Serving a 2.8T-parameter open model requires 8x B300 and virtual memory from the 1960s; licensing the pipeline that built it cost Nvidia $7B. Two faces of the same tension.
Moonshot's Kimi K3, 2.8T-parameter MoE, just 16 of 896 experts active per token, isn't easy to serve: it wants an 8x B300 footing. What makes a 3T-class open model hostable at all is MXFP4 weights quantized from SFT onward and a KDA rework baked into vLLM. A user's plain vLLM run on 8x B300 lands near 92 tok/s decode at ~$190 per million tokens; Modal's custom speculator pushes interactive serving to 460 tok/s.
A new reasoning model, "Ox Alpha," is now live on OpenRouter as a free stealth offering from an anonymous third-party provider. It's aimed at long-horizon software engineering, sustained agentic work, and production loads, with a 1M-token context and support for text-plus-visual workflows. OpenRouter routes it but stresses it isn't the developer, prompts stay with the provider and aren't used for training. The anonymity makes it both intriguing and hard to benchmark against known models.
Naive inference reserves a full max-context KV slot per user, so a short request can hoard thousands of tokens it never uses, both internally and externally fragmented. PagedAttention lifts virtual memory straight from 1960s OS design, paging cache into fixed blocks so kernels still run on non-contiguous memory. The payoff: 2-4× more users served per GPU without bigger hardware.
DiG-bench is a benchmark of scientific discovery built as 70 text-based games, each with transformation rules an agent must uncover purely through interaction and experimentation. Twenty-one games are public; the rest stay private for secure evaluation, and every game has been beaten by at least one human on a first attempt. The point: these tests isolate a prerequisite for creativity, autonomously finding useful, undocumented truths about genuinely novel setups. Early results suggest some frontier models already approach human performance.
Hugging Face found that several of the best open-source speech-recognition models "benchmax": they echo VoxPopuli and LibriSpeech reference transcripts even when the audio contradicts them. In one probe, six of 11 models reproduced a benchmark's erroneous transcript instead of what was actually said, behavior that vanished on fresh recordings. The post introduces three tests for quantifying this kind of benchmark gaming, alongside the held-out sets HF already added to its ASR leaderboards.
IBM Research's ALTK-Evolve distills reusable guidelines from an agent's own trajectories and injects them at inference, with no weight updates or human labeling. Tested across eight models, the dose matters: capable models want the full set, weaker ones do best with a compact core plus per-task retrieval, and saturated models gain nothing. Notably, gpt-oss-120b jumped +16.1pp on task completion for just +5% tokens. The lesson: agentic memory is a calibrated dosage, not a default feature.
Standard PCA's scree plot lies about noisy, mixed-type tabular data, shattering nonlinear dependencies into spurious orthogonal dimensions that inflate the true rank. Entropic Scree replaces variance with normalized mutual information to collapse that to the real generative rank, and even sizes autoencoder bottlenecks from it. The standalone R function is live with CRAN and PyPI packages on the way.
Nvidia is handing Poolside $6B for a license to Model Factory, its open-weights-building pipeline, plus $1B in equity at a $12B valuation, and pulling 100+ staff onto Nemotron while the founders stay put. It's the third run of the same playbook after Groq and Enfabrica: absorb the mechanism of AI production under a licensing deal that sidesteps antitrust review. The catch, Poolside's Laguna models still trail DeepSeek and Claude on coding benchmarks, so Nvidia is buying leverage, not the frontier.
Benchmarking Inco AI's DFlash 2 block-diffusion drafter in a llama.cpp PR build on Qwen3.8-27B shows a 2.26× speedup on real coding prompts, 4.68× with an n-gram drafter, and up to 8× in specific cases. Speculative decoding leaves output provably unchanged, so users get near-doubled throughput without retraining or touching the base model. It's Apache 2.0 and open on GitHub.
Loyal Bear strips the invisible SynthID watermarks OpenAI and Gemini now bake into every AI-edited image, doing minimal damage to the picture itself. The author keeps its methods secret to dodge countermeasures, bans corporations outright, and rigs a tripwire to catch anyone snooping. Runs on any Python 3.10 machine with 8GB RAM, and text-SynthIDs could be next.
You’re caught up.
That was all 10 stories in the 23 August issue.