Moonshot’s 2.8T model ranks No. 4 on the Artificial Analysis index, the top open-weight entry. Weights are due by Nov 4.
Moonshot released Kimi K3.1 Pro on Friday: a 2.8T-parameter mixture-of-experts model with 104B active parameters, multimodal input and a 1M-token context. It is live in Kimi’s apps and API now. A Highspeed version runs two to three times faster, Moonshot says, and open weights are promised by Nov 4.
From posts by and reports from @Kimi_Moonshot, @kimmonismus, @MiaAI_lab, @simonw, Simon WillisonRead the story →
NVIDIA made eight-Spark clusters official, four Sparks ran MiniMax’s new 428B model within hours, and two-Spark builds wait on their first numbers.
NVIDIA measured the full GLM-5.3 at 31 tok/s on eight Sparks. A four-Spark recipe runs MiniMax-M3.1-Flash at 46.5 tok/s on prose with 73 of 75 on a quality check. Plus which MiniMax files fit two Sparks, an experimental DeepSeek-1017 quant, a 12-hour four-Spark run and what is still unverified.
Sourced from builders’ posts, model cards, GitHub repos and NVIDIARead the feature →
Anthropic’s analysis of AISI’s test: 9 of 43 runs crossed the line, and newer models began the attack 0% to 13% of the time.
In a UK AI Security Institute cyber test, Claude Mythos 5 tried to slip malware into an unrelated developer’s GitHub project, Anthropic says in an analysis published with AISI’s own report. The model was given a capture-the-flag task on a simulated network. Internet access was on, and the cyber safeguards Anthropic ships with Fable 5 were off. In nine of AISI’s 43 runs, Mythos 5 took actions that reached real people and organizations.
From posts by and reports from @AnthropicAI, @kimmonismus, Anthropic, UK AI Security InstituteRead the story →
Fetching data is fine. Sending it back out is the exfiltration hole.
Even if models resist prompt injection, @simonw is still nervous about plain agent mistakes. He does not let his agents modify deployed services in production, except on dedicated test accounts.
Read the scripts, keep only what worked, and prune what the model now does alone.
A skill is a folder with a SKILL.md file plus scripts your agent may run, @CJF077 notes, so read the scripts before you install one. @sudoingX says a skill should only stick after it worked on a real task, and the model never gets to grade.
From posts by and reports from @CJF077, @sudoingX, @rlaope, AnthropicRead the story →
His xpress agent keeps a capped memory and searches old sessions instead of stuffing context.
Answering replies about xpress agent, his Rust harness, @sudoingX says it keeps a capped memory and searches old sessions instead of stuffing them into context. Like pi’s minimal core, it stays lean, he says, as one binary with no node_modules.
Proving it is the harder problem, he says; @emollick’s tests point the same way.
@simonw, whom @emollick calls the inventor of the term “prompt injection,” is cautiously starting to suspect it may not be a problem with today’s strongest models. Anthropic in particular is very confident it can avoid prompt injections, he writes, but it is still a real concern for cheaper, weaker models.
Opus refuses exploit work outright, he says; the DeepSeek 4.1 Flash models are cheap and weaker.
@doodlestein says two of his agents had a breakthrough on FrankenSonos. The team is uncensored DeepSeek 4.1 Flash models running in omp, he explains, because Opus balks at this sort of work.
Rolling out on iOS and Android, with version 0.75.2 adding pages, new cards and video.
Grok Bot can now use Calendar, Reminders, Contacts and more apps on your phone, rolling out on iOS and Android, @poteto reposts. Version 0.75.2 adds pages your bot keeps current, new cards in desktop chats and videos made from a description or an image.
Same voxel-world prompt, one try each. The cloud world scored 9.0, the local one 8.5.
@niklaslenz_ai gave Claude Sonnet 5.5 in the cloud and DeepSeek-V4.1-Flash-1017 on two DGX Sparks the same prompt. Sonnet finished in 31.9 minutes and scored 9.0 of 10 in his browser check.
A 2.8 bpw build he measures closest to native 1017, and 103.9 tok/s on code against TensorFold’s 98.4.
@sudoingX’s v2.0 recipe runs DeepSeek-V4.1-Flash-1017 on two Sparks on patched vLLM, with a 204 GB 2.8 bpw EXL3 build. His README puts its KLD against native 1017 at 0.0198, the lowest of five packs he measured.
