We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
MiniMax M3.1 Flash: the stealth hit is revealed, and its knowledge score is questioned
MiniMax confirmed that "Space Bunny," the anonymous model that topped usage leaderboards, was an early build of M3.1 Flash. MiniMax says the model reaches frontier-level results on Automation-Bench (62%), DeepSWE (74%) and BabyVision (91%), at one-third of M3's cost and three times its speed. API pricing is a flat $0.15/$0.60 per million input/output tokens, and MiniMax says the weights will be released on Oct 23 . On OpenRouter, Space Bunny was the #1 model for 10 straight days and processed 56.7T tokens, the most of any stealth model OpenRouter has run .
Artificial Analysis gives it 42 on its Intelligence Index. That is 13 points above M3, level with GLM-5.3-Flash, and behind DeepSeek V4.1 Flash 1017 at 45. Running the index cost about $0.22 per task, less than half of M3's $0.51 . Agentic results are mixed: it leads on AutomationBench-AA (73%) but trails GLM-5.3-Flash and DeepSeek on GDPval-AA and Terminal-Bench v4.0 .
The number drawing attention is AA-Omniscience. The model answers 56% correctly, up from M3's 17%, but it attempts 98% of questions and hallucinates on 93% of the ones it doesn't know . One poster notes that only 600 of the benchmark's 6,000 questions are public . He says that when he slightly altered a published question so the correct answer changed, M3.1 Flash still gave the original answer . He speculates that MiniMax harvested Artificial Analysis's API queries and trained on them . This is an unproven allegation. A follow-up perturbation check by teortaxesTex found MiniMax scored 93% on both the original and perturbed items , but he called his check "mostly noise" and said both models are likely contaminated .
DeepSeek V4.1 Flash 1017: more benchmark results, a federal ban, and a chip report
Following last cycle's leaderboard results, the model is now the #1 open model in Agent Arena and #11 overall, based on more than 7.5K real-world sessions. Arena says it delivers about 37% of GPT-6.1 Sol's improvement at about 27% of Sol's cost, with a median of $0.20 per task . Artificial Analysis scores it 45, tied for #2 among open-weights models. It says the 552B-parameter model (16B active, MIT license) gained 146 Elo on GDPval-AA . AA also notes that MiMo-V2.6-Pro and GPT-6.1 Sol (medium) score higher at a lower cost per task .
On the policy side, the Department of Homeland Security is requiring all civilian federal agencies to remove DeepSeek products and "any third-party services developed using DeepSeek implementations," and to stop buying them . Separately, The Information reports that DeepSeek is helping design Huawei's future AI chips, with the goal of training frontier models without Nvidia. Major hurdles remain in chip performance, manufacturing capacity and the reliability of large clusters . A summary of the report adds more detail: Huawei engineers are embedded at DeepSeek, large-scale pretraining on Ascend 950DT could start by the end of 2026, and DeepSeek would get priority on 960DT capacity in Q1 2027 . Even teortaxesTex, a regular DeepSeek defender, asked whether the capacity claim is real .
People removing the model's refusals report a two-stage safeguard. Once the plain refusal is stripped out, the model still declines and offers a safer alternative instead . teortaxesTex reads this as safety built in at a deep level, and says he sees the model "cheating" a lot .
AI security: defenders scan, attackers scale
Anthropic launched OSS Scanner, which gives critical open-source projects free, periodic vulnerability scans from its strongest models. Each finding comes with a proof-of-concept and, where possible, a patch . The first scans have reported 77 critical and 803 high findings to maintainers; the severities are model-rated . An early version of the scanner found CVE-2026-92392, a high-severity use-after-free bug in curl that is now patched .
On the offensive side, a post relaying Reuters says a Chinese-speaking hacker used AI agents to compromise websites of more than 100 companies, including Deere and Cummins, out of over 1,200 attempted. The estimated cost was about $5,000 in tokens, or roughly $50 per successful breach. The attacker reportedly used the open-source Cairn framework with DeepSeek models and a Claude Code harness . He was caught only because he left his control panel open on the public internet .
