The pulse of open source AI

Tracking the tools that matter — terminal agents, coding models, autonomous systems, and the infrastructure behind them.

TerminalCrush7 May 2026

Crush (formerly OpenCode) Relaunches as a Fully Autonomous Terminal Agent

The open-source terminal coding agent previously known as OpenCode has rebranded to Crush and shipped a major architecture overhaul. The new release moves away from a pure code-completion model toward a persistent agent loop that can plan, execute, and verify multi-step tasks entirely from the terminal. Crush now supports tool use natively — including file system access, shell execution, and web search — without requiring a separate orchestration layer. The team behind Crush says the rename reflects a broader pivot: rather than assisting developers with individual edits, Crush is designed to own tasks end-to-end. Early benchmarks on SWE-bench show it outperforming several proprietary alternatives at a fraction of the cost when paired with DeepSeek V3 as the underlying model.

GitHub
ModelsDeepSeek6 May 2026

DeepSeek V4 Preview Drops with 256K Context and Improved Reasoning

DeepSeek has quietly released a preview of V4 to API customers, and the early reports from the developer community are causing significant excitement. The new model extends the context window to 256K tokens — double V3's 128K — while reportedly delivering substantial improvements on multi-step reasoning benchmarks. Notably, DeepSeek V4 appears to close the gap with frontier proprietary models on tasks requiring long-horizon planning, making it particularly attractive for autonomous agent workloads where context length is a hard constraint. Pricing remains in line with V3: roughly $0.14 per million input tokens and $0.28 per million output tokens, keeping it dramatically cheaper than GPT-4o or Claude Sonnet for high-volume agent pipelines.

DeepSeek API
ModelsQwen35 May 2026

Qwen3 Released: 235B MoE Model with Switchable Thinking Mode

Alibaba's Qwen team has released Qwen3, the third generation of their open-weight model family, and it represents a significant leap forward. The flagship Qwen3-235B uses a Mixture-of-Experts architecture that activates only 22B parameters per forward pass, keeping inference costs low while matching frontier model quality on most benchmarks. The most interesting new feature is a switchable "thinking mode" — similar to the approach used in DeepSeek R1 — which lets the model toggle between fast instruct responses and slower, more careful chain-of-thought reasoning depending on the task. Developers can control this via a simple system prompt flag. The full suite includes models ranging from 0.6B to 235B, all released under the Apache 2.0 licence.

HuggingFace
TerminalAider4 May 2026

Aider v0.90 Ships Architect Mode for Large-Scale Refactors

The popular AI pair-programming tool Aider has released version 0.90 with a new "Architect mode" that separates the planning step from the editing step across two distinct model calls. In architect mode, a high-capability model (such as o3 or Claude Opus) first produces a detailed implementation plan describing which files to change and how, and then a faster, cheaper model (such as DeepSeek V3 or Qwen3-32B) executes the actual code edits. This two-stage approach significantly reduces hallucinations on complex multi-file changes while cutting costs by 60–70% compared to using a single frontier model throughout. Aider v0.90 also introduces improved repo-map generation that handles monorepos more gracefully.

GitHub
ModelsCodestral3 May 2026

Mistral Releases Codestral 25.01 with Improved Fill-in-the-Middle

Mistral AI has updated Codestral to version 25.01, their specialised code generation model. The update brings meaningful improvements to fill-in-the-middle (FIM) completion — the capability that allows the model to insert code given both the prefix and suffix context around a cursor position. This makes it particularly well-suited for IDE plugin integration where the model needs to complete code mid-function rather than simply appending to the end. Codestral 25.01 also shows strong performance on the recently published Multi-SWE-bench evaluation, and Mistral has added support for 30 additional programming languages including Kotlin, Swift, and Elixir.

Mistral AI
ModelsPharia2 May 2026

Aleph Alpha Open-Sources Pharia-1, Europe's Privacy-First 7B Model

German AI company Aleph Alpha has open-sourced Pharia-1, a 7B parameter language model designed specifically for enterprise deployments where data sovereignty is a legal requirement. The model was trained entirely on European infrastructure using data that meets GDPR compliance standards, and the weights are released under a licence that permits commercial use within EU member states. While Pharia-1 does not compete with frontier models on raw benchmark scores, it is specifically optimised for German, French, Italian, and English, making it the strongest openly available model for multilingual European enterprise tasks. Aleph Alpha is positioning this as a building block for organisations in regulated industries — finance, healthcare, public sector — where sending data to US-based API providers creates compliance risk.

Aleph Alpha

Sign in for the full feed, full articles, and category filters