A focused feed of models, agents, research and open-source releases for people building with AI.
OpenAI unveiled Pando , a Git-compatible backend powering its entire monorepo. Median push latency dropped ~80%, system now serves nearly a petabyte of Git data daily. Engineers and Codex agents keep using standard Git commands with no workflow changes.
Prefrontal AI and ETH Zurich introduce Latent Sparse Attention (LSA), targeting the indexer bottleneck at 10M-token context At ultra-long context, the key-selection indexer consumes about 98% of a sparse attention layer's runtime LSA clusters keys hierarchically in a learned latent space so queries
SGLang HiCache now integrates with context parallelism, pipeline parallelism, speculative decoding, and prefill-decode disaggregation serving iFLYTEK Astron production: cache hit rate jumped from 30% to 95%, median TTFT under 3 seconds on GLM-5.1/5.2 CP sharding raised GPU token hit rate ~40 points
Anthropic launched message threads in public beta on the Messages API. Server caches conversation between turns, so clients send only new messages plus the previous response ID. Median time to first token drops about 35% on 500K to 1M token requests inside Claude Code.
Next-gen Meta Andromeda hits 1B+ ads/sec per server, a 100x jump over 2024. Retrieval MoE architecture reaches ranking-level quality at ~300x fewer inference FLOPs. Fused GPU kernel keeps tensors in registers and shared memory, hitting ~85% of peak HBM bandwidth.
Bionic now has a built-in terminal powered by libghostty, Ghostty's embeddable emulator core. Toggle with Cmd+J on Mac or Ctrl+J on Windows and Linux. Available in public Bionic builds on Mac, Windows, and Linux, free.
AA-AgentPerf adds Throughput per $/hr , a value-for-money metric for inference systems. Pricing comes from Silicon Data daily rental indices for H100, H200, B200 and B300. NVIDIA B300 leads overall; B200 matches it at faster output speeds on some models.
Instagram Live ingest on iOS and Android now runs on Media over QUIC , replacing Meta's in-house RUSH protocol. MoQ is an emerging IETF standard co-developed with Google, Akamai, Cisco, Cloudflare, and academia.
AutoTrust ships a lossless MTP speculative-decoding recipe for Qwen3.8-27B-NVFP4 , taking DGX Spark from 12.5 to 48.2 tok/s. Weights are bit-exact identical to NVIDIA's upstream NVFP4 release; only the drafter changes.
Zyphra introduces ZHA , a tree-structured sparse attention that is asymptotically subquadratic. Queries descend a tree of compressed KV summaries, skipping entire branches instead of scoring every key. Projected speedups: 35x over dense at 10M tokens, up to 417x at 1B tokens.
Perspective paper reframes subsurface ML: deliverable is calibrated uncertainty and auditable risk envelopes, not point forecasts. Identifies four bottlenecks: scarce labels, uncorrelated uncertainty, poor scale bridging, weak QA for regulators.
KAIST's L3G-World generates explorable 3D worlds at 9 FPS from a single image. Each 3D Gaussian stores a VAE latent, rendering directly into the diffusion model's conditioning space. Memory grows with scene extent, not stream length, keeping GPU use flat during long exploration.
Alibaba's Wan team released Wan-Animate-2 , an end-to-end character animation model with native ComfyUI support . No pose extraction or skeleton preprocessing: the driving video goes straight into a redesigned Diffusion Transformer.
Zed announced DeltaDB , a Git-compatible version control system for agent context and conversations. Every human or agent edit becomes a delta in a causal DAG, linking code to the conversation that produced it.
A research team proposes using thin-film lithium niobate modulators as analog replacements for Softmax and Sigmoid inside transformer attention.
Meta built a scheduler that scavenges idle GPU and CPU slivers across its inference fleet to improve ads ranking without buying new hardware.
SYSTRAN's English-only base checkpoint repackages OpenAI Whisper for CTranslate2, delivering roughly 4x faster transcription at the same accuracy on modest hardware.
PGX runs thousands of isolated Postgres databases inside a single Rust process, built for throwaway agent sandboxes, CI jobs, and PR previews.
vLLM matches DeepSeek's own inference engine on V4.1-Flash and scales to 954K tokens/s per GPU on a GB300 NVL72 rack.
Vals AI's new Inference Provider Index shows the same open-weight model can cost 2.3x more and run 5x slower depending on where you host it.
Cohere's Model Vault now runs on hardware-isolated NVIDIA H100 enclaves on Azure, letting enterprises verify their data stays private even from Cohere.
Humans& rebuilt the NVFP4 decode kernel on Blackwell by keeping weights at 4 bits while promoting activations to BF16, gaining both speed and accuracy.
You.com's new Inference Router pairs open-weight models like Kimi K3 and GLM 5.3 with a 99.9% uptime SLA, zero data retention, and built-in web search.
A new two-stage training recipe lets looped language models reuse a single KV-cache across every recurrence, cutting HBM use up to 66%.
No stories match these filters.