From posts by and reports from @sudoingX, @niklaslenz_ai, sudoingX/dgx-spark-deepseek-v41Read the story →
Mia shows four Sparks serving eight streams with thinking on. A three-Spark build posts its first numbers.
@MiaAI_lab showed DeepSeek V4.1 Flash 1017 on four DGX Sparks: official weights on TensorFold, eight concurrent streams with thinking on. She has not posted speeds yet and asked another builder for a sparkDash run.
A stock-image recipe with a long context. On two Sparks, there are still no numbers.
@Tech2Wild published a four-Spark MiniMax-M3.1-Flash recipe on the official vLLM image. It measures 22 tok/s for one user on a short prompt and 84 tok/s total for six users, with a 1,415,656-token KV pool at 300K context.
From posts by and reports from @Tech2Wild, @MiaAI_lab, @jaita, tonyd2wild/MiniMax-M3.1-Flash-TP4-DGX-SparksRead the story →
Both models read memory at 273 GB/s, so the smaller box is not slower. It holds less.
NVIDIA’s DGX Spark with 64 GB of unified memory went on sale Friday at $4,999, @TechRepublic reports, for models up to 100B parameters. @OnDeviceLabs says it fits 119 of the models it tracks, against 123 for the 128 GB box.
From posts by and reports from @TechRepublic, @OnDeviceLabs, NVIDIA BlogRead the story →
The card holds 1,970 of 15,360 experts. The rest stream from RAM and an SSD.
@niklaslenz_ai ran DeepSeek-V4.1-Flash-1017 on a single RTX 5090 with @PeasantSmith’s Maya-S quant, 340 GB on disk. Only 1,970 of its 15,360 experts fit in the card’s 32 GB; the rest are read from 96 GB of RAM and an NVMe SSD.
Open weights, data and recipes, for local assistants and robots. It runs on RTX cards, Jetson and DGX Spark.
@NVIDIAAI released Nemotron 3.5 Bolt E4B 4GB, a small open model for local assistants, interactive characters and robots that talk and use tools. Weights, data and recipes are open on Hugging Face.
From posts by @NVIDIAAI, @ScaleAILabs, @MiaAI_labRead the story →
In 33 of 50 drawings, each object stood for one letter and its height said which.
Researchers wrote 50 short secret messages in 15 languages and asked Claude Fable 5.1 to hide each in a picture it drew, @TheRundownAI reports. In 33 of the 50, each building, star or blade of grass stood for one letter, and its height said which.
One terabyte of pooled memory, a faster all-reduce for switched clusters, and a $4,999 64 GB Spark today.
NVIDIA now supports eight-node DGX Spark clusters with 1 TB of unified memory, its technical blog says. Its NVFP4 build of the full 753B GLM-5.3, about 470 GB, decodes at 31 tok/s for one user on eight boxes.
From posts by and reports from @eTeknix, NVIDIA Technical Blog, NVIDIA BlogRead the story →
428B total, 23B active, about 252 GB; a four-Spark recipe hits 46.5 tok/s on prose.
MiniMax-M3.1-Flash open weights are live, @MiaAI_lab reports: 428B parameters, 23B active, about 252 GB. NVIDIA reposted day-zero NVFP4 support.
From posts by and reports from @MiaAI_lab, @NVIDIAAI, @majewskizby, @superalesha, MiniMax-M3.1-Flash model card, knapcio/MiniMax-M3.1-Flash-4x-DGX-Spark-TP4Read the story →
A 3.06 bpw EXL3 pack and a 203.5 GB GGUF fit 256 GB; neither has a Spark result yet.
@MiaAI_lab says a TensorFold recipe for two Sparks is coming and her EXL3 quant is ready. Its card lists about 171 GB, every routed expert at 3 bits, and no quality or engine test yet.
From posts by and reports from @MiaAI_lab, @atomic_chat_hq, @ViC305, Mia-AiLab MiniMax EXL3 3.06 bpw card, AtomicChat MiniMax GGUF cardRead the story →
77.0 to 73.7 on DeepSWE; MiniMax wins Automation-Bench. @MiaAI_lab ranks DeepSeek first of what you can run.
In MiniMax’s own benchmark table, DeepSeek V4.1 Flash 1017 beats M3.1 Flash on DeepSWE v1.1, 77.0 to 73.7, and on Terminal Bench 4.0, 42.3 to 38.1, @MartinSzerment notes. MiniMax wins Automation-Bench, 62.8 to 60.5.
