI models, following a safety partnership announced last month with Baseten's Base Labs and Hugging Face. Users can choose which risks to flag, such as hacking attempts or weapons-related misuse, and select automated responses ranging from logging to blocking requests.
Fireworks Research released Ember-1, a specialized model built on Kimi K3 that matches its quality while using 40% fewer tokens. The company says it trained the model to reason more efficiently rather than simply reducing reasoning effort, using more than 50 training experiments and 200 evaluations run on its Serverless Training platform. Ember-1 is available now and is described as the first in an ongoing series of specialized models from Fireworks.
American AI developers have alerted US authorities that foreign actors, likely from China and Russia, are using distillation techniques to replicate the capabilities of Western frontier models at much lower cost. These attacks reportedly involve buying logs of conversations from legitimate accounts to extract training data, making the practice difficult to fully stop despite lab collaboration efforts started earlier in 2026. China has denied the accusations and warned it will impose 'countermeasures' if the US uses this issue to justify restricting Chinese AI development.
Chinese AI startup Moonshot announced Thursday that its Kimi models now integrate directly with financial data providers including S&P Global Market Intelligence, Crunchbase, Wind, Tianyancha, SEC EDGAR filings, the IMF, World Bank and FRED. The company said clients such as CICC and venture firm Hong Shan (formerly Sequoia China) are already using Kimi's new financial-services tools, which are offered through tiered subscriptions ranging from 49 to 699 yuan per month.
Cognition released SWE-2, a coding model built by post-training on a 2.8-trillion-parameter Kimi K3 base with 104B active parameters, marking its first RL push into the multi-trillion-parameter range. The model posts a class-leading 92.8 on Terminal-Bench 2.1 and competitive FrontierCode scores at lower cost than rivals like Claude Fable 5.1 and GPT-6 Astra, but falls well behind both on the harder Terminal-Bench 4.0 benchmark (27.3 vs 55.8 and 57.9). SWE-2 is closed-weight and available now through Devin Desktop and CLI, with no local deployment option.
Cognition unveiled SWE-2, a coding-focused AI model post-trained from the 2.8-trillion-parameter Kimi K3 base. The company says it scores 50.0% on FrontierCode 1.1 Main, nearly matching Fable 5.1 while costing 64% less, and comes close to GPT-6 Astra at roughly a quarter of that model's price. SWE-2 also outperforms Cognition's earlier SWE-1.7 model and Grok 4.6 across several coding benchmarks.
Legal AI startup Harvey has closed a $550 million funding round valuing it at $15.5 billion, co-led by Diffusion and Lightspeed Venture Partners. The raise comes just months after an $11 billion valuation in March and an $8 billion mark last December, bringing Harvey's total funding past $1.55 billion. The deal follows the company's launch of its own in-house model, Harvey Tenet, built on an open-weight foundation with help from Fireworks.
A follow-up experiment tested four open models—DeepSeek V4 Flash, Inkling, Kimi K3, and Qwen3.8 A95B—by inserting the first 1% of GPT-5.5 Pro's reasoning trace into each model's own reasoning channel before letting it generate answers freely. Researchers then measured how much of GPT-5.5 Pro's visible answer text overlapped with each model's output. Qwen3.8 showed the largest jump, with overlap rising from 33.92% unprefilled to 54.50% with the GPT-5.5 Pro prefill, a 20.58 percentage-point increase, while other models showed much smaller shifts.