DeepSeek V4 Flash 0731 gets ARC Prize-verified benchmark scores
ARC Prize independently verified DeepSeek's V4 Flash 0731 on its ARC-AGI reasoning benchmarks: 89.0% on ARC-AGI-1 and 61.4% on ARC-AGI-2 at max reasoning effort, for $0.02 and $0.04 per task respectively. Lower-effort settings trade a few points of accuracy for roughly half the cost (84.0%/46.0% at low effort). It's the first independently verified scorecard for the model since its API/weights launch a week ago, and the cost-per-task framing is the notable bit — these numbers land well below what comparably-scoring proprietary models have charged on the same leaderboard.
Sources & depth
SGLang v0.5.17 ships day-0 Kimi K3 support
582 PRs from 194 contributors. The headline is day-0 serving support for Kimi K3 — a 2.8T-parameter multimodal MoE (896 experts, top-16, 1M-token context, 69 KDA linear-attention layers interleaved with 24 MLA layers) — including speculative decoding, chunked-prefill, and prefix caching tuned for its architecture, verified running on both NVIDIA GB300 and AMD MI35x. Beyond that: MoE prefill throughput up to 1.92x via a new dispatch scheme, engine restart time cut from 6+ minutes via weight caching, and a Rust-based frontend replacing parts of the Python request path for lower latency.
Sources & depth
- github.comv0.5.172026-08-08 07:13 IST
OpenAI says it can't rule out "critical" cyber capability in its next model
OpenAI disclosed that internal evaluations of Astra, an unreleased upcoming model, showed strong enough agentic-coding and cybersecurity performance that the company "cannot rule out critical cyber capabilities" under its own Preparedness Framework — the framework's highest defined tier, reached if a model can find and weaponize zero-days across many hardened real-world systems without human help, or plan and execute a novel end-to-end attack from just a high-level goal. Every prior OpenAI model, including GPT-5.6 Sol, topped out at the lower "High" tier. In response, OpenAI says it's paused internal work on Astra that doesn't yet meet stricter security requirements (isolated test environments, restricted network/tool access, weight encryption, chain-of-thought monitoring that can interrupt risky agent actions) and will bring in outside government and safety-org testers. Astra was not involved in the OpenAI/Hugging Face security incident from earlier this year. This is a disclosure about internal evaluation results and process, not a model launch — no release date, benchmark table, or capability demo accompanies it.
Sources & depth
xAI's Imagine Image 2.0 ranks #2 in the world on image generation and editing
xAI shipped Imagine Image 2.0 as the new "Quality Mode" on grok.com/imagine and its iOS/Android apps, built around precise editing rather than one-shot generation: a magic-wand tool edits only the region you point at, segmentation selects precise areas, background removal exports a transparent-background cutout, and multi-reference editing takes up to 5 input images in one generation. It also plans typography and layout for dense multi-part visuals and supports smart resize to any aspect ratio from one source image. By xAI's own citation of the Arena Text-to-Image and Image Edit leaderboards (Aug 7 snapshot), it ranks second in the world on both boards, behind OpenAI's gpt-image-2 and ahead of Meta's Muse Image, Google's Gemini image models, and ByteDance's Seedream 5.0 Pro — third-party leaderboard numbers, not xAI's own benchmark claims.
Sources & depth
TutorMoments: why AI tutors over-help, and how to fix it
AllenAI built an evaluation framework from 462 real one-on-one math tutoring transcripts (grades 2-7), with teachers annotating 1,500+ moments where a tutor has to choose between giving a student the answer and pushing them to work it out. Every model tested, given only a generic "tutor well" instruction, defaulted to over-explaining instead of holding back — a real failure mode for anyone deploying LLMs as tutors, not a benchmark curiosity. Spelling out the scaffold-vs-rigor tradeoff explicitly in the prompt measurably improved judgment, though no model reached human-tutor consistency.
Sources & depth