π€π» AI Daily Digest β July 16, 2026
Mistral Leanstral 1.5: Apache-2.0 Lean 4 Proof Engineering, 100% miniF2F
Mistral AI released Leanstral 1.5 on July 2, an open-weight model purpose-built for Lean 4 proof engineering that delivers a dramatic cost-performance breakthrough in formal verification. The 119B-parameter MoE (6B active per token) saturates both the miniF2F validation and test sets at 100%, solves 587 of 672 PutnamBench problems, and achieves new highs on FATE-H (87%) and FATE-X (34%) graduate algebra benchmarks β all at an estimated $4 per solved problem, compared to $300+ for competing high-budget provers.
The model's test-time scaling behavior is remarkable: performance on PutnamBench rises monotonically from 44 problems at 50K tokens to 587 at 4M tokens per attempt, demonstrating that Leanstral keeps reasoning rather than plateauing. Beyond pure math, Mistral built a pipeline that translates Rust to Lean via Aeneas, then has Leanstral infer correctness properties and attempt proofs. Across 57 repositories, it flagged 47 violated properties and 11 genuine bugs β 5 previously unreported on GitHub. One caught an integer overflow in a varint decoding library's sign function that crashed in debug mode and silently corrupted data in release. The model is available on HuggingFace under Apache 2.0, with a free API endpoint as leanstral-1-5.
β Mistral AI Β· HuggingFace
π Mistral Leanstral 1.5 Announcement Β· HuggingFace Model
Poolside Laguna XS 2.1 & M.1: Open-Weight Agentic Coding Models
Poolside AI released its first public open-weight coding models on July 2, marking a significant entry into the agentic coding tools market. The lineup includes Laguna XS 2.1, a 33B-parameter MoE (3B active per token) design small enough to run locally on a single desktop GPU, and Laguna M.1, a 225B-parameter flagship (23B active) optimized for long-horizon enterprise software engineering. Both models were trained entirely in-house on 30T tokens using 6,144 interconnected NVIDIA H200 GPUs, with async on-policy reinforcement learning in Poolside's agent harness.
Benchmarks are competitive: XS 2.1 scores 70.9% on SWE-bench Verified and 63.1% on SWE-bench Multilingual, outperforming comparable small MoE models. The lightweight model is distributed under OpenMDW-1.1 (fully permissive), and Poolside also offers DFlash speculator models that double local inference throughput. With $2B raised at a $12B valuation backed by NVIDIA, the company explicitly positions this release as an open-weight counterweight to Chinese AI labs like Alibaba and DeepSeek in the coding assistant sector.
β Poolside Β· NVIDIA Β· OpenMDW
π Poolside Laguna XS 2.1 Blog Β· Poolside Models Page Β· Open Source For You Coverage
NVIDIA & HuggingFace: Open Robot Foundation Models + Data for Agents
NVIDIA and HuggingFace announced a joint initiative to develop open-source foundation models for robotics, integrating NVIDIA's Isaac GR00T 1.7 reasoning vision-language-action model and the Isaac Teleop data-collection framework directly into HuggingFace's LeRobot ecosystem. The collaboration connects NVIDIA's GPU hardware ecosystem and CUDA software stack with HuggingFace's massive model library and developer community, dramatically reducing the barrier to entry for robotics AI training and deployment. Cosmos 3 integration is planned next.
In a parallel move, the partners also announced the Open Data for Agents initiative, publishing over 10 trillion pre-training tokens and millions of post-training samples specifically designed for building AI agents. The release includes region-specific synthetic personas and an interactive Nemotron Post-Training v3 Prompt Atlas, enabling organizations to fine-tune agent models without exposing proprietary data. The combined effect standardizes post-training, evaluation, and deployment for humanoid robotics inside a widely used open-source stack while simultaneously addressing a systemic bottleneck in agent development: access to high-quality, diverse training trajectories.
β NVIDIA Β· HuggingFace Β· Yahoo Finance
π NVIDIA + HuggingFace Robotics Β· HuggingFace Models Blog
Meta Muse Spark 1.1 Enters the Coding Arena
Meta marked a pivotal strategy shift this month with the release of Muse Spark 1.1, its most powerful agent model now specifically targeting agentic coding. The event was notable enough to pull CEO Mark Zuckerberg out of a three-year social media hiatus β his first post on X since July 2023 β where he described Spark as "a very low-cost but powerful agent and coding model" excelling at "agentic performance, tool use, and computer operation." Meta AI head Alexandr Wang claims Spark 1.1 is "currently the most capable model in agentic tasks and coding."
