Thinking Machines ships Inkling 2 Small, Qwen3.8-27B runs locally for Copilot, Mistral Studio gets managed agents

Inkling 2 Small is Thinking Machines Lab's second-generation model. It's a mixture-of-experts model with 594B total parameters and 17B active, and it takes image input and supports adjustable effort. For now the weights are closed: you can use it in the Tinker Playground and for sampling and fine-tuning on Tinker, the company's training platform.

The company reports a score of 31 on the Artificial Analysis Intelligence IndexVendor claim, well above its first generation, plus 86.3% on SWEBench Pro v2 (Hard)Vendor claim and 72.1% on DeepSWE v1.1Vendor claim. It plans to release the weights under Apache 2.0 next week once external safety testing wraps up, which would add another large open model for coding and agent work that you can use commercially. The company also says early checkpoints fabricated results and hid actions, and that the model still trails the top closed models on alignment evals.

Microsoft published a setup for running Qwen3.8-27B locally with Foundry Local, its on-device inference tool, and connecting it to GitHub Copilot through BYOK (bring your own key). It runs on ONNX Runtime and uses speculative decoding, where a small drafter model guesses several tokens ahead, plus KV-cache quantization to keep it fast and within memory. You need Foundry Local CLI 0.13.0 or later. Install it, load the model, then pick Foundry Local under Model providers in the Copilot app.

winget install Microsoft.FoundryLocal
foundry --version
foundry model load qwen3.8-27b
foundry server status

At 4-bit the model shrinks to about 18 GBVendor claim, which Microsoft says fits a 24 GB GPU with up to 64K context, or the full 262K on 32 GB and up. Microsoft lists Qwen3.8-27B at 73.0 on Terminal Bench 2.1Vendor claim and 61.7 on SWE-bench ProVendor claim. Teams that can't send code to the cloud, or developers who work offline, get a coding agent with no per-token bill.

Mistral added managed agents and app hosting to Mistral Studio, its enterprise platform. Agents can carry skills that encode a team's procedures, sandboxes for running code, subagents to hand off parts of a task, and hooks for custom logic during execution. Workflows resume where they stopped, can start from enterprise system events, and show their logic as a graph. A new Mistral MCP server lets you build with Mistral Vibe or your current coding agent and deploy to Studio with one command.

The same agent can run in Mistral's cloud, your own VPC or on-premises, which matters for banks and public bodies that can't move data out. Tracing and evaluations ship in the box, so there's less to bolt on when a proof of concept goes to production. Mistral says BNP Paribas, Capgemini and CMA CGM already run Studio in production.

Anthropic shipped a batch of security and admin features for Claude Enterprise. A security center compares your chat, Cowork and Claude Code settings against Anthropic's recommended configuration on one page, and you can download it as an audit report. Inference hooks are now generally available: they send prompts and tool calls to your existing DLP (data loss prevention) tools for a check before they run. There's also a central admin console at admin.claude.com, an Admin API call to sign a member out everywhere, and a way to block personal Claude accounts on managed machines.

The usual blocker for rolling out coding agents is whether existing security tools can see and stop what the agent sends and runs. GA inference hooks answer that. On the API side, a new beta lets an org allowlist up to 1,000 IPs or CIDR ranges, so a leaked key is useless outside them. Security teams can also hand auditors the security center report instead of collecting settings by hand.