Particle-1 is now available in OpenClaw. @particle_api says that compared with DeepSeek V4.1 Flash, its output tokens are reduced by about 45%, with comparable benchmark accuracy; standard inference does not retain any data. Users can try it by upgrading to OpenClaw v2026.10.10.
Particle-1 is now available in OpenClaw. @particle api says that compared with DeepSeek V4.1 Flash, its output tokens are reduced by about 45%, with comparable benchmark accuracy; standard inference does not retain any data. Users can try it by upgrading to OpenClaw v2026.10.10.
The MIT CSAIL team proposes a weight-space programming method that lets GPT-6 Astra directly read and modify neural network weights without training or computing gradients. On Qwen3.5-0.8B, Astra modifies only about 0.016% of the weights, raising MATH-500 from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and IFEval from 21.0% to 62.3%, mainly because it fixed the model's failure of getting stuck in reasoning loops. The research was led by doctoral students Rikiya Takehi and Jayden Teoh, with participation from Omar Khattab and others, and the results were published on the research blog "Sparks of Weight-Space Programming"; additional experiments on Gemma 3 270M yielded improvements of no more than about 1 percentage point, and the team says more independent replications are still needed.
The MIT CSAIL team proposes a weight-space programming method that lets GPT-6 Astra directly read and modify neural network weights without training or computing gradients. On Qwen3.5-0.8B, Astra modifies only about 0.016% of the weights, raising MATH-500 from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and IFEval from 21.0% to 62.3%, mainly because it fixed the model's failure of getting stuck in reasoning loops. The research was led by doctoral students Rikiya Takehi and Jayden Teoh, with partic
In its ICML 2026 blog, LG AI Research points out that tabular foundation models such as TabPFN suffer from label shift: among 253 OpenML datasets, about 85% (214) are class-imbalanced, and in extreme cases 98.3% of test samples are predicted as the majority class, with minority-class recall nearly collapsing. The team proposes DistPFN, which at inference time multiplies predicted probabilities by a correction factor based on relative training class proportions, without changing the architecture, retraining, or estimating the test label distribution. ICML 2026 received a total of 23,918 submissions and accepted 6,352, for an acceptance rate of 26.6%.
In its ICML 2026 blog, LG AI Research points out that tabular foundation models such as TabPFN suffer from label shift: among 253 OpenML datasets, about 85% (214) are class-imbalanced, and in extreme cases 98.3% of test samples are predicted as the majority class, with minority-class recall nearly collapsing. The team proposes DistPFN, which at inference time multiplies predicted probabilities by a correction factor based on relative training class proportions, without changing the architecture,
At the end of October, Electric Power Golden Warehouse released the KES AI database, incorporating semantic construction, computational reasoning, and native data security control into the database kernel, emphasizing that AI applications can be supported without data leaving the database. Alibaba Cloud PolarDB launched capabilities such as the AI data lakehouse (Lakebase) in January this year and launched PolarDB Agentic Database in July, while OceanBase packaged its AI direction into the OceanBase AI Database product portfolio in June. IDC data shows that the market size of China's AI data infrastructure in 2025 is approximately 14.9 billion US dollars, and will reach 73.8 billion US dollars by 2030.
At the end of October, Electric Power Golden Warehouse released the KES AI database, incorporating semantic construction, computational reasoning, and native data security control into the database kernel, emphasizing that AI applications can be supported without data leaving the database. Alibaba Cloud PolarDB launched capabilities such as the AI data lakehouse (Lakebase) in January this year and launched PolarDB Agentic Database in July, while OceanBase packaged its AI direction into the Ocean
Reports suggest that DeepSeek's next-generation large model V5 will be pretrained on Huawei Ascend chips; previously, Ascend was mainly used for inference tasks on DeepSeek models. On September 30, DeepSeek open-sourced the TileLang compilation tool for the Ascend platform, computing libraries such as DeepGEMM, and the DeepEP communication library, corresponding one-to-one with NVIDIA platform components. The article also mentions that teams including Shenzhen Hetao College completed full-parameter post-training of the 1.6-trillion-parameter DeepSeek-V4-Pro on an Ascend 910C cluster, but this did not involve foundational pretraining.
