Gemini Takes Video, DeepSeek Hits 1M Context on One GPU, and AI Writes Its Own Kernels

Gemini Omni Flash Runs Away With Video

Google's Gemini Omni Flash has jumped to number one on the Video Arena leaderboard with a 1404 Elo rating, opening a 101-point gap over Seedance 2.0 Mini. That's one of the largest single leaps the leaderboard has seen, and it signals that text-to-video is maturing fast into a genuinely usable creative tool rather than a novelty.

DeepSeek V4 Flash Fits 1M Context on a Single RTX 5090

A new llama.cpp patch wires DeepSeek V4 Flash's lightning indexer into the model graph with a custom CUDA kernel, letting the model run up to a 1 million token context on a single RTX 5090. At 256K context the compute buffer drops from around 67 GiB, which used to overflow the card, down to just 3.2 GiB, while prefill speed jumps from 56 tokens per second to roughly 263. It's a striking example of software engineering unlocking capability that previously demanded far more expensive hardware.

An AI Agent Writes a Record-Breaking Megakernel

In a KernelBench-Mega result, an AI coding agent reportedly wrote the first authentic single-launch megakernel for a decode workload, hitting 18.7x over the reference implementation and beating prior multi-kernel entries. It fused attention, routing, mixture-of-experts, normalisation and more into one launch, and even benchmarked itself, reverting regressions along the way.

Test-Time Compute Reshapes the Scoreboard

The UK AI Safety Institute showed that starving an agent of tokens makes it look far weaker than it is. Give a frontier system 2.5 million tokens and its task horizon sits around 2 hours; give it 50 million and that horizon stretches to roughly 14 hours. How much thinking time you grant a model is now as important as the model itself.

Memory Becomes a Trainable Skill

Stanford's AutoMem work treats memory management as something a model learns rather than a bolt-on database, deciding what to store, retrieve and reorganise. Optimising memory alone reportedly delivered 2x to 4x gains on game-based benchmarks like Crafter and NetHack.

Continual Learning Gets Measured

ByteDance Seed's EdgeBench studies day-long tasks across 134 real-world environments and claims learning speed is doubling roughly every 3 months. Epoch's EBR-bench, where models replay a game to learn from failure, found little improvement without dedicated reinforcement learning, a useful reality check.

A Small Fine-Tuned Model Beats the Giants

A collaboration between Bridgewater and Thinking Machines produced a fine-tuned Qwen3-235B that reached 84.7% on a document-filtering task, outperforming frontier models prompted directly, at around 1/14th the inference cost. It's fresh evidence that specialised post-training can beat brute-force scale for narrow jobs.

Inference Keeps Getting Cheaper

The vLLM team reported a 5x drop in token costs on DeepSeek V4 in a single month, and a serving redesign for Qwen3-Omni's speech pipeline that delivers first audio in about 0.6 seconds instead of 6, with 5.4x more throughput.