Alibaba open-sourced its internal code reviewer. Same Claude Opus model, 200 PRs: - Open Code Review: 37.8% precision, 352K tokens - Claude Code: 15.9% precision, 2,062K tokens Code picks the files. The model only judges. Would you trade some recall for fewer false alarms?

A 428B open model now runs on 150GB of memory. Unsloth's MiniMax M3.1 Flash GGUFs: - 2-bit: 143.9GB, 91.5% top-1 agreement with the full model - 4-bit: 256GB, 98.0% - 23B active, 1M context My take: serious local coding is now a memory budget, not a GPU budget.

Baidu's new OCR model parses a whole book in one request. Unlimited OCR 2: - 3.3B params, MIT license - Memory and latency stay nearly flat as pages grow, per Baidu - Hundred-page benchmark: 77.1 vs 33.7 for the next best in its chart Open weights. What doc pile goes in first?

1 researcher. 3.5 months. Under $10M of compute. That's how Mirendil says it pretrained its base model from scratch. Agents wrote over 99% of the code. Its first model, Fern 0.1, takes up to 10M tokens of context. My take: the pretraining team just became optional. Buy it?

Claude Enterprise users without Claude Code spent under $75 a month on average last quarter, per Anthropic's S-1. The top 5% using it spent about 20x that. $100K+ customers: ~1,500 at end of 2025, 11,900+ by Sept 30. Chat is the sample. Coding agents are the business.

Anthropic's S-1 is public. Ticker: ANT. Revenue per dollar of compute: Q1 2025: 0.40x Q2 2026: 2.08x Q2 revenue: $11.6B on $5.6B of compute. Adjusted operating margin: -221% to +19%. "AI loses money on every token" just got harder to argue.

Nvidia quietly cut Vera Rubin's specs. - HBM4 pin speed: 11.7 to 9.6 Gbps - NVL72 memory bandwidth: 1,580 to 1,400 TB/s - NVLink per GPU: 3.6 to 3 TB/s TechPowerUp says AMD's MI455X now leads on memory: 432 GB at 23.3 TB/s per GPU vs Rubin's 288 GB at 19.2 TB/s.

Vercel's AI gateway, April to September: Price per token: down to 43% of April Spend: 3.4x Usage: 7.9x Tokens got way cheaper. Total bills went up anyway.

Cerebras showed how it serves a 3-trillion-parameter model. Kimi K3 at 4-bit, spread across 59 wafers: two linear-attention layers per wafer, one dense layer per wafer. Weights never move. Only activations hop, about 2 microseconds each. Claimed speed: 1,000 to 2,000+ tokens/s.

A 33B open model just beat Qwen3.8-27B on most agentic coding benchmarks. Nanbeige4.3-33B, Apache 2.0: - Terminal-Bench 2.1: 80.9 vs 73.0 - DeepSWE: 65.5 vs 42.5 - SWE-Bench Pro: 60.3 vs 61.7, a loss The lab's own numbers. A 3B distill is next. Anyone running it locally?

"No human can write better code than LLMs." Xiangpeng Hao on bbtree, a Rust B-tree where LLMs wrote all the code. His benchmarks: 5x faster than std's BTreeMap on one thread, 12x faster than crossbeam-skiplist at 96 threads. Speed is easy. Proving it correct is the hard part.