Reports suggest that DeepSeek's next-generation large model V5 will be pretrained on Huawei Ascend chips; previously, Ascend was mainly used for inference tasks on DeepSeek models. On September 30, DeepSeek open-sourced the TileLang compilation tool for the Ascend platform, computing libraries such as DeepGEMM, and the DeepEP communication library, corresponding one-to-one with NVIDIA platform components. The article also mentions that teams including Shenzhen Hetao College completed full-parame
ARC Prize announced DeepSeek V4.1 Flash 1017's results on the ARC-AGI series of benchmarks: 5.81% on ARC-AGI-3, 78.1% on ARC-AGI-2, and 94.5% on ARC-AGI-1, leading among open-weight models. Compared with the original DeepSeek V4.1 Flash released on September 10, 2026, 1017's ARC-AGI-3 semi-private set score rose from 2.68% to 5.81%, while its public set score fell from 11.05% to 8.97%. At maximum reasoning effort, the reasoning tokens used by 1017 for its first action decreased by an average of 36% across 25 public games and by 48.6% across 55 semi-private games.
ARC Prize announced DeepSeek V4.1 Flash 1017's results on the ARC-AGI series of benchmarks: 5.81% on ARC-AGI-3, 78.1% on ARC-AGI-2, and 94.5% on ARC-AGI-1, leading among open-weight models. Compared with the original DeepSeek V4.1 Flash released on September 10, 2026, 1017's ARC-AGI-3 semi-private set score rose from 2.68% to 5.81%, while its public set score fell from 11.05% to 8.97%. At maximum reasoning effort, the reasoning tokens used by 1017 for its first action decreased by an average of
ARC Prize announces the verified results of DeepSeek V4.1 Flash 1017 on ARC-AGI: ARC-AGI-2 at 78.1% with $0.16 per task, ARC-AGI-1 at 94.5% with $0.06 per task, and ARC-AGI-3 at 5.8% with $1.2K under the provider adapter harness and 0.2% with $1.0K under the standard harness. The model leads open-weight models on ARC-AGI-3 and ARC-AGI-2, and matches the best score of open-weight models on ARC-AGI-1. Compared with the original V4.1 Flash from September 10, 2026, 1017's ARC-AGI-3 semi-private set score rose from 2.68% to 5.81%, while its public set score fell from 11.05% to 8.97%; at maximum reasoning effort, the reasoning tokens consumed by its first action decreased by an average of 36% across 25 public games and by 48.6% across 55 semi-private games.
ARC Prize announces the verified results of DeepSeek V4.1 Flash 1017 on ARC-AGI: ARC-AGI-2 at 78.1% with $0.16 per task, ARC-AGI-1 at 94.5% with $0.06 per task, and ARC-AGI-3 at 5.8% with $1.2K under the provider adapter harness and 0.2% with $1.0K under the standard harness. The model leads open-weight models on ARC-AGI-3 and ARC-AGI-2, and matches the best score of open-weight models on ARC-AGI-1. Compared with the original V4.1 Flash from September 10, 2026, 1017's ARC-AGI-3 semi-private set
ARC Prize announces the evaluation results of DeepSeek V4.1 Flash 1017 on ARC-AGI: it scores 5.81% on the ARC-AGI-3 semi-private set, 78.1% on ARC-AGI-2, and 94.5% on ARC-AGI-1, leading among open-weight models. Compared with the original V4.1 Flash released on September 10, 2026, its ARC-AGI-3 semi-private set score rose from 2.68% to 5.81%, while its public set score fell from 11.05% to 8.97%. At maximum reasoning effort, the model's reasoning tokens consumed for the first action decreased by an average of 36% across 25 public games and by 48.6% across 55 semi-private games.
