llama.cpp shipped DSpark for DeepSeek V4 Flash, and almost everyone called it the wrong name
llama.cpp release b10228 merged speculative-decoding support for DeepSeek V4 Flash, but the new 0731 checkpoint embeds DSpark and ships no MTP at all, so the widely repeated "MTP support landed" advice points users at a head their model does not contain.
Quantizing V4 Flash's KV cache in llama.cpp changes which tokens it picks
A community experiment found that llama.cpp's generic 8-bit KV cache leaves DeepSeek V4 Flash's average perplexity almost unchanged while altering which tokens make the model's shortlist about one time in eight, because the model's attention makes discrete block-retrieval decisions that a small numerical error can flip.
A 284-billion-parameter model with a 3-gigabyte working set, and a 96-gigabyte disk bill
An open-source engine called Mference runs DeepSeek V4 Flash on a 24 GB Mac with an effective memory footprint of about 3 gigabytes by streaming each token's experts off the SSD, but the checkpoint still occupies 90 to 98 gigabytes of disk and the test ran at a 4,000-token context.
DeepSeek's low effort setting writes more than its high setting, because the dial is just a prompt
A reproduction across both a local copy and DeepSeek's hosted API found V4 Flash consuming substantially more tokens on its low reasoning-effort setting than on high, and the released encoder explains why: low injects no instruction at all while high prepends a paragraph demanding exhaustive deliberation.
An attacker's own AI agent exposed his entire operation to researchers
Palo Alto Networks' Unit 42 reconstructed an autonomous attack campaign from the operator's own session logs after his AI agent accidentally started a public file server from its home directory, revealing an open-source agent harness driving a hosted DeepSeek API through a Telegram channel.
ByteDance's Seedance 2.5 generates a 30-second single take, and still cannot promise a face across a cut
ByteDance Seed announced Seedance 2.5 on 31 July, a joint audio-video model producing single clips of up to 30 seconds extendable twice, roughly doubling its predecessor's maximum take length while documenting no persistent character identity across separate shots.
DistillAlign explains why fast video models get prettier and more repetitive at the same time
A new preprint shows that the standard technique for speeding up video generators optimises a mathematical objective that deliberately concentrates on safe, common outputs, and that adding a diversity-preserving loss back during that stage more than doubled the paper's coverage measure without giving up the quality gain.
NeurIPS rebuttal week ended with authors, reviewers and chairs all reporting the same silence
The NeurIPS 2026 author discussion window closed on 3 August with multiple independent reports of reviewers and area chairs never responding to rebuttals, in a cycle the conference redesigned specifically to make that discussion consequential.
Beijing says U.S. firms distilled Chinese models, and names none of them
China's Ministry of Commerce said in a written statement on 27 July that many U.S. AI companies had distilled Chinese models during research and training, identifying no company, no model and no evidence, mirroring a U.S. accusation five days earlier that named two companies but published no logs either.
Kimi K3 runs in 8 gigabytes of RAM, at 33 seconds per token
A hand-written C engine generates text with Moonshot's 2.8-trillion-parameter Kimi K3 using a peak of 8.24 gigabytes of RAM and no GPU, by reading the model's four-bit experts directly off disk, at a rate of roughly one token every 33 seconds.
The "2x GB200 bandwidth" Chinese chip claim is a 2027 projection, and the arithmetic gives 1.67x
A widely shared claim that a Chinese accelerator delivers twice the memory bandwidth of NVIDIA's GB200 traces to a roadmap part expected in early 2027, compared 64-at-a-time against a full NVIDIA rack, and the published numbers work out to 1.67 times at rack level while the single chip lands below a shipping GB200.
Four agent-memory papers landed in a week, and none tested what happens when an attacker controls the writes
Four papers published within days define an AI agent's memory as four incompatible things - a pretrained module, a rewritten lesson, a folder of files, and a reliability ledger - and three of them introduce writable state that determines future behaviour without evaluating an adversary who controls what gets written.