Tag
Qwen3.8-2.4T-A95B by Alibaba Qwen and Alibaba Cloud is now available on Modal, served with a custom DFlash speculator trained on tool-call-heavy data and a full 1M context window.
A user shares surprising results testing Liquid AI's new LFM2.5-VL-3B vision-language model across many languages, noting strong visual capabilities but weaker instruction following; Liquid AI announces the model can read screens, documents, and ground objects to coordinates.
This paper introduces the Latent Critic, a lightweight LoRA adapter that detects hallucinated agent actions in real time by restructuring the transformer's residual stream into localized natural-language feedback, achieving 0.966 AUROC and enabling self-correction.
Cactus Compute releases Needle 2, a 45M-parameter agentic LLM compressed to a 14MB binary for phones, wearables, smart home and robots, achieving 500+ tokens/sec on a Raspberry Pi 5 and running in 28MB RAM.
Cactus releases Needle 2, a 14MB agentic LLM for phones, wearables, smart home devices, and robots, achieving fast inference on low-end hardware and supporting structured extraction and fine-tuning.
Santiago Valdarrama reports that Kimi K3 is the best open-weight model he has tested, a 2.8T-parameter vision model with tool calling, reasoning, and a 1M context window.
Liquid AI released LFM2.5-2.6B, a 2.69B parameter model with 128K context and tool calling, optimized for multi-step agent workflows and capable of running at 30 tok/s on a phone with a 1.67GB Q4_K_M GGUF, though coding and knowledge-heavy tasks remain weak compared to larger models.
Liquid AI releases LFM2.5-2.6B, a compact agentic model designed for on-device deployment, supporting tool calling and multi-step workflows with efficient inference on CPUs and GPUs.
The author shares experience building a human-in-the-loop approval system for an enterprise agent platform, emphasizing that the approval step must be a true blocking pause with editable parameters and first-class rejection/editing outcomes, and asks how others structure agent suspension.
Introduces OoO-Spec, a method that accelerates LLM tool calling by computing semantic slots out of order via a small sidecar model, achieving up to 5.34x speedup over autoregressive decoding and outperforming existing drafter methods across multiple targets and benchmarks.
NVIDIA released NemotronLabs VoiceChat 11B, an open end-to-end full-duplex speech model enabling real-time conversational AI with ~450ms turn-taking latency, barge-in, and live tool calling, the first open full-duplex model to support tool calling.
The author praises DeepSeek Flash's greatly enhanced long-horizon and agentic abilities, which can automatically discover and combine subagent swarm tool calls in the harness.
This paper investigates the claim that 4-bit weight quantization is nearly lossless for LLM agents, showing that while aggregate benchmark scores stay flat, quantization amplifies existing tool-call failures (e.g., hallucination), hidden by the benchmark's error budget. The authors suggest reporting per-channel error rates and success under shrinking budgets to reveal the masked damage.
Ant Group's AntLing-3.0-flash model is now live on OpenClaw via OpenRouter, requiring only a model ID setup. It features 256K context, RL training for long-horizon tool calling, is API-only, and free through August 3.
Grafana released an AI SDK for Go that provides a unified API for LLM calls, streaming, tool execution, and structured output, with built-in compatibility with Vercel's AI SDK React frontend.
A developer built a GBNF grammar compiler in Rust for llama.cpp that enforces tool-calling JSON schemas, making small models (like 8B parameters) reliably call tools. The system narrows grammar to only matched tools per turn, improving reliability.
The author shares a strategy for reducing costs in multi-model agent loops by using a cheap fast executor for repetitive nodes and a strong planner for high-level reasoning, with experience using Ling-3.0-flash on OpenRouter.
The author benchmarks local Qwen models of various sizes on a 4GB RTX 3050 Ti laptop GPU within the Bike4Mind workspace, finding the 2B model at Q4_K_M quantization is the sweet spot for fitting in VRAM, achieving 96 tok/s. Smaller models struggle with tool selection, artifact generation requiring multiple models, and RAG embeddings causing model swap overhead.
A practical technique for preventing AI content generators from making up values: provide the allowed list as a tool/function that the model can call, rather than embedding it in the prompt as plain text.
Evaluation of Laguna-S-2.1 against Qwen3.5-122B on RTX Pro 6000 shows it is the fastest 100B+ model tested and best at tool calling, but prone to inventing facts under pressure.