ARC Prize announces the evaluation results of DeepSeek V4.1 Flash 1017 on ARC-AGI: it scores 5.81% on the ARC-AGI-3 semi-private set, 78.1% on ARC-AGI-2, and 94.5% on ARC-AGI-1, leading among open-weight models. Compared with the original V4.1 Flash released on September 10, 2026, its ARC-AGI-3 semi-private set score rose from 2.68% to 5.81%, while its public set score fell from 11.05% to 8.97%. At maximum reasoning effort, the model's reasoning tokens consumed for the first action decreased by
NVIDIA has released the Cosmos Framework and Cosmos NIM for physical AI. The former provides modular tools and agent skills for training, reinforcement learning, optimization, and inference, while the latter provides accelerated inference services in container form. The Cosmos 3 Nano simulation recipe can generate synchronized camera video and LiDAR observations from released checkpoints, covering autonomous driving, robotics, and surgical scenarios. NIM uses quantization, speculative decoding, and KV-cache-aware routing to achieve a measured speedup of about 2–3x on H200, and Cosmos3-Nano generation tasks require 58 GiB of VRAM under BF16.
NVIDIA has released the Cosmos Framework and Cosmos NIM for physical AI. The former provides modular tools and agent skills for training, reinforcement learning, optimization, and inference, while the latter provides accelerated inference services in container form. The Cosmos 3 Nano simulation recipe can generate synchronized camera video and LiDAR observations from released checkpoints, covering autonomous driving, robotics, and surgical scenarios. NIM uses quantization, speculative decoding,
An AWS official blog post explains how to fine-tune a shopping agent on Amazon SageMaker AI using sequence-level knowledge distillation (SeqKD): using GLM-5.3 on Amazon Bedrock as the teacher model to generate trajectories, using the ShoppingBench scorer to retain only trajectories that solve the task, and then performing serverless SFT on the Qwen3.5-4B student model. On 324 held-out test questions, the distilled 4B student's average absolute success rate increased from 0.297 to 0.541, essentially on par with the teacher model and mid-sized frontier models; when the teacher's reasoning process was removed and only tool calls were learned, the average success rate dropped to 0.415.
An AWS official blog post explains how to fine-tune a shopping agent on Amazon SageMaker AI using sequence-level knowledge distillation (SeqKD): using GLM-5.3 on Amazon Bedrock as the teacher model to generate trajectories, using the ShoppingBench scorer to retain only trajectories that solve the task, and then performing serverless SFT on the Qwen3.5-4B student model. On 324 held-out test questions, the distilled 4B student's average absolute success rate increased from 0.297 to 0.541, essentia
In October, Google launched its frontier reasoning model Gemini 4 Argon globally, covering Google AI Studio, the Gemini API, Google Cloud, Search, and the Gemini app. Two weeks later, it announced Gemini 4 Carbon, focused on long-horizon engineering and cybersecurity defense, which is currently undergoing pre-release testing with the U.S. government and select trusted cybersecurity parties through the Fairwind Program. In the same period, it released Genie 4, a real-time interactive world model that supports native audio and natural-language scene control; SynthID Detector opened globally in English and can identify AI content from partners such as OpenAI, NVIDIA, Kakao, and Apple.
