Inkling's 975B Open Weights, xAI Open-Sources Grok Build, and a DeepMind Departure

Good morning. It's Wednesday, July 16, and we're covering a 975 billion parameter open-weights model from Thinking Machines, xAI finally opening the Grok Build source code, a Google DeepMind researcher explaining exactly why he left, and a proposal that SQLite should borrow Rust's edition system.

The Inkling announcement hit 992 points on HN by morning - the largest open-weights model release in 2026 so far.

In today's brief:

• Thinking Machines ships Inkling, a 975B MoE with 41B active parameters and open weights
• xAI publishes Grok Build source after the data-upload controversy dies down
• Alex Turner's "Why I Left Google DeepMind" details the Pentagon AI deal
• SQLite editions proposal gains traction as a backwards-compatibility escape hatch

THE BIG ONE

Inkling: A 975B Open-Weights Model With Controllable Thinking

Thinking Machines released Inkling, a 975 billion parameter Mixture-of-Experts model with 41B active parameters and fully open weights. The model supports a 1 million token context window and was pretrained on 45 trillion tokens spanning text, images, audio, and video.

What makes Inkling notable beyond scale is the efficiency claim. Thinking Machines says Inkling "reaches a given score at fewer tokens" - matching Nemotron 3 Ultra on Terminal Bench 2.1 at roughly a third of the token cost. The model features controllable thinking effort, letting developers dial between reasoning depth and token efficiency depending on the task.

Benchmark numbers show competitive performance across the board:

Benchmark	Score
AIME 2026	97.1%
SWEBench Verified	77.6%
MMMU Pro (vision)	73.5%
Terminal Bench 2.1	63.8%

The post-training involved large-scale reinforcement learning with over 30 million rollouts. Thinking Machines claims the improvements in reasoning and chain-of-thought efficiency "emerged organically during training" rather than being explicitly engineered.

A smaller variant, Inkling-Small (276B total, 12B active), is available in preview. Both models ship with Apache 2.0 weights on Hugging Face. Inference is available through partnerships with Together AI, Fireworks, Modal, and Databricks.

The HN thread at 992 points focused on what this means for the open-weights landscape. At 41B active parameters, Inkling is smaller to run than it sounds - competitive with models that require far more hardware. The SWEBench Verified score of 77.6% puts it in the same tier as Fable 5 on coding tasks.

Why it matters: The largest open-weights release of 2026 arrives with efficiency claims that could reshape the hosted-vs-local cost calculation for teams deciding where to run inference.

Our coverage: Inkling: Thinking Machines' 975B Open-Weights Model

PLATFORMS

xAI Open-Sources Grok Build

xAI published the Grok Build source code, following through on promises made after the data-upload controversy in June. The repository contains 844,530 lines of Rust.

The release came after weeks of scrutiny over Grok Build's aggressive telemetry. The original implementation was uploading repository contents and shell history to xAI servers with insufficient consent flows - behavior that Simon Willison documented extensively at the time.

The HN discussion hit 458 points and 503 comments, with much of the conversation focused on what the code reveals about modern agent architecture. Willison noted that the codebase includes tools adapted from competing systems like Codex and Claude Code, showing "how modern agents adapt across different environments."

One practical outcome: Willison built grok-mermaid, a browser-based tool that compiles the Grok Build Rust code to WebAssembly to convert Mermaid diagrams into Unicode box art. The kind of thing you can do once the code is actually available.

Why it matters: Open-sourcing after a privacy incident is damage control, but the code itself is now a resource for understanding how a well-funded agentic coding tool is built.

Our coverage: Grok Build Goes Open Source: What the 844K Lines Reveal

RESEARCH

Why I Left Google DeepMind

Alex Turner, a former Research Scientist at Google DeepMind, published "Why I Left Google DeepMind", a detailed account of his departure centered on Google's classified Pentagon AI contract.

The post claims Google signed a military AI deal in April 2026 granting the Pentagon "all lawful use" of Gemini without binding restrictions against lethal autonomous weapons or mass surveillance. Turner argues this directly contradicts Google's 2018 AI principles and internal promises made during DeepMind's 2014 acquisition.

Turner describes extensive but unsuccessful internal advocacy: a 250-person petition, a 25-page oversight framework proposal, and direct approaches to senior leadership. His concern extends beyond ethics to AI safety: he argues that military deployments lack adequate mechanisms to detect AI deception through chain-of-thought monitoring, creating risk if advanced systems become misaligned.

The post names specific executives - Jeff Dean, Demis Hassabis, Stuart Russell - who Turner says pledged against autonomous weapons but "failed to take visible action despite having significant leverage."

The HN thread reached 332 points, with debate splitting between those who see this as whistleblowing and those who view military AI applications as legitimate.

Why it matters: A documented first-hand account of how AI safety concerns interact with commercial and government contracts inside a major lab - rare transparency about decisions usually made behind closed doors.

DEVELOPER TOOLS

SQLite Should Have Rust-Style Editions

A post arguing that SQLite should adopt Rust-style editions hit 278 points on HN. The proposal addresses SQLite's collection of problematic defaults that cause real bugs in production.

The author experienced actual crashes from SQLite's default behaviors: foreign key constraints disabled by default (allowing dangling references), loose typing that accepts wrong data types, concurrent write locks that fail immediately instead of waiting, and Write-Ahead Log mode disabled despite better performance.

The proposed solution is a "super pragma" using year-based editions:

PRAGMA edition = 2026;

This single declaration would enable sensible defaults - foreign key enforcement, a 5-second lock timeout, WAL mode, strict typing - without breaking existing databases. Future editions could adopt new improvements as they emerge.

The HN discussion debated whether the complexity is worth it. Some argued that SQLite's simplicity is the feature and adding edition complexity fights against the core design. Others pointed out that the current state forces every SQLite user to rediscover the same set of recommended pragmas.

Why it matters: SQLite ships in roughly everything. A backwards-compatible way to improve defaults at the pragma level would compound across billions of deployments.

TOOLS

Claude Code v2.1.211

Claude Code v2.1.211 shipped yesterday with a set of fixes for parallel session management and plugin stability. Key changes:

• Added --forward-subagent-text flag for stream-json output
• Fixed parallel Claude Code sessions logging out simultaneously after wake-from-sleep
• Fixed plugin MCP servers not reconnecting after idle web session wake
• Fixed nested .claude/rules/*.md files loading incorrectly
• Fixed /loop hiding sessions from /resume

The most-discussed change in recent releases has been the fix for isolation: 'worktree' subagents accessing the main repo instead of isolated worktrees - a bug that affected teams running parallel agents on the same repository.

WHAT ELSE IS HAPPENING

• Running Gemma 4 26B at 5 tokens/sec on a 13-year-old Xeon with no GPU (289 points): CPU-only inference on old server hardware - useful for understanding the floor of what's possible without accelerators.

• Governments should invest in open-source AI (201 points, PDF): The Siegel Foundation argues for public investment in open AI infrastructure as a counterweight to concentrated commercial power.

• G# language announcement (100 points): A .NET language combining Go, Kotlin, and Swift ergonomics - another entry in the "what if we redesigned C#" category.

• Claude web_fetch exfiltration vulnerability patched: Researcher Ayush Paul found that Claude's web_fetch tool could be tricked into following nested links on attacker-controlled sites to exfiltrate context. Anthropic has since restricted link-following in fetched content.

• Anthropic pledges $10M to Canadian AI research: Part of Anthropic's broader push into non-US research partnerships this month.

---

Every link above goes to a primary source or our sourced coverage. Tomorrow's brief lands when the news does - subscribe to get it by email.