Easy AI Daily News Digest | November 26, 2025

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily for November 26, 2025 covers major model releases including Black Forest Labs' FLUX.2 series (Pro, Flex, Dev with 32B open weights, and upcoming Klein), Anthropic's Claude Opus 4.5 with lower pricing ($5/$25 per million tokens) and strong coding performance, and Google's Gemini 3 API updates featuring reasoning depth control and a 93% GPQA Diamond score. Community highlights include FP8 reinforcement learning from Unsloth enabling training on consumer GPUs with 60% VRAM savings, FLUX.2 running on 24GB VRAM via 4-bit quantization, and Claude Opus 4.5 arriving on Perplexity Max and LMArena. Hardware news covers RTX GPU pricing trends and leaked NVIDIA B200 benchmarks. Research items include NVIDIA's Nemotron-Flash small models, the DiP pixel-space diffusion model with 10x faster inference, and LLM-as-a-Judge calibration studies.

Easy AI Daily News Digest — November 26, 2025

*Translated and edited from the zhichai.net forum post (source: Easy AI teaching project).*

Model Releases and Updates

Black Forest Labs Releases FLUX.2 Series

The FLUX.2 family ships in four tiers: Pro (API-only), Flex (quality/speed control), Dev (32B open weights), and Klein (open weights coming soon), plus a FLUX.2 VAE. The Dev weights are live on Hugging Face, supporting multi-reference-image generation and 4K resolution output.

FLUX.2 official blog
FLUX.2 Dev open weights

Anthropic Releases Claude Opus 4.5

$5/million input tokens and $25/million output tokens

Anthropic announcement

Google Gemini 3 Series Updates

93% on GPQA Diamond

Gemini 3 documentation

AI Twitter Highlights

Claude Opus 4.5 evaluation: Leads Gemini 3 Pro on SWE-Bench Verified; 96.5% accuracy on research tasks (paper QA, systematic reviews); supports BrowseComp-Plus tool calling. (scaling01, stuhlmueller)
Gemini 3 benchmarks: 93% record on GPQA Diamond, strong in organic chemistry; comparable text reasoning to Opus 4.5, better visual input, slightly weaker jailbreak robustness. (EpochAIResearch, hendrycks)
FLUX.2 ecosystem: Day-one support on Replicate, Together AI, Vercel AI Gateway; open-source pipeline on Hugging Face; OSTris AI offers day-0 inference/editing and LoRA training tools. (replicate, huggingface)

AI Reddit Highlights

FP8 reinforcement learning on consumer GPUs: Unsloth reports 1.4x training speedup and 60% VRAM savings; Qwen3:4B runs on 5GB VRAM. (discussion)
FLUX.2 on 24GB VRAM: Community members run FLUX.2 on an RTX 4090 using diffusers and 4-bit quantization with remote text encoders. (discussion)
Non-technical subreddit feedback on Opus 4.5: Users note improved complex coding ability but a more hedged response style; a benchmark chart comparing it favorably to Opus 4.1 drew criticism over its y-axis truncation. (ClaudeAI, GeminiAI)

AI Discord Highlights

Claude Opus 4.5 on Perplexity Max: Available to Max subscribers; detailed specs undisclosed, but users report strong coding/research performance. (announcement)
FLUX.2 on LMArena and OpenRouter: LMArena added flux-2-pro and flux-2-flex (text-to-image and editing); OpenRouter lists FLUX.2 [pro] and FLUX.2 [flex]. (LMArena, OpenRouter)
Unsloth FP8 RL with NVIDIA support: NVIDIA officially supports FP8 RL on Blackwell RTX-50 and DGX Spark with setup docs. (Unsloth, docs)

Hardware and Infrastructure

NVIDIA RTX GPU pricing trends: Used RTX 3090 (24GB) around $750; RTX 4090 at $2,000–3,500, with users favoring the 3090's value; RTX PRO 6000 Blackwell dropped to $7,999. (discussion)
Leaked NVIDIA B200 benchmarks: Under CUDA runtime with Torch 2.9.1+cu130, a 16384x7168 matrix multiply took 33.6±0.05 µs; 7168x4096 took 124±0.1 µs. (GPU MODE thread)

Research and Development

NVIDIA Nemotron-Flash: Uses evolutionary search over hybrid attention/operator mixes to improve the small-model accuracy–latency frontier: +5.5% accuracy over Qwen3-0.6B with 1.3–1.9x lower latency and 45.6x higher throughput. (overview)
DiP pixel-space diffusion: Two-stage DiT backbone with a Patch Detailer Head achieves 10x faster inference with only 0.3% parameter overhead; FID 1.90 on ImageNet 256x256. (overview)
LLM-as-a-Judge calibration: Most LLM-as-a-Judge results use biased estimators and require calibrating evaluator error rates; CoT explanations may increase blind trust and reduce error detection. (Kangwook Lee, Maarten Sap)

Community and Events

Psyche Office Hours: December 4 (Thursday), 1PM EST on Discord. (event)
DSPy Pune meetup: In-person gathering in Pune, India, announced via X. (announcement)
MCP Dev Summit: Upcoming, though some community members cannot attend due to scheduling conflicts. (discussion)

*Source: Easy AI teaching project.*