In October, Google launched its frontier reasoning model Gemini 4 Argon globally, covering Google AI Studio, the Gemini API, Google Cloud, Search, and the Gemini app. Two weeks later, it announced Gemini 4 Carbon, focused on long-horizon engineering and cybersecurity defense, which is currently undergoing pre-release testing with the U.S. government and select trusted cybersecurity parties through the Fairwind Program. In the same period, it released Genie 4, a real-time interactive world model
IEEE Spectrum reported on the PHP2Shell vulnerability, which affects websites running WordPress or Drupal, including those of NASA, Wikipedia, and the White House; it can read sensitive data and execute malicious code on servers. After discovering it, researchers reported it to the PHP Foundation and completed patching before it was exploited. Suman Jana of Columbia University believes that exploiting the vulnerability requires a great deal of tedious, specific reasoning, and that AI performs well in this regard but cannot complete it independently; he mentioned that a new tool released by Anthropic on October 22 allows LLMs to generate patches and have LLMs generate correctness proofs, with only the proof checker guaranteed by humans. Matthew Green of Johns Hopkins University judges that, until a more serious incident occurs, all parties will not truly invest in fixes.
IEEE Spectrum reported on the PHP2Shell vulnerability, which affects websites running WordPress or Drupal, including those of NASA, Wikipedia, and the White House; it can read sensitive data and execute malicious code on servers. After discovering it, researchers reported it to the PHP Foundation and completed patching before it was exploited. Suman Jana of Columbia University believes that exploiting the vulnerability requires a great deal of tedious, specific reasoning, and that AI performs we
Browser Use has opened its fastest browser agent, BU Ultrafast, from the waitlist to all Cloud plans in the US, including the free tier. Official data shows a median task time of 22 seconds and a cost of $0.0064 per task, compared with 79 seconds and $0.0098 for the default agent on GPT-6 Luna. To use it, simply set model to bu-ultrafast; if you need more step-by-step reasoning, you can switch to bu-fast from the same agent, which is billed at standard rates.
Krea announced that its research lead @sleenyre will speak at the SGLang Summit, covering reinforcement learning and inference scaling for Krea 3. The summit takes place November 12–13 at Fort Mason in San Francisco.
AWS presents a five-step closed loop for managing Amazon Bedrock costs: attribute by IAM principal, record request observability, control spending with budget alerts and hard blocks, optimize platform defaults such as model routing and caching, and review on a regular cadence. Amazon Bedrock now supports automatically attributing inference costs to the IAM principal that initiated the call, with the caller ARN written to the line_item_iam_principal column in CUR 2.0, covering IAM users, roles, federated identities, and API keys. In shared gateway scenarios, role session names and session tags can be set per user, credentials are cached for up to 1 hour, and attribution by user and tenant is restored.
AWS presents a five-step closed loop for managing Amazon Bedrock costs: attribute by IAM principal, record request observability, control spending with budget alerts and hard blocks, optimize platform defaults such as model routing and caching, and review on a regular cadence. Amazon Bedrock now supports automatically attributing inference costs to the IAM principal that initiated the call, with the caller ARN written to the line item iam principal column in CUR 2.0, covering IAM users, roles, f
OpenAI releases the GPT-Rosalind-6 system card addendum. This model uses the same set of weights as GPT-6 Astra, with no new training data or post-training added, and is provided only through trusted access to vetted research institutions, enterprises, and government partners. At the highest reasoning budget, MedChemBench v7 averages 49.6% (GPT-Rosalind-5.5: 28.9%), GeneBench Pro v13 single-attempt success rate is 37.7% (previous generation: 16.7%), and LifeSciBench Gold average score is 61.3% (previous generation: 44.0%). Under the Preparedness Framework, this release is rated High capability in the biological and chemical domains, below the Critical threshold; rated Critical in the cybersecurity domain; and below High in AI self-improvement.
OpenAI releases the GPT-Rosalind-6 system card addendum. This model uses the same set of weights as GPT-6 Astra, with no new training data or post-training added, and is provided only through trusted access to vetted research institutions, enterprises, and government partners. At the highest reasoning budget, MedChemBench v7 averages 49.6% (GPT-Rosalind-5.5: 28.9%), GeneBench Pro v13 single-attempt success rate is 37.7% (previous generation: 16.7%), and LifeSciBench Gold average score is 61.3% (
UCSD Hao AI Lab has open-sourced JetSpec++, which pipelines the draft stage and verification through causal parallel drafting, a hidden-state predictor, and an anchor-token decoder, achieving 1,361 TPS on Qwen3-8B, up to 1.54x end-to-end speedup over DSpark, and 7.45x speedup over autoregressive decoding while remaining lossless. Combined with NVFP4 and kernel optimizations, Qwen3-8B exceeds 4,000 TPS and Qwen3.8-27B exceeds 900 TPS (on two GB200 GPUs). The model, code, and inference engine have all been open-sourced.
