AgentdexModel Pulse #2 · Saturday, Oct 24, 2026
Today in AI: open weights close in, and agents leave the editor
The seven AI stories from the last day or so that matter for AI coders and indie game devs. Each one links to its primary source.
Sponsored Sponsor this issue: reach AI coders and solo founders in Model Pulse.
Oct 23, 2026
NEW: Kimi K3.1 Pro lands, with open weights promised by Nov 4
Moonshot AI released Kimi K3.1 Pro on Oct 23. It is a 2.8T-parameter mixture-of-experts model (104B active) with a 1M-token context, live now on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, plus a Highspeed variant Moonshot says runs 2-3x faster. Moonshot reports 65.7% on Terminal-Bench 4.0, just behind Claude Opus 5.5 (66.4%) and ahead of GPT-6 Astra (58.2%). It also pitches full-project game builds in Unity, Unreal and Godot, and scripted 3D modelling in Blender. The trade-off: these are Moonshot's own numbers, and the weights are not out yet. Treat it as a strong API option today and an open model only once the weights land.
Oct 23, 2026
NEW: Cursor agents can now drive your whole computer
Cursor shipped Computer Use on Oct 23. Agents can click and type across any desktop app, not just the browser, so they can open the iOS Simulator and test the feature they just built. Cursor picks the tool on its own when a task needs it, or you can @-mention Computer Use in your prompt. This closes the gap between writing code and checking that it works. It also means the agent can see your screen and press any button. Run it under a test account or in a VM, and keep banking tabs and password managers closed.
Oct 23, 2026
NEW: Claude Code adds MCP Tasks and Elicitations
Anthropic added MCP Tasks and Elicitations to the Claude Code CLI on Oct 23, with more Claude products to follow. Tasks let a plugin take a long job, such as deep research or a full test run, and report back when it is done. Elicitations let a plugin pause and ask you for input mid-task, like confirming a refund. For builders, this means slow tools no longer block your session, and you can put a human check exactly where it matters. The catch is that your MCP server has to support both features first.
Oct 23, 2026
OPEN WEIGHTS: MiniMax M3.1 Flash is now downloadable
MiniMax published the weights for M3.1 Flash on Hugging Face on Oct 23. It has about 428B total parameters (23B active), a 1M-token context, and takes text, image and video input. MiniMax reports 73.7% on DeepSWE v1.1 and 61.7 on OSWorld V2, and its API is listed at $0.15 / $0.60 per 1M input/output tokens. To be fair, its own model card shows it well behind Opus 5 on Terminal Bench 4.0 (38.1 vs 51.8). It is a cheap workhorse for agent loops, not a frontier replacement. Read the license before you ship it commercially.
Oct 23, 2026
NEW: NVIDIA Nemotron 3.5 Bolt puts a talking, tool-using NPC brain in about 4GB
NVIDIA released Nemotron 3.5 Bolt E4B 4GB on Oct 23. It is a small open model built on Gemma 4 E4B for local assistants, robots and in-game characters. It supports multi-turn tool calls, a 64K context and 35 languages, and NVIDIA demoed it as a blacksmith NPC that checks payment, crafts an item and hands it over. Weights, data and recipes are on Hugging Face. For indie game devs, this is a realistic way to run on-device NPCs with no per-call API bill. NVIDIA's own tables show it trailing Qwen 3.5 4B on some tool benchmarks (Tau2-Bench 55.1 vs 75.1), so test it on your own game loop.
Source: https://research.nvidia.com/labs/nemotron/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB/
Oct 22, 2026
GAME DEV: Epic's AI MegaJam pays up to $50K to keep building your agent-made game
Unreal Engine announced the 2026 Epic AI MegaJam, which runs Nov 12-15. Teams of up to five people plus as many AI agents as they like build a playable game in Unreal Engine or UEFN, with agents hooked into the editor through Unreal MCP. There is $65,000 in cash prizes, plus continuation grants of up to $50,000 for standout entries. It is open to anyone 18+ with an Epic account, online or in person in San Francisco. This is the clearest sign yet that a major engine maker wants agent-built games. Using AI is optional, and players are still wary of AI art. Lean on agents for code and systems, and keep a human eye on the look.
Source: https://www.unrealengine.com/news/2026-epic-ai-megajam
Oct 22, 2026
RESEARCH: Claude formally proves 123 open-source libraries crash-free
Anthropic's Frontier Red Team and Theorem used Claude to prove that 239 fuzzing entry points across 123 OSS-Fuzz libraries never crash on any input under 64 MB. That took about 180 million lines of Rocq proofs written over a few weeks, and Claude found and fixed bugs that years of fuzzing had missed. The limits are real: the proofs cover single-threaded static RISC-V binaries, not the x86 or ARM builds most people run, and each proof costs hundreds of millions of tokens. Still, AI-written proofs are starting to look like a practical safety net for AI-written code.
Source: https://www.anthropic.com/research/formal-verification
Ship faster this weekend
Grab the free Agentdex starter kit with prompts and agent configs for Cursor, Claude Code and Codex.
Get the free kitWe store your email with Resend and use it only for Model Pulse. See privacy. Missed yesterday? Read Model Pulse #1.