Wednesday, November 18, 2026


Funding and industry
Companies and models
Model adaptation

Hugging Face releases a Qwen3.8‑27B adapter for Pi

The team trained a LoRA adapter with reinforcement learning for the Pi coding agent. In the team’s pass@1 tests using the same harness, SWE‑bench Pro Hard 51 rose from 29.4% to 62.7%, and Terminal‑Bench 2.1 rose from 39.5% to 55.8%.

Agent products
Robotics and simulation

Antigravity plugin connects robot arms and MuJoCo

Google DeepMind demonstrated natural-language control of a supported robot arm with Gemini 4 Hydrogen and the Antigravity robotics plugin. Users can first try the task in MuJoCo and then run it on physical hardware.

Developer tools and APIs
Tool use · @ClaudeDevs

Claude Code experiments with chaining tools in scripts

Code Mode lets Claude write scripts that connect multiple tools and keep responding while long commands are still running. Set CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 to opt in.

Billing · @ClaudeDevs

Claude Platform tests linked billing

Customers of tools such as Factory, Linear and Ramp can charge Claude API usage to their own Claude Platform organization without giving the integration an API key.

Monitoring · @OpenAIDevs

OpenAI updates its API usage dashboard

Analytics lets teams track usage and spending, inspect rate limits and rate-limited requests, and investigate errors by model, project or API key.

Model gateway · @perplexity_ai

Perplexity introduces the Orchestrator API

The API brings open-weight and proprietary models under one interface and bill. Callers can choose a model, while the service selects providers based on availability, capacity and latency and supports failover.

Agent teams · @perplexity_ai

Perplexity rolls out Multiagent

Computer and Teammates are getting a Multiagent effort setting. A lead agent delegates work, agents using different models trade findings, and the lead assembles the answer.

Local execution · @replit

Replit Desktop gives Agent access to selected folders

Users choose which folders the Agent can use to build and run apps. On supported Windows and macOS devices, Microsoft Execution Containers provide a local sandbox.

Open source · @replit

Replit open-sources DataCat for self-hosted observability

DataCat handles metrics, logs, traces, dashboards and alerts, and accepts OpenTelemetry data. It supports PostgreSQL and ClickHouse and can be self-hosted from the public repository.

Framework · @rauchg

Vercel releases eve 1.0

The open-source framework is generally available and targets agents that keep running across multiple steps. Developers can integrate and deploy their own agents in applications.

Enterprise tooling · @gneubig

OpenHands ships a lighter enterprise edition

The OpenHands team highlighted using cloud coding agents for security audits and remediation, software testing, and code review.

Desktop apps · @ClaudeDevs

Claude Desktop adds a Latest release channel

Windows and macOS users can opt into Latest in Desktop app settings for earlier fixes. Stable remains the default, and Latest builds can be less reliable.

Plugins · @sarahookr

Adaption adds a Hermes Agent plugin

Adaption announced a plugin for preparing data and training custom models from inside Hermes Agent.

Video generation research

Apple releases FastFrames video generation models

The model cards list FastFrames‑S at 1.3 billion parameters and FastFrames‑L at 14 billion. Both use two-step autoregressive video diffusion. The published configuration generates 832×480 video at 16 fps for 81 frames, with interactive latency as the project’s goal.

Inference and evaluation
Open inference · @sgl_project

SGLang summit recap details version 1.0 and nine founding members

The SGLang team recapped the announcement of SGLang 1.0, Miles 0.2 for reinforcement learning and the SGLang Consortium. Its founders include NVIDIA, AMD, Intel, Google, Meta, Thinking Machines, Cloudflare, Modal and RadixArk.

Speculative decoding · @vllm_project

vLLM updates DSpark for Kimi K3.1 Pro

Inferact and Red Hat collaborated on the revised speculative decoding draft model. vLLM reports up to 30% longer accepted drafts than the preview on HumanEval and a cap on the draft model’s KV cache.