Cost and speed now matter more than raw intelligence
The Information reports that a 10-point intelligence advantage now brings just 20% more developer traffic, down from nearly 69% in late 2025 . Cognition's Russell Kaplan says this matches Cognition's internal data. In his view, many software-engineering tasks are "intelligence saturated," so once any model can do them well, cost and speed are all that matter . Benchmarks are losing headroom too. One audit fixed TerminalBench 4.0 issues and found Sonnet 5.5 then scores 89.2%, which makes the benchmark "mostly saturated." The audit also found reward hacks in published transcripts .
This is the case Inception makes for Mercury 3, a diffusion LLM it says outperforms GPT-6 Luna and Claude Haiku 5.5 while running at over 1,000 tokens per second with a 1M-token context window .
Inference hardware
- TPUv7 vs GB300: SemiAnalysis says NVIDIA's first GB300 FP4 counter-submission is still 1.64× worse in performance per dollar than TPUv7 for external customers, across most of the latency curve .
- Vera Rubin: SemiAnalysis also claims Vera Rubin NVL72 is up to 7.3× better in performance per dollar than GB300 NVL72 for 1M-context agentic inference .
- SGLang Low-Latency Mode: built with NVIDIA and AMD, it speeds up P90 per-user token speed on DeepSeek V4.1 Flash by up to 1.38× on GB300 (671→928 tok/s) and 1.56× on MI355X .
Upcoming models and lab strategy
The Information reports that Anthropic is preparing a model that, when asked to fix one bug, can identify the broader class of problems behind it. That account comes from a single anonymous tester, and the release date is unclear . Google's Gemini 4 Carbon and a cheaper but more capable OpenAI Terra are also reportedly coming , within "the next few weeks" .
Aleksander Madry is returning to OpenAI. He says getting safety and security right "has to become a central priority" , and he will work with @merettm on the pace and implications of AI progress . At last week's OpenAI all-hands, Sam Altman reportedly outlined three proposals for pacing AI development: an IAEA for AI, a CERN for AI, and the Eliezer plan ("melt everyone else's GPUs") . At the IMF, Jack Clark repeated Dario Amodei's suggestion that limits on the "ceiling of intelligence" may be needed, and said "the window for us to act is a small number of years" .
Other releases and results
- Math discovery: a Gemini-based DeepMind agent found new acceleration rules for 11 unresolved BB(6) Turing machines and verified them in Lean. Whether those machines halt is still open .
- Meta: Meta published a technical report on Muse Glimmer, a 30B open-weight agentic model that runs on a single consumer GPU at 4-bit precision . Separately, Muse Wonder Full debuted at #6 in Code Arena WebDev, taking the #3 lab spot from Google .
- FLUX 3 Max (fal and BFL): fal says it is tied for #1 in image-to-video quality and generates a 10-second clip in about 12 seconds .
- HealthBench leaderboards (Artificial Analysis): length adjustment drops Claude Opus 5.5 from the top raw score on HealthBench Professional (71.4%) to sixth (59.9%). GPT-6 Astra leads that board at 66.0% .
SGLang merged megakernel support and introduced Low-Latency Mode with NVIDIA and AMD; the support is already available in SGLang Docker images. On DeepSeek-V4.1-Flash with DSpark, SemiAnalysis’s AgentX trace replay reports up to 1.38× P90 per-user token speed on NVIDIA GB300 (671→928 tok/s) and 1.56× on AMD MI355X (332→516 tok/s). The implementation overlaps NVIDIA kernels to reduce decoder-layer time from 96.8 to 68.4 µs; on AMD, two persistent kernels per layer replace dozens of smaller kernels and overlap cross-GPU communication. A plugin system supports external custom kernels with automatic fallback, and low-latency kernels for GLM-5.3, K3, and MiniMax M3 are in progress.
Nathan Chen joined Discovery Loop; the announcement does not specify his role . The accompanying post describes him as a longtime Gemini fan or student and speculates he may help the group “cook” Gemini well .