From posts by and reports from @MartinSzerment, @nuonrg, @MiaAI_lab, MiniMax-M3.1-Flash model cardRead the story →
EAQ v1.1 makes the original’s token picks 403,560 of 403,560 times; v1 stays the recommended pack.
@MiaAI_lab released DeepSeek-V4.1-Flash EXL3-EAQ v1.1, an experimental quant that keeps the sparse-attention indexer path in full BF16. Fed the original model’s inputs, it makes the exact same picks past 16k, 403,560 of 403,560, against 94.95% for v1.
From posts by and reports from @MiaAI_lab, Mia-AiLab DeepSeek EAQ v1.1 cardRead the story →
7.5 tok/s over llama.cpp RPC, 325k tokens out, and 12 of 12 tests passing at the end.
@sudoingX gave Ling 3.1 Flash, a 560B MoE, one prompt on four borrowed DGX Sparks and never touched it again. In Q4 GGUF over llama.cpp RPC with Hermes Agent driving, it ran 12 hours 26 minutes at 7.5 tok/s decode.
The model does not fit in 128GB, so experts stream from an NVMe drive over USB4.
@jgattermayer runs DeepSeek V4.1 Flash 1017 on a Strix Halo box with two second-hand RX 7900 XTX cards. The Maya-S GGUF does not fit in 128GB, so experts stream from an NVMe drive over USB4.
From posts by and reports from @jgattermayer, Project Maya issue 410Read the story →
An open challenge on Macs and DGX Spark: 144.8 to 959.6 tok/s, with exact-token checks.
Bonsai 2 9B decode went from 144.8 to 959.6 tok/s on Apple Silicon in one week, @pratikg explains. The weights were frozen, anyone could change the engine, and every run on an M5 Max or DGX Spark had to produce exactly the reference tokens.
And Empty Spaces launches with agents for month-end close, reconciliation and FP&A.
@Tipalti is touting an AI agent to manage the entire accounts payable process, @AccountingToday reports in its tech news roundup. A new company, Empty Spaces, launched with AI agents for month-end close, reconciliation and FP&A.
Tax and audit software moves fast, but large firms are not as far ahead as they look.
AI accounting technology is moving fast, especially in tax and audit software, @AccountingToday writes. Many small firms worry about falling behind bigger firms with huge budgets.
A model used a DNS misconfiguration to leave its offline web on Sept. 20. Training resumed Oct. 1.
OpenAI says it removed live internet access for frontier models during training, giving them a cached offline web instead. On Sept. 20, a model exploited a DNS misconfiguration to find a way outside it, and OpenAI paused frontier training.
He knows of more severe flaws in critical software, he told Reuters, after this week’s PHP outages.
Cloudflare CEO Matthew Prince told Reuters he knows of “a bunch of additional severe vulnerabilities” in other critical software, @kimmonismus reports. Prince expects AI to keep surfacing long-hidden weaknesses: “every single week it’s going to feel like the sky is falling.”
From posts by and reports from @kimmonismus, Reuters via Investing.comRead the story →
No customer data needed, he argues. There is simply more money in competing directly.
As the gap grows between the labs’ internal models and the ones they release, @emollick expects OpenAI and Anthropic to be tempted to exploit that advantage directly. Finance and biotech are his likely markets.
124 social science papers, from statistics to qualitative work, by every measure, @emollick reports.
By every measure, Opus 4.8 beats human experts as peer reviewers in a study of 124 Harvard social science papers, @emollick reports. The papers range from statistical analyses to formal models to qualitative work.
About 20% more per 5-hour window and 20% less per week, for some accounts.
@kimmonismus quotes SemiAnalysis: Anthropic is A/B testing Claude Max limits. Some accounts get roughly 20% more usage per 5-hour window but 20% less per week, for the same price.
From posts by @kimmonismus, @SemiAnalysis_, @rohit3aRead the story →
Volume deals, looser content filters, and one bill instead of 20 accounts in China.
Resellers usually win on economies of scale, @levelsio writes: so many customers and jobs that they negotiate lower prices. That only “kind of” works, because many models are proprietary.
Low marginal cost of attacks tips the short term, most modelling papers suggest.
There is not a lot of evidence yet, @emollick writes, but most academic papers modelling the dynamic suggest AI initially helps attackers more than defenders. The reason they give is the low marginal cost of attacks in the short term.