Tag: #agentic16 results
Anthropic launches Claude Opus 4.7, its top model with 64.3% on SWE-bench Pro (vs GPT-5.4's 57.7%), multi-agent coordination for long workflows, 3x image resolution, and 14% better agentic reasoning with fewer tool errors. Pricing: $5 input/$25 output per million tokens.
Gemma 4 31B achieved 100% survival and +1,144% median ROI on FoodTruck Bench, outperforming GPT-5.2, Gemini 3 Pro, and all tested Chinese open-source models. Only Opus 4.6 beats it, but at 180x the cost of $0.20 per run. Best cost-performance ratio among 22 tested models.
After 2025 struggles, Kimi launches K2 agentic model in July 2025 and K2.5 2.5T multimodal in 2026. Attention Residuals paper challenges Transformer foundations, earns Karpathy praise. Adopted by Cursor (scandal), Perplexity, Cloudflare.
Mistral Small 4 PR on transformers unifies Instruct, Reasoning, and Devstral into one hybrid model. Features 119B MoE with multimodal input and configurable reasoning. Open-source under Apache 2.0 with agentic capabilities.
A user tasked Qwen3.6-35B with building a tower defense game using screenshots from MCP, and it successfully implemented and tested features like upgrades. The model self-detected and fixed bugs in canvas rendering and wave completions. High excitement for the upcoming Qwen Coder model.
Liquid AI released LFM2.5-350M, a 350M parameter model optimized for data extraction, tool use, and agent workflows. Trained on 28T tokens with RL, it beats Qwen3.5-0.8B while fitting <500MB quantized. Runs on CPU/GPU/mobile with reliable outputs.
Cursor has released a new AI coding evaluation benchmark that surpasses SWE-Bench. It specifically assesses the 'agentic' intelligence of different models within Cursor. Claude performs poorly on this challenging new standard.
User shares 'Browser OS' implementation using Qwen 3.6 35B, calling it the best result from any local model. Post links to details but content is minimal.
Alibaba released CoPaw-Flash-9B, an official agentic finetune of Qwen3.5 9B. It matches Qwen3.5-Plus performance on some benchmarks. Available on Hugging Face.
Autonomous pipeline generates playable Godot games from text prompts using LLMs. Overcomes GDScript scarcity with layered references, agentic lazy-loading, and multi-stage verification including screenshots. Open-source code for reproducibility.