The model is available via public preview on Meta's API portal and delivers strong performance on multi-application computer-use workflows, maintaining context across long sessions and intelligently choosing between scripts, direct UI interaction, and batch operations with minimal human intervention. Wang confirmed that Meta is simultaneously training Watermelon, a larger model that has already matched GPT-5.5 on key benchmarks. The Muse Spark release sits within a broader restructuring: Meta shut down its Llama API service on July 6 (ending a 14-month experiment in selling API access), adopting a dual-track strategy where open-source Llama continues for the community while closed-source Muse powers Meta's proprietary ecosystem across WhatsApp, Instagram, Facebook, and smart glasses. Meta also released Muse Image (codename Mango), a generative image model deeply integrated into its social platforms.
β Meta Β· TechCrunch Β· 36Kr
π TechCrunch via 36Kr Β· Meta Muse Spark Blog
ZCode: Z.ai Launches Agentic Development Environment
Beijing-based Z.ai (formerly Zhipu AI) launched ZCode on July 2, an "Agentic Development Environment" purpose-built around its flagship GLM-5.2 model. The desktop application directly challenges Cursor, Claude Code, and GitHub Copilot by organizing work around multi-step "Goal" tasks with multi-agent collaboration β allowing several AI agents to work in parallel on different parts of a project simultaneously. The environment ships with a built-in file manager, terminal, Git panel, and live browser preview, plus MCP (Model Context Protocol) integration.
A distinctive feature is support for remote task management via WeChat and Feishu messaging bots, enabling developers to trigger and monitor long-running coding tasks from mobile devices β a design choice reflecting workflows common in Chinese enterprise environments. GLM-5.2, the model powering ZCode, is a 753B-parameter MoE (approximately 40B active) under MIT license with 1M-token context window and 131K output tokens. Its IndexShare sparse-attention technique cuts per-token FLOPs by 2.9Γ at full context, enabling API pricing at $1.40/M input tokens β roughly one-third of Claude Opus 4.8. Subscription plans start at $18/month.
β Z.ai Β· TechTimes
π ZCode Launch Coverage Β· ZCode on moccet.ai
OpenSquilla 0.4.0: AI Coding with Self-Verification
OpenSquilla, an open-source AI Agent framework from Shanghai-based startup εΊε εΎε¨ (valuation $100M), released version 0.4.0 on July 1 β introducing a "self-verification" mechanism that may be the most important trust innovation in AI coding this year. The core idea is a red-green-regression evidence chain: the agent first writes a deliberately failing test that proves it can catch the bug, then implements the fix to turn the test green, and finally runs the project's existing test suite to confirm nothing broke. All three must pass before delivery.
In the demonstration, OpenSquilla's Coding mode added correct gradient computation to Andrej Karpathy's micrograd library β with forward values and every gradient matching PyTorch to 10 decimal places. The framework also includes automatic repair loops (revert-and-retry on failure), isolated sandbox execution (changes happen on a private fork), and a learnable cost-routing system that claims 60β80% cost reduction by automatically selecting cheaper models for simpler tasks. An accompanying signed desktop installer supports macOS and Windows. The project's GitHub stars grew to 5,300+ within weeks of launch.
β OpenSquilla Β· ζΎζΉζ°ι» Β· Baidu Baike
π OpenSquilla Baike Β· 36Kr Coverage Β· GitHub
MetaSkill-Evolve: Recursive Self-Improvement for LLM Agents
A research paper from LMU Munich and collaborators published on arXiv (July 6) introduces MetaSkill-Evolve, a two-timescale framework that makes LLM agent skill improvement recursive rather than one-shot. Current self-improving agents rewrite their task skills from execution traces β but the improvement procedure itself remains hand-authored and fixed. MetaSkill-Evolve breaks this ceiling by parameterizing the entire improvement pipeline (Analyzer, Retriever, Allocator, Proposer, and Evolver) as a meta-skill that evolves on a slower timescale while task skills evolve on a faster one β both using the same frozen backbone model, with no additional training.
The results are notable: MetaSkill-Evolve outperformed no-skill, static-skill, and single-level evolution baselines across three agentic benchmarks, improving held-out test accuracy over the raw backbone by +23.54 points on OfficeQA, +16.09 on SealQA, and +1.92 on ALFWorld. The implications extend beyond benchmarks β a framework where agents can evolve not just what they do but how they improve themselves suggests a path toward genuinely autonomous long-term capability growth without human redesign of the improvement loop.
β arXiv Β· LMU Munich
π arXiv:2607.05297
Next digest: July 17, 2026
Top comments (0)