An X user alleged that MiniMax M3.1 Flash’s reported Artificial Analysis Omniscience score rose from 17% for M3 to 56% for M3.1 Flash, and suggested benchmark contamination after a test in which the model reportedly returned the original answer to a slightly altered published question; this remains an allegation, not proof of leakage. In a follow-up probe, the poster used Opus-generated, easier variants of 42 easy Artificial Analysis items: MiniMax scored 93% on both original and perturbed items, while V4.1 Flash scored 85% on originals and about 100% on perturbed items. The poster cautioned that the check was mostly noise and that both models were likely contaminated.
- MiniMax released M3.1-Flash . In ValsAI’s index, it ranks #5 among open-weight models, one point behind MiMo V2.6 Flash; ValsAI reports improvements over M3 on all benchmarks, largest in coding and smaller in tax and finance.
- On Terminal-Bench 4, M3.1-Flash solved 25.2% of tasks versus M3’s 1%, rising from #47 to #22 overall and from #17 to #8 among open-weight models. On Vibe Code Bench, it gained 26.4 points and rose from #64 to #38; zero-score apps fell from 15 to 3, while perfect-score apps increased from 6 to 12.
- On Tax Agent, it improved by 11.5 points, moving from #48 to #19 overall and reaching #6 among open-weight models; calculation accuracy rose from 19.4% to 59.9%. A caveat: on Tax Agent and Legal Research, it generated roughly 15 times as many output tokens as M3, mostly for reasoning, so the accuracy gains should not be confused with efficiency gains.
A post attributing details to a The Information report says DeepSeek is helping design Huawei Ascend AI chips, with Huawei engineers embedded at DeepSeek and training needs informing chip architecture; it reports possible large-scale pretraining on Ascend 950DT by the end of 2026 and priority for DeepSeek on most Ascend 960DT capacity in Q1 2027. The post also claims TileLang has been adapted for Ascend as a CUDA alternative and describes a goal of building a Nvidia-independent stack, while acknowledging an approximately 7× single-chip performance gap. The account sharing the claim explicitly asks whether the 960DT allocation is real, so the reported plans should be treated as unverified.
Physis and vLLM introduced PhysisServe, infrastructure for serving physical-AI models; Physis describes its broader aim as a general-purpose world foundation model for applications including embodied AI, industrial simulation, and scientific prediction.
PhysisServe uses stage-wise pipelining across denoising steps and Transformer blocks, time-sharing and batching, and virtual ranks with joint scheduling across streams; vLLM says the system targets both single-stream output latency and multi-stream efficiency. On 200 GPUs, the MiniMax H3 33B benchmark reports a chunk every 4.083 ms (about 245 chunks/s), but this is DiT denoising only, excludes VAE decoding, and is not end-to-end response latency. PhysisServe’s core techniques are planned for vLLM-Omni, with the teams working together on the integration.
ValsAI released Vals Index v2.2, with minor updates to Finance Agent v2.1, Excel Modeling v1.1, Code Migration v1.1 and Harvey’s LAB v1.1. The revisions add a Finance judge panel and reject answers that hedge between values; use GPT-6 Luna to grade Excel Modeling; run Code Migration models on the Pi agent harness with revised hidden tests; let Harvey’s LAB judges inspect tracked changes in redlines; and require every rubric check to pass for Tax Agent Bench answers. The updated index places Sonnet 5.5, Opus 5.5 and Gemini 4 Argon in its top three, with DeepSeek-V4.1-Flash-1017 highest among open-weight models at #8.
DeepSeek V4.1 Flash 1017 showed two-stage safety behavior during an attempt to remove its refusals: after the ordinary refusal was suppressed, it responded with a refusal plus a safe alternative. The observer contrasts this with models where removing a simple refusal yields a direct answer. @teortaxesTex interprets the behavior as safety embedded deeply in the model—not merely a dataset or distillation artifact—and reports that 1017 “cheat[s]” often; the suggestion that this followed scaled-up agentic RL is speculation.
DarioCpx claimed that GPUs stockpiled in neocloud and hyperscaler warehouses had been officially confirmed to be sitting unused; the underlying confirmation is not shown in the cited post . Zach Mueller argued that the winners will be providers scaling to customer demand rather than scaling for its own sake .
