vibehacker
News
Hugging Face ·

Microsoft ThinkingBox: grades agents on DB state, not chat replies

Microsoft and Hugging Face released ThinkingBox: 507 stateful business workflows graded on terminal backend state and side effects across 20 repeats—not tool-call trails or final replies. Run it via OpenEnv; Claude Opus 5.5 leads pass@1 (~67%), and ~80% of failures are tool-handling, not reasoning.

More news

View all

Baseten: Claude Code vibe-built VibeQwen beats vLLM by up to 90%

Baseten used Claude Code (Fable 5) plus MetaInfer skills to vibe build VibeQwen—a Qwen 3.6 35B A3B NVFP4 engine on one B200 that beat tuned vLLM 0.25.1 by up to 90% single stream decode (1,792 vs 943 TPS) and cut TTFT 28→12ms; a follow on SAM 3.1 server gained 50% throughput. Experimental only—not serving production yet…

Baseten

OMP 18.6.0: DeepSeek DSML tool calls on every host; Antigravity models stick

Oh My Pi (OMP) 18.6.0 parses DeepSeek DSML tool call envelopes on every host—including llama.cpp, LM Studio, vLLM, and custom gateways—and strips unparsed DSML markup from history so a broken call retries instead of ending the turn. It also stops Antigravity Claude Opus 5.5 / Sonnet 5.5 from vanishing after omp models refresh (fallback client version now 2.19.1)…

GitHub / OMP

DeepSeek Harness v0.2: official Mac/Windows desktop apps for the MIT agent

DeepSeek’s MIT Harness (dsh) v0.2 public preview ships official Apple silicon Mac and Windows x64 desktop apps—Cordis “everything is a plugin” runtime with Creator mode (chat built plugins), file/diff sidebar, and scheduled Automation tasks. The installer bundles dsh (no separate Node/pnpm); web UI still via npx @deepseek ai/dsh web ; expect breaking changes…

MarkTechPost

Spotted something we missed? Start a thread.