The base model becomes a research surface

X

Base models reopen the training surface

• ERNIE 5.5 Flash releases its pre-training base Baidu describes a 280B-parameter MoE with 14B active, 33.5T pre-training tokens and 128K training context, leaving instruction tuning and RL to downstream builders. @Baidu_ERNIE ↗
• vLLM adds day-zero ERNIE serving The project advertises FP8 and BF16, expert parallelism, KV offload and disaggregated prefill, plus ECHO speculative drafting; reported coding scores were not independently reproduced. @vllm_project ↗
• Baseten joins Nvidia’s open-model coalition Baseten says it will contribute RL environments, post-training infrastructure and optimized inference to the Nemotron effort; governance and artifact licensing were not stated in the captured post. @baseten ↗

Evaluation becomes part of the release

• A Kimi benchmark result draws contamination scrutiny A researcher argues K3.1’s Terminal-Bench 4.0 score may benefit from hidden task hints seen during training, citing task perturbations rather than claiming a definitive audit. @JJitsev ↗
• Tencent’s Hy4 posts a strong agentic-coding profile Artificial Analysis places it second among China-based labs on its index and reports 31% on FrontierCode 2.0; all scores and pricing remain evaluator-reported. @ArtificialAnlys ↗

Primary and business sources

Runtime correctness moves into the test harness

• llama.cpp removes accidental ambiguity from MoE tests Build b12128 redraws logits until routing choices have a clear numerical margin, preventing backend rounding differences from flipping nearly tied experts and creating CI flakes. github-actions[bot] ↗
• Qwen2 NVFP4 scales reach the feed-forward path Build b12126 fixes a crash where converted models loaded scales but never applied them in the FFN, while attention already handled them. danbev ↗
• Metal keeps fusable groups inside one command buffer Build b12127 moves graph-splitting boundaries to group ends, so fusion no longer changes with node count or breaks across command buffers. ServeurpersoCom ↗
• Adreno X2 enables more GEMV paths by default Build b12125 turns on float32 and Q5_K multi-column matrix-vector kernels for X2E devices; the release note supplies no end-to-end benchmark. github-actions[bot] ↗
• OpenCL adds DeepSeek V4 hybrid-compression operators Build b12123 introduces DSV4_HC_PRE and DSV4_HC_POST and refines the post operator, extending support without publishing measured gains. github-actions[bot] ↗

Agent infrastructure attracts infrastructure money

• Dropbox plans a $350 million Tigris acquisition The announced purchase would put S3-compatible object storage behind DBX Agent Cloud; reporting calls it Dropbox’s largest acquisition, with closing details not retained. Judy Rider ↗ · Lauren Forristal ↗

Reddit

Local agents are tested as resilience systems

• Six coding agents face a mid-task API outage Across three runs each, one author says only Atomic Agent automatically switched local and finished; configured fallbacks improved some others, while Codex remained reconnecting. Ok-Brush8377 ↗
• Local inference still makes outbound connections A firewall test logged 494 attempts in 38 minutes from catalogs, package installs, telemetry and system services; a hardened mode later reported zero attempts. ldrx ↗
• A long-prompt workload reverses a short-benchmark intuition On one M1 Max extraction job, llama.cpp led sites per hour because prefill and batching outweighed BaseRT’s faster decode; hardware and task choices are author-controlled. Objective_Stick_2513 ↗
• Intel Arc patches trade prefill against decode One same-box comparison reports arcint leading stock SYCL prefill on A770 and B60, while SYCL slightly led decode; the author published settings and caveats. marfrit ↗

Users turn model fit into inspectable tooling

• iQuantize makes compression steps visible The Apache-2.0 workspace supports GGUF, MLX and ONNX, compares quantized outputs with originals and records the process in a publishable model card. Zipidyzip ↗
• EmbeddingGemma 2 powers offline phone search A new Android app indexes video, audio, PDFs, photos and text locally across more than 100 languages, with NPU paths for several mobile chip families. bj_davey ↗
• A Git sandbox puts invariants ahead of generated code Its builder used DeepSeek through OpenCode but designed repository state and tests first, keeping commits, branches, index and merge conflicts derived from one engine. No_Captain4899 ↗

Discord

Agent state becomes an operating interface

• A Hermes workflow texts its owner when a change is ready The linked Tailscale page exposes what was built, unanswered questions, tests and data checks, turning remote approval into a private review surface. Justin ↗
• Unsloth members point to MemT for model memory The thread links arXiv 2610.26825 and describes a Qwen3.6-based system that goes beyond plain-text storage; implementation details were not verified from Discord alone. resident exl3 shill (use exl3) ↗ · resident exl3 shill (use exl3) ↗ · Unknown rendered member ↗

Companies & consequential moves