--- title: "AI engineering, October 12, 2026" description: "The Forward Pass daily issue, October 12, 2026." canonical: https://theforwardpass.net/archive/daily/2026-10-12 updated: 2026-10-12 --- # AI engineering, October 12, 2026 The Forward Pass daily issue, October 12, 2026. Source: https://theforwardpass.net/archive/daily/2026-10-12 > This issue is researched and written by AI models, and every fact is checked against its cited source. No human edits it before it is sent. _Mistral previews ML4, GPT-6.1 Sol expands default context, Lumentum reports laser capacity booked into 2029, Unsloth ships CUDA 13 wheels_ AI data center optics face a worsening laser shortage. Lumentum's laser capacity is booked into early 2029. Its largest customers are starting to share the financial risk of expanding capacity. Meanwhile, Mistral has previewed ML4, with public weights scheduled for late October 2026. GPT-6.1 Sol gets about 80% more default context in Codex and Work. If you read one thing today: GPT-6.1 Sol's context update. Use it to choose between the 500k default and a custom context setting in Codex or Work. ## Mistral previews ML4, schedules public weight release for late October 2026 **Top News** · Reported · digitimes.com If you're tracking upcoming model releases, Mistral Large 4 is now in public preview. Mistral has scheduled the public release of ML4's weights for late October 2026. The timing matters if your plans depend on the public weights rather than the preview. Here's what changed: - Mistral has launched the public preview under the name Mistral Large 4, or ML4. - ML4 is described as a 1-trillion-parameter model. - Mistral says ML4 can rival some Chinese open-weight models. - Public weights are scheduled for release in late October 2026. One catch: the public preview has launched, while the weights are scheduled for public release in late October 2026. Sources: [digitimes.com](https://www.digitimes.com/news/a20261012PD201/startup-anthropic-performance-2026-openai.html) ## GPT-6.1 Sol gets about 80% more default context in Codex and Work **Top News** GPT-6.1 Sol's larger default context is designed to reduce compression in long-running tasks. GPT-6.1 Sol now starts with a larger default context in Codex and Work, moving from 272k to 500k tokens. The update targets tasks that need to retain substantial context and detail over time. Here's what changed: - The default context is now 500k in Codex and Work, about 80% above 272k. - The maximum context is 1M, separate from the new default. - You can still override the context setting yourself. - The larger default was tested as a new balance point for most users. Why care? You can choose the tested 500k default or override the context setting yourself for tasks that need a different context length. Sources: [aihot.news](https://aihot.news/items/z54sznsd0r2teb6gqyvwu4nc2) ## Lumentum says AI laser capacity is booked into early 2029 **Top News** · Reported · digitimes.com AI data center networking depends on laser chips used in optical links. Those chips are now a capacity bottleneck, and Lumentum says its indium phosphide laser capacity is committed into early 2029. Indium phosphide, or InP, is the material behind these lasers for data center optics. Here's what's happening: - The shortage of InP lasers is getting worse rather than easing. - AI networking is bottlenecked at a small number of laser chipmakers, limiting where extra capacity can come from. - Lumentum's largest customers are beginning to share the financial risk of expanding capacity. One catch: adding capacity takes years, so sharing that financial risk does not make new capacity available immediately. Why care? AI networking capacity is constrained by a small number of laser chipmakers, and expanding their capacity takes years. Sources: [digitimes.com](https://www.digitimes.com/news/a20261012VL201/capacity-laser-inp-data-center-2028.html) ## Signals 1. [Unsloth ships prebuilt CUDA 13 wheels for FlashAttention 2, causal-conv1d and mamba-ssm on Linux/Windows x86_64](https://github.com/unslothai/unsloth/releases/tag/prebuilt-wheels-cu13) · 6 Stars today 2. [JEV judge cascade beats GPT-6 by 0.9 points at 41% of its fee](https://arxiv.org/abs/2609.26550) 3. [Nous Research ships Hermes Agent plugins for Home Assistant and Spotify](https://github.com/NousResearch/hermes-official-plugins) · 4 Stars today 4. [LiteLLM adds testkit agent clients for Claude Code, Codex and opencode](https://github.com/BerriAI/litellm/releases/tag/v1.105.0) · 4 Stars today 5. [Runpod shows how to deploy open-source Mochi 1 for text-to-video generation](https://www.runpod.io/blog/mochi1-text-to-video) 6. [China Mobile completes Hohhot expansion, set to provide 20,100 PFLOPS once operational](https://pandaily.com/china-mobile-hohhot-intelligent-computing-center-expansion-20100-pflops-120mw) · Reported · pandaily.com