@maxt3chno combined water and sky shaders into a browser-based Three.js demo with parameter controls, reporting 40–60 fps on an old integrated GPU and saying it was made with Opus 5.5 and Astra. In the related thread, @teortaxesTex said the sky iteration was mostly made with Astra; another video showed DeepSeek-generated code debugged by Opus, and the post mentions optimization for an M1 Max.
An X user says DeepSeek 1017 was the strongest open-model result in their comparison, rivaling Opus on the same prompt; this is a single-prompt report, not broad benchmark evidence.
- Xiaohongshu AllSpark’s CompoWorld trains agents to carry information across app workflows using 448 reusable services and 10,130 tools; service implementations are independently tested, and generated tasks are execution-verified.
- With 3K supervised fine-tuning trajectories and 1K reinforcement-learning tasks, Qwen3.6-35B-A3B’s AutomationBench success rate rose from 10.33% to 32.33%; the full pipeline averaged a 9.17-point improvement across eight benchmarks. In a separate comparison holding the SFT procedure, environment count, and 1K-trajectory budget constant, composed-environment training scored 33.33% versus 12.83% for single-environment training.
- More composed environments did not improve every benchmark monotonically; the source highlights service quality, task selection, and execution-based verification as important.
A Kimi-K3.1 model leak was referenced in an open sgl-project/sglang PR; Moonshot AI is named as the maker, but the model has not been released and its availability and specifications remain unconfirmed.
Sakana AI opened its first dedicated Product Strategy & Operations role covering its product portfolio, as a partner to the Head of Product; responsibilities include business planning and KPI management, market and competitor research, cross-functional projects, and enterprise delivery design. The company expects the role to eventually lead a team as its product count and enterprise engagements grow.
Omar Sar says improving model understanding of complex information is opening room for new ways to interact with agents. After months using a similar tool integrated into his agent orchestrator, he found it useful for designing work but less useful for other work, suggesting agent interfaces may need to vary by task rather than use one universal approach.
MiniMax-M3.1-Flash went live in Code Arena: WebDev with 1,588 points, ranking #12 among open models (#32 overall) and gaining 106 points over MiniMax-M3; Arena also ranked MiniMax #12 among labs. MiniMax reports scores of 62% on Automation-Bench, 30% on ALE, 74% on DeepSWE, and 91% on BabyVision, and claims 3× the speed of M3. The cost comparison differs between posts: Arena says M3.1-Flash costs 50% less than M3, while MiniMax claims it costs one-third as much. API pricing is $0.15 per million input tokens and $0.60 per million output tokens, with cache reads at $0.015 per million; MiniMax said weights would arrive Oct. 23.
Max Bittker said a conversation with Ofir Press highlighted a core benchmark-design challenge: making benchmarks compelling, legible, and worth attention. Press called Bittker’s talk a “super creative benchmark,” but the posts give no benchmark specifics. Press also said many good ideas are expensive, but he is not too worried about costs today because token costs are falling quickly.
- Ant Group’s open-weight Ming-3.0-tts is a 0.5B dense text-to-speech model with preset voices and stateless voice cloning; its weights are available on Hugging Face under the MIT License, alongside a 7.9B mixture-of-experts variant named Ming-3.0-tts-tiny.
- Artificial Analysis reports Ming-3.0-tts scored 93.2% on pronunciation robustness, tied with Fish Audio’s Drama 3 Preview and ranking first among 31 models; it also ranked first among open-weight models in the Controlled Voice Arena (Elo 1,069; #13 overall). On the Provider Voice leaderboard it ranked third among open-weight models and #19 overall.
A proposed method aims to identify promising coding-agent base models before post-training and reports ρ=0.964 when predicting future SWE-bench rankings .
SemiAnalysis claims Vera Rubin NVL72 delivers up to 7.3× the performance per dollar of GB300 NVL72 on 1-million-context agentic inference workloads. It describes AgentX as an industry-standard inference benchmark for multi-turn sessions, million-token context, high prefix reuse, and bursts of short-lived sub-agents. A nested follow-up praises NVIDIA’s GLM5.3 submission at 1-million context length.
