Open weights arrive with their training bills attached, coding agents get both new powers and new warnings, and an IDE veteran tries to govern the whole agentic workflow. Seven stories, each summarized in our own words with a link to the original reporting.
Xiaomi open-sources MiMo-V2.6 and publishes the $3.47M reinforcement-learning bill
Xiaomi released the MiMo-V2.6 series as open weights, led by a Pro mixture-of-experts model with about a trillion total parameters alongside a smaller Flash sibling, both handling text, image, video, and audio in a long context window. The unusual part is not the weights but the paperwork: the team streamed its reinforcement-learning run in public, reporting roughly 750,000 trajectories per model over six days with itemized costs near $2.62 million for Pro and $850,000 for Flash, including restarts caused by memory and infrastructure failures. The package also includes thousands of training environments, the RL framework pieces, and a small distilled model for further experiments.
For builders, the release is two assets in one: capable permissively licensed checkpoints plus a rare public record of what post-training at this scale actually costs and where it breaks. Treat vendor benchmark figures as progress markers from the team's own harness rather than independent leaderboard entries, and verify them on your own workloads before committing.
Read the full story at TechNode (September 22, 2026).
China Telecom ships Xing4.0-29B, an agentic model that fits on one consumer GPU
China Telecom AI released Xing4.0-29B-A4B, a 29-billion-parameter agentic model with only 4 billion parameters active per token, aimed at running tool-using agent workloads outside the data center. With low-bit quantization the team says it fits in about 15 GB of memory, so a single consumer-grade graphics card can host it while the same weights also serve enterprise deployments. The model is available through public code and weight hubs for developers to download and try.
The builder signal is the shrinking gap between hosted agent APIs and hardware you already own: planning, tool calls, and multi-step tasks no longer strictly require a cluster. As always with vendor releases, confirm the agentic benchmarks on your own tasks and check the license file before shipping anything commercial.
Read the full story at Markets Insider (September 22, 2026).
MiniMax open-sources MiniMax Code, a terminal coding agent under MIT
MiniMax published the source for MiniMax Code, a terminal coding agent that reads projects, edits files, and runs tests from the command line, with both an interactive interface and a headless mode for scripts and CI. The first-party code carries the MIT license and speaks the Agent Client Protocol, so it can plug into editors that support the standard, work with the vendor's accounts or third-party model endpoints, and run against bring-your-own keys. The separate desktop application stays closed and is distributed on its own.
For local-first teams this is a practical new option next to the closed CLIs: inspectable terminal-agent code you can fork, wire into pipelines, and point at your own models. Review the repository scope before adopting, since only the terminal, headless, and protocol components are open.
Read the full story at TechNode (September 21, 2026).
JetBrains bundles its agent tooling into JetBrains Air: IDEs, teams, governance
JetBrains gathered its agentic development work under one banner called JetBrains Air, spanning the coding experience inside its IDEs, a new team-coordination layer, and a governance layer for policy, auditability, and AI cost control. The pitch is that generating code was the easy part and the real bottleneck is verifying, coordinating, and accounting for agent work across an organization. The system stays open to third-party models and agents through a standard client protocol plus a registry for discovering compatible agents.
Teams drowning in agent sprawl should watch the governance and shared-context pieces rather than the code-generation demos: one policy set, one cost view, and provenance that survives switching vendors. Availability rolls out in stages, so confirm what is shipping versus preview before planning a migration.
Read the full story at The JetBrains Blog (September 22, 2026).
Plugin4Shell: one git naming trick defeats plugin locks on four coding agents
Researchers at Air Security disclosed a flaw affecting plugin installation in Claude Code, Codex, GitHub Copilot, and the Gemini CLI, where the SHA pinning meant to lock each plugin to reviewed code can be bypassed with crafted branch names. Because some agents refresh plugins automatically in the background, a trusted plugin could be swapped for malicious code with no click from the victim, running with the developer's own file and credential access. Two vendors have shipped fixes while one ships none and the last tool is being retired rather than patched.
The operator takeaways are immediate: update the affected agents, audit which marketplaces your plugins come from, and treat auto-update plus broad plugin permissions as a real attack surface. The deeper lesson is architectural: the same design error appeared in four products from four companies, so verify the update mechanism itself instead of assuming a hash check means safety.
Read the full story at The Next Web (September 21, 2026).
Claude Code reads AGENTS.md, ending the two-instruction-file era
Anthropic added support for the shared AGENTS.md instruction format in Claude Code, so projects without a CLAUDE.md file now have their agent instructions picked up from the same file that Codex and other tools already read. Teams running several coding agents on one repository can stop maintaining two parallel instruction files and drifting out of sync. Project-specific Claude features and a feature-flag dependency for some enterprise setups remain caveats, with configuration options to control which files are read.
The practical move is simple: consolidate shared conventions into AGENTS.md and keep genuinely Claude-specific instructions separate where supported. Small standardization wins like this compound across every agent run in CI and on every developer machine.
Read the full story at InfoWorld (September 21, 2026).
Grok 4.7 targets coding and knowledge work at lower prices
SpaceXAI introduced Grok 4.7 as its strongest model for coding and knowledge tasks, available through its editor integrations, build tools, API, and third-party platforms. Pricing starts at a few dollars per million tokens with a faster tier trading double speed for double price, while running on a larger base model at the same cost and speed as its predecessor. The release frames throughput and price per task, not just benchmark scores, as the headline.
For builders choosing a coding backend, evaluate it the same way the vendor prices it: timed runs on your own repositories measuring tokens per merged change rather than headline claims. A cheaper fast tier is only a win if output quality holds on your stack.
Read the full story at Dataconomy (September 22, 2026).