English translation is being prepared.
Matched fieldsTitle、Summary、Body
←Search results
Search “DeepSeek”
English translation is being prepared.
Matched fieldsTitle、Summary、Body
English translation is being prepared.
Matched fieldsTitle、Summary、Body
English translation is being prepared.
Matched fieldsTitle、Summary
CWE-bench v2: Comprehensive improvements in coverage, quality, and transparency
CWE-bench v2 publishes evaluation results: Among 11 models, Gemini 4 Argon leads with 67.8% pass@1 and 81.0% pass@4, followed closely by Claude Opus 5.5 with 67.3% pass@1. 95 tasks were solved by at least one model, but the top model still fails 19% of attempts. On cost, DeepSeek V4.1 Flash, with 57.0% pass@1 and $0.04 per rollout, forms the Pareto frontier together with Claude Opus 5.5 and Gemini 4 Argon; the dataset, scorer, and harness are all publicly available and can be run directly from the Hub via Harbor.
shes evaluation results: Among 11 models, Gemini 4 Argon leads with 67.8% pass@1 and 81.0% pass@4, followed closely by Claude Opus 5.5 with 67.3% pass@1. 95 tasks were solved by at least one model, but the top model still fails 19% of attempts. On cost, DeepSeek V4.1 Flash, with 57.0% pass@1 and $0.04 per rollout, forms the Pareto frontier together with Claude Opus 5.5 and Gemini 4 Argon; the dataset, scorer, and harness are all publicly available and can be run directly from the Hub via Harbor.
Matched fieldsSummary、Body
Particle-1 launches on OpenClaw
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.
Matched fieldsSummary、Body
NVIDIA uses a 12,000-line Megakernel to optimize Kimi K3 inference, and vLLM says AI code is now faster than human review
Benjamin Chislett, an engineer at NVIDIA CentML, used a 12,000-line Megakernel written in CuTe DSL to optimize Moonshot AI's Kimi K3 inference. The vLLM blog reports over 840 tokens/s at batch 1 and a 1.5x to 2x increase in decoding throughput across multiple configurations, while SemiAnalysis estimates 423 tok/s/user on GB300 NVL72, a 1.68x improvement. The vLLM blog says agent-optimized inference code is now outpacing human review and maintenance speed, so the team proposed a fast-track model optimization RFC that lets reviewers judge designs from high-level abstractions rather than reviewing code line by line. The article also mentions that NVIDIA is developing optimizations for other models for vLLM, such as DeepSeek V4.1.
oughput across multiple configurations, while SemiAnalysis estimates 423 tok/s/user on GB300 NVL72, a 1.68x improvement. The vLLM blog says agent-optimized inference code is now outpacing human review and maintenance speed, so the team proposed a fast-track model optimization RFC that lets reviewers judge designs from high-level abstractions rather than reviewing code line by line. The article also mentions that NVIDIA is developing optimizations for other models for vLLM, such as DeepSeek V4.1.
Matched fieldsSummary、Body
Baidu Search Achieves Recursive Self-Evolution, Baidu App Upgrades Four Major AI Features
Baidu Search has achieved multi-Agent collaborative self-evolution through an online feedback loop of "retrieval—feedback—retraining," improving overall iteration efficiency by more than 80%. At the Baidu App product upgrade briefing on October 30, Baidu also upgraded its search AI API, priced as low as 4.9 yuan per thousand calls during the promotional period, lower than the DeepSeek search API; the API's content layer covers more than 80 languages and a high-quality index of hundreds of billions of entries, with data credibility improved by over 50%, and the Ernie Assistant saw a 20% performance improvement after integration.
improving overall iteration efficiency by more than 80%. At the Baidu App product upgrade briefing on October 30, Baidu also upgraded its search AI API, priced as low as 4.9 yuan per thousand calls during the promotional period, lower than the DeepSeek search API; the API's content layer covers more than 80 languages and a high-quality index of hundreds of billions of entries, with data credibility improved by over 50%, and the Ernie Assistant saw a 20% performance improvement after integration.