UCSD Hao AI Lab has open-sourced JetSpec++, which pipelines the draft stage and verification through causal parallel drafting, a hidden-state predictor, and an anchor-token decoder, achieving 1,361 TPS on Qwen3-8B, up to 1.54x end-to-end speedup over DSpark, and 7.45x speedup over autoregressive decoding while remaining lossless. Combined with NVFP4 and kernel optimizations, Qwen3-8B exceeds 4,000 TPS and Qwen3.8-27B exceeds 900 TPS (on two GB200 GPUs). The model, code, and inference engine have
UCSD Hao AI Lab has launched JetSpec++, which, through causal parallel drafting and pipelined verification, achieves lossless end-to-end speedups of 1.54× over DSpark and 7.45× over autoregressive decoding on Qwen3-8B, reaching 1,361 TPS in BF16. On GSM8K, Qwen3-8B's per-round latency drops from 9.92 ms with DSpark to 6.84 ms, with drafting time compressed from 3.76 ms to 0.61 ms. The model, code, and inference engine have all been open-sourced.
Hao AI Lab proposes JetSpec++, which pipelines draft generation and verification through causal parallel drafting, a hidden-state predictor, and an anchor-token decoder, achieving lossless speedup on Qwen3-8B and Qwen3.8-27B. Compared with DSpark, it achieves up to 1.54x end-to-end speedup, and 7.45x speedup over autoregressive decoding, reaching 1,361 TPS and 367 TPS respectively under BF16; with NVFP4 and kernel optimizations, it exceeds 4,000 TPS and 900 TPS on two GB200s. Across 7 benchmarks, it accepts 8.94 tokens per round, compared with 6.89 for DSpark; the model, code, and inference engine have all been open-sourced.
Hao AI Lab proposes JetSpec++, which pipelines draft generation and verification through causal parallel drafting, a hidden-state predictor, and an anchor-token decoder, achieving lossless speedup on Qwen3-8B and Qwen3.8-27B. Compared with DSpark, it achieves up to 1.54x end-to-end speedup, and 7.45x speedup over autoregressive decoding, reaching 1,361 TPS and 367 TPS respectively under BF16; with NVFP4 and kernel optimizations, it exceeds 4,000 TPS and 900 TPS on two GB200s. Across 7 benchmarks
Anthropic's revenue in the second quarter of this year was 11.555 billion US dollars, achieving an operating profit of 974 million US dollars for the first time, with about 80% of revenue coming from pay-as-you-go billing. IDC data shows that in the first half of 2026, China's public cloud MaaS market revenue reached 10.63 billion yuan, up 889.4% year on year, and its share of public cloud PaaS revenue rose from 5.2% in the second half of last year to 21.0%. Zhipu's open platform and API revenue was 825 million yuan, accounting for 86.5% of total revenue; Volcano Engine ranked first with a 45.1% share. The article argues that model services can now generate gross profit, and the unit of competition has shifted to inference optimization and delivery capability.
Anthropic's revenue in the second quarter of this year was 11.555 billion US dollars, achieving an operating profit of 974 million US dollars for the first time, with about 80% of revenue coming from pay-as-you-go billing. IDC data shows that in the first half of 2026, China's public cloud MaaS market revenue reached 10.63 billion yuan, up 889.4% year on year, and its share of public cloud PaaS revenue rose from 5.2% in the second half of last year to 21.0%. Zhipu's open platform and API revenue
