Ollama is getting a native /v1/systemone decision-model endpoint
Jev-style scoring goes mainstream: open engines give any local model calibrated yes/no and score judgments today — plus fresh llama.cpp and Qwen Code drops.

Copy markdown
The endpoint: read a decision, don't parse one
Ollama PR #18606 adds POST /v1/systemone — submit context plus labeled questions and get three typed answers back: choice (a label with per-option probabilities), noul (probability a condition is true), and score (an expected value over a rubric). It reads next-token logits directly, so there's no JSON to coax out or regex-parse. Caveat: it's still in review and Nimble-via-MLX only for now — but the signal is that the runtime you already run is about to speak the decision-model protocol Nowline readers have tracked through Tev1 and GLM-Flash.
Don't wait for the merge — serve it today
jev-rs already exposes a Jev-compatible /v1/systemone over any GGUF via llama-server (plus DeepSeek, vLLM and SGLang backends), returning calibrated noul/choice/score in a single prefill; it's MIT/Apache-2.0 and plugs into Claude Code, Codex and OpenCode over MCP. Kev does the same, self-hosted on Ollama or any OpenAI-compatible endpoint. The weekend build: swap a brittle 'answer in JSON' router or LLM-as-judge prompt for probabilities you can actually threshold.
Elsewhere: llama.cpp adds Ling 3.0 VL
The latest llama.cpp builds land Ling 3.0 VL support, plus a CUDA path for Nemotron 3's Puzzle SSM and tiled mul_mat for k-quants on CPU — a new vision-language model to run locally and faster quantized inference on the hardware you own.
Elsewhere: Qwen Code v0.24.6 supervises agent crews
Qwen Code shipped v0.24.6 across CLI, Desktop and the TypeScript SDK with persistent shared-thread agent collaboration — interjection, cancellation and review history — so you can steer a running fleet of agents mid-task instead of firing one off and hoping.
Elsewhere: Pi v0.87.1 gets Codemode + MCP
Pi merged Codemode sandboxed execution and MCP tool support in v0.87.1, alongside an experimental Virtual Models feature — widening what the agent harness can safely run and which tools it can reach.