Encoder-decoder is back 😈!!! In DeepSeek V4.1 Flash the first 20 layers build the global KV that the next 20 read from, so prefill costs about half. Rolled out over the last 24h: throughput doubled, and we are approaching 1T tokens/day on OpenRouter. 30% off to celebrate. Enjoy. Cheapest on the market, as always.

A compute-based Vulkan APV encoder has just been merged into FFmpeg. It runs at 1200fps on a 6900XT for 1080p422 10-bit encoding. Meant for low-overhead high-quality recordings, or low-latency 4K streaming over a 1Gbps link.

TL;DR Advanced encoders were losing to older models on sparse retrieval for one surprising reason: a vocabulary mismatch. Fixing it with just 20k adaptation steps sets a new BEIR state of the art. Title: Why Advanced Encoders Lag on Sparse Retrieval? The Answer and an Approach to Bridging Vocabulary Gaps URL: Key points 🔤 Raw case-sensitive vocabularies were splitting single meanings across redundant surface forms 🧩 The proposed Vocabulary Transfer method migrates vocabularies in 3 lightweight steps 📊 ModernBERT-VT hits nDCG@10 52.4 on BEIR, a new SOTA 💥 A collapsed RoBERTa-large jumps from BEIR score 1.4 to 51.3 ⚡ Uses under 0.2% of the original pretraining token budget 🧪 Near-optimal performance reached in just 500 MLM steps 🧬 Also works for domain-specific vocabularies like chemistry It's a nice reminder that what looked like an architectural ceiling turned out to be a fixable vocabulary design problem. #InformationRetrieval# #LLM#

Seeing InternVideo-Next explicitly characterize its architecture as "Encoder–Predictor–Decoder (EPD)" and defining the Predictor as a "latent world model" confirms a clear convergence in the field. 📉 From our early work on Context AutoEncoder (CAE, — where we decoupled the Encoder, Predictor ("Regressor" in the paper), and Decoder — to Meta's I-JEPA/V-JEPA shifting entirely to latent prediction, and now this. It seems we are all validating the same core intuition: Understanding ≠ Pixel Reconstruction. 💡 Decoupling representation learning from high-frequency detail generation to build World Models in latent space is evidently the path forward. Glad to see our early intuition resonating with the latest SOTA. 🫡 #AI# #ComputerVision# #WorldModel# #JEPA# #CAE#

Meet Gemma 4 12B! A unified, encoder-free multimodal model designed to bring high-performance intelligence directly to your laptop, and released under an Apache 2.0 license. Bridging the gap between edge efficiency and advanced reasoning. Here is what’s new with Gemma 4 12B: 👇

DeepSeek V4.1 Gives Prefill and Decode Different Compute Paths Prompt tokens mostly traverse 20 layers; generated tokens traverse all 40. This Causal Encoder-Decoder design nearly halves long-input prefill while preserving autoregressive generation. Zhihu contributor 潜龙勿用, Changxin Ke(柯昌鑫), a graduate researcher at ICT, CAS, explains how its architecture and post-training were designed together. 1️⃣ A causal encoder, not T5 The 40-layer backbone is split into a 20-layer causal encoder and a 20-layer decoder. Both remain causal. The encoder processes the prompt and supplies the decoder’s global KV. Most prompt tokens avoid the decoder stack, while generated tokens still run through all 40 layers. That means 8B active parameters per prefill token versus 16B per decode token. 2️⃣ Most layers share global memory CSA2 uses three modes: Full creates global KV and an index; Reindex shares the KV but selects new positions; Reuse shares both. Only four layers create independent global KV, four reindex it, and 30 reuse both KV and the latest index. Each layer retains its own query and local SWA state. The first Full layer builds up to 16,384 candidates. Later layers search this pool for their Top-512. From 4K to 1M context, decode FLOPs per token rise only about 25%. 3️⃣ Serving approximations enter training Exact reconstruction of decoder SWA states would replay 2,560 prompt tokens. V4.1 replays only the final 128 encoder outputs and trains the model to tolerate the approximation. With FP4 global KV, cache falls to 890 bytes per token, persistent cache to roughly one eighth of V4-Flash, and prefill compute close to half. Bounded replay and constrained retrieval are not last-minute serving tricks. The model experiences them during post-training. 4️⃣ Post-training is an evolving Agent system Each RL task combines a problem, environment, and verifier. New trajectories can reveal shortcuts, broken environments, or verifier errors and send the task back for repair. V4.1 trains across multiple harnesses, merges checkpoints between RL runs, and finishes with on-policy distillation from more than 40 teachers. Raising reasoning effort from 25 to 100 increases output length about 2.5×, while average Pass@1 across eight benchmarks rises from 67.1% to 76.3%. ✅ The real design choice V4.1 aligns model structure, cache policy, retrieval limits, training environments, and inference around long-running Agents. The model learns under the same constraints the deployed system will actually impose. 🔗 Full analysis: #DeepSeek# #DeepSeekV41# #LLMArchitecture# #AIAgents# #LongContext# #AIInfra#

Looped transformers are a popular architecture topic right now. This new technical report extends the loop across tokens. Recurrent Looped Transformer (RLT) makes the decoder recurrent over every token, prompt and response included. A causal encoder builds global KV memory. For each new token, the decoder combines the token's encoder representation with its own final hidden state from the previous token and a sliding-window cache of recent activations. With a 48-layer decoder, the computation path after t tokens runs through 48t decoder blocks, while each token still executes a fixed number of blocks. Depth grows with the sequence and per-token cost stays the same. The same state transition is used for pretraining, SFT, sampling and RL replay, and nothing resets at the prompt-response boundary. RL replay rebuilds states under the current weights instead of reusing stale rollout states. The report is a design proposal. The author states that reasoning gains, hardware speedups and RL scaling are goals that have not been measured yet. Paper: Chat with Paper:

“SenseNova-U1.5: Towards Native Unified Visual Intelligence” Most multimodal models still use separate representations for seeing and generating images. SenseNova-U1.5 instead uses one 8B encoder-free, VAE-free model to understand, reason about, generate, and edit images directly in pixel space, including native 4K generation. It then trains specialist RL experts for aesthetics, text, infographics, and editing, and distills them back into one unified model. The result is a stronger case that visual understanding and generation can share the same native representation rather than being separate systems.