DFlash 2 is here! Qwen3.8-27B at 70 tok/s on an M5 Max MacBook Pro.
⚡ Up to 4.6× the speed of autoregressive decoding, with the same output.
This is the next generation of DFlash, seeded at Z Lab and upgraded at Inco AI. Get one more accepted token on every pass, for free!
https://t.co/We0lwYPSBl
We will be at the ACM Conference on AI and Agentic Systems (CAIS) tomorrow @CAISconf
Come find the OpenMesh team at our poster station. We’d love to talk about your agentic workflows.
Posters: “AgentPulse”+“IntelligenceArena”
Multi-agent systems are now routine, but one decision goes unmeasured: when an orchestrator hands a subtask to a peer model, does it pick the right one?
We built DecisionBench to score the handoff itself, not just the final answer, across an 11-model, 7-vendor pool over 23,375 task instances. The headline finding: there is enormous performance left on the table. A perfect-delegation ceiling sits 15 to 31 points above where today's systems land, on every benchmark we ran. That gap is the prize, and skill-aware orchestration is how you capture it.
Read more about it in our blog post:
https://t.co/U8dbTlWMNF
As LLMs become increasingly powerful, the ability to effectively delegate and collaborate with other models may itself become a critical capability for long-horizon agentic systems.
It was a lot of fun working with these passionate young researchers on DecisionBench! @MichaelYGao@meganwxng@YYL_GB
Three strong signals today: DecisionBench (delegation quality), UCCI (calibrated cascade routing), and Gemini 3.5 Flash (agentic speed). Common thread: advantage is moving from model IQ to orchestration policy—who handles which step, at what confidence, at what cost.
Excited to share what we’ve been working on @OpenMeshAI
DecisionBench aims to measure how well LLM agents emergently delegate sub-tasks to peer models during long-horizon workflows, using process-level metrics that quality-only evaluation misses.
just went back to the data, and Gemini-3.1-Pro turns out to have a serious thing for frontier peers: of its 86 delegations, 48 go to GPT-5.5 and 23 to DeepSeek-V4-Pro
the kicker is that Flash actually beats GPT-5.5 on the 2 skills Pro delegated the most, right on the profile card Pro just read, but Pro picked 5.5 anyway!
Your multi-agent system might be delegating to the wrong model on every single subtask, and your eval would never tell you 🧵
Introducing DecisionBench: the first benchmark that measures whether agents actually delegate well, not just whether the task got done.
https://t.co/rRQbP930OB
DecisionBench fixes only the tasks, model pool, interface, and metrics, so learned routers, multi-step delegation, and adaptive profiles all evaluate on the same harness. Bring your own method and compare head to head.
11 models, 7 vendors (Claude Opus 4.7, GPT-5.5, Gemini-3.1-Pro, DeepSeek V4, and more). We release the substrate, annotation layer, reference suite, analysis pipeline, and 220 run archives.
Joint work with @meganwxng , @YYL_GB , @CarlMaZijian , and @ao_qu18465
Supported by @OpenMeshAI
Code + data: https://t.co/7ILSKYC971
How much is left on the table? A perfect-delegation ceiling sits 15 to 31 points above measured performance on every suite, mostly on weak and mid-tier agents.
And the agents have their own bias: some over-delegate to their own vendor at up to 3.65× chance. The delegation analogue of LLM-as-judge self-preference 👀