Streaming video · @vllm_project

vLLM‑Omni adds day-one support for AURA 2.0

The 9-billion-parameter video model follows a scene as it unfolds, can initiate a conversation and can call tools in response to what it sees.

Evaluation · @rauchg

DeepsecBench adds GPT‑6.2 vulnerability detection results

Vercel added GPT‑6.2 to its application security leaderboard. Under Codex, Astra at xhigh reached 56.2% recall and 95.0% precision, with a reported test cost of $59.85.

Model access · @perplexity_ai

Perplexity reports per-task costs for both GPT‑6.2 models

Both models are in Search and Computer. On Perplexity’s WANDR benchmark, Astra scored 0.783 at about $12.54 per task and Sol scored 0.746 at about $2.65 per task.

Security · @gdb

Calif uses GPT‑6 Cyber to help find a Java flaw

Greg Brockman said security team Calif used GPT‑6 Cyber to help discover a major Java security flaw under the Patch the Planet initiative.

Caching · @ollama

Ollama outlines cache pitfalls in long coding sessions

Ollama shared observations from Claude Code, Codex and Pi. Timestamps in system prompts and changes to reasoning effort can invalidate the cache. Telling the agent that output was truncated can prompt it to read what it missed.

Inference architecture · @woosuk_k

Inferact revisits pipeline parallelism for prefill

The team argues that long-running agent tasks may change prefill traffic patterns. Idle time from pipeline bubbles is worth measuring again under those workloads.

Draft models · @_akhaliq

NVIDIA releases a speculative decoding draft for Qwen3.8‑27B

The 4-billion-parameter draft model proposes candidate tokens in batches for the full Qwen3.8‑27B model to verify.

Scientific reasoning · @vllm_project

vLLM supports Sage S2.1 Flash

The model combines scientific inputs such as molecular graphs and protein structures with retrieval, coding and tool calls. It can check intermediate results and refine its next step.

Research and engineering
Paper · @_akhaliq

An energy function gives recurrent reasoning a stopping signal

Energy‑Based Reasoners learn an energy function over latent reasoning states. Gradient descent refines those states and helps decide when to stop. The paper compares the method with earlier recurrent models on four algorithmic reasoning benchmarks.

Data · @_akhaliq

Ropedia Xperience 2 aligns scenes with human motion

The paper releases 1,589 metric-scale 3D scans and 4,230 first-person recordings covering 74 tasks over 131 hours. Hand and full-body motion is aligned to the scenes for simulation and robot trajectory retargeting.

Benchmark · @_akhaliq

Accelerate Bench tests whether agents can speed up programs

The benchmark includes 32 CPU program optimization tasks, with independent verification of correctness and measured speedup. Across 16 model-and-harness configurations, the authors report Score@1 results from 25.5% to 67.4%.

Video understanding · @_akhaliq

V‑JEPA‑LM trains on counterfactual video to read motion

The work pairs a frozen V‑JEPA 2.1 encoder with language models. The experiments cover models from 1 billion to 8 billion parameters and use counterfactual video to train for motion understanding.

Isolation · @simonw

Simon Willison demonstrates external sandboxes

The coding agent stays on the Mac, while its generated code executes in WebAssembly, QEMU or Apple Container. The demonstrations cover Claude Code, Codex and Pi.

Training recipes · @rasbt

Raschka discusses constant learning rates and weight averaging

Raschka noted that some pretraining recipes keep the learning rate constant and average late checkpoints or maintain a running weight average. That removes the need to schedule learning-rate decay ahead of time.

Hands-on test · @emollick

Mollick reports uneven results with Astra

Ethan Mollick found GPT‑6.2 Astra harder to steer in early use. In the same trials, it could tackle a demanding prompt to build a space elevator scene from scratch using Blender and Unity.

Demo · @simonw

One prompt, two GPT‑6.2 takes on a Three.js scene

Simon Willison gave Astra and Sol the same prompt: use OpenStreetMap data to turn Yonkers into a mech that walks through several New York City boroughs. He shared both outputs.

Creative tools

© 2026 FAV0 · AI Daily