Open Source AI Digest

Model Releases

• 🔥 Leanstral 1.5: Mistral released Leanstral 1.5, an Apache-2.0 open-weight Lean 4 code agent model solving 587/672 PutnamBench...

At the AI Engineer World's Fair, 7,000 engineers converged on harness and loops as the real battleground, not models.

• Open layers win: OpenAI...

DeepSeek continues as the key benchmark for open-weight frontier models, with its V4 family (1.6T-parameter Pro and 284B Flash variants) offering...

Poolside released Laguna XS 2.1 on July 2 as a free download on Hugging Face and free OpenRouter tier, retiring XS.2 on July 9.

• MoE design...

• ctx open-sources a Rust CLI that indexes months of multi-agent histories into local SQLite, delivering ~50x token-efficient searches
• Agents using...

Two Beijing labs are pushing high-performance open-weight coding models into mainstream tools, highlighting a clear cost-performance edge alongside...

• PAW compiles natural-language specs into 23MB LoRA adapters via a 4B compiler model, letting a frozen 600M interpreter handle repeated tasks...

Mature teams typically combine both: closed frontier models for hard tasks and open-weight models for high-volume work.

• Cost: Closed APIs suit low...

A new CLI tool lets developers bootstrap production-ready AI agent apps in under a minute.

• Install via uv/pipx, run wizard to generate FastAPI +...

Mistral AI is executing a dual strategy: shipping powerful open-weight models while raising massive capital to champion European AI sovereignty.

-...

OpenClaw, the AI agent formerly known as Moltbot and Clawdbot, launches on iOS and Android. This expands agent accessibility to smartphones, though the announcement urges caution.

Google's TabFM delivers zero-shot predictions on structured data by processing entire tables in a single forward pass through frozen weights,...

Model Releases

• 🔥 MiMo-V2.5 on DeepInfra: Xiaomi’s unified agentic+multimodal 310B MoE model with 1M context is now available on DeepInfra under...

Kimi K2.7 Code's arrival in GitHub Copilot highlights the open-weights vs open-source gap in enterprise tools.

• Cost edge: Moonshot's MoE design...

Open-source inference is advancing on two fronts at once.

• DSpark delivers 60–85% faster generation on DeepSeek-V4 with no quality loss, using...

Two fresh benchmarks reveal the push for smarter agent testing.

• PACE builds cheap proxy sets from non-agentic tasks that predict full agentic...

Chinese open-weight models are claiming leadership in targeted niches with strong cost and deployment advantages.

• GLM 5.2 outperforms Opus and...

Hugging Face CEO Clement Delangue signals strong optimism for the open-source AI ecosystem with his tweet: "Are you ready for the open-source AI summer™️?".

Enterprise Integrations

• 🔥 Kimi K2.7 Code in GitHub Copilot: GitHub made Kimi K2.7 Code, an open-weight model, generally available as a low-cost...

• LLMs make data referencing errors (DREs) by incorrectly citing or omitting table values, even when they understand structure
• These errors...