Matched fieldsSummary、Body
Garry Tan recommends pairing Particle-1 with Capy
Capy announced that Particle-1 is now live on its platform: the model is post-trained on DeepSeek V4.1 Flash, supports 1M context and image input, and is priced at $0.25 / $1 per million tokens. Garry Tan reposted, saying that this "most cost-effective code generation model" can be paired with Capy, the "easiest-to-use code generation agent cluster harness."
Matched fieldsSummary
Artificial Analysis releases the Progress Index…
Artificial Analysis releases the Progress Index and simultaneously launches Intelligence Index v5. The Progress Index uses item response theory to estimate a general capability score for each model from all evaluations (including tasks that have been retired or not included in the Intelligence Index), for cross-time comparison; the Intelligence Index is still used to measure the intelligence level of current models. Its data shows that after the emergence of reasoning models, the frontier has improved by about 65 to 70 points per year, roughly twice the 2023 rate; open-weight models typically lag closed-source models by about 4 months, and DeepSeek V4.1 Flash 1017 ranks highest among open-weight models with 324 points.
g tasks that have been retired or not included in the Intelligence Index), for cross-time comparison; the Intelligence Index is still used to measure the intelligence level of current models. Its data shows that after the emergence of reasoning models, the frontier has improved by about 65 to 70 points per year, roughly twice the 2023 rate; open-weight models typically lag closed-source models by about 4 months, and DeepSeek V4.1 Flash 1017 ranks highest among open-weight models with 324 points.
Matched fieldsSummary、Body
Reinforcement Learning Terminology for LLM Readers: How a Single Score Is Distributed to Each Token
The article uses DeepSeek-R1-Zero's pass@1 on AIME 2024 rising from 15.6% to 71.0% as an example to show that a rule-based judge gives only a single score for the entire response, whereas training needs to adjust the probability of each token. The author uses the sampling example "Is OpenAI open?" to explain in turn policy, state, action, reward, return, value, Q, advantage, baseline, TD error, GAE, clipping, and KL, and points out that PPO uses a critic to learn V, while GRPO uses the average score within a group to replace V.
The article uses DeepSeek-R1-Zero's pass@1 on AIME 2024 rising from 15.6% to 71.0% as an example to show that a rule-based judge gives only a single score for the entire response, whereas training needs to adjust the probability of each token. The author uses the sampling example "Is OpenAI open?" to explain in turn policy, state, action, reward, return, value, Q, advantage, baseline, TD error, GAE, clipping, and KL, and points out that PPO uses a critic to learn V, while GRPO uses the average s
Matched fieldsSummary、Body
Artificial Analysis Intelligence Index v5 is live, Claude…
Artificial Analysis Intelligence Index v5 is live, featuring 2 new benchmarks and 3 upgrades, of which 60% are based on private test sets or held-out answers. Claude Opus 5.5 (max) leads with 58 points, followed closely by GPT-6 Astra (max) at 57 points and GPT-6.1 Sol (max) at 55 points, with Gemini 4 Argon (high) further back at 50 points. The gap between open-weight models and the leader has widened from 3 points to 12 points; among released weights, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) lead with 32 points.
gence Index v5 is live, featuring 2 new benchmarks and 3 upgrades, of which 60% are based on private test sets or held-out answers. Claude Opus 5.5 (max) leads with 58 points, followed closely by GPT-6 Astra (max) at 57 points and GPT-6.1 Sol (max) at 55 points, with Gemini 4 Argon (high) further back at 50 points. The gap between open-weight models and the leader has widened from 3 points to 12 points; among released weights, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) lead with 32 points.
Matched fieldsSummary
Hurun List: AI Creates New Chinese Tech Billionaires
The Hurun Rich List shows that AI is spawning a new generation of Chinese tech billionaires: ByteDance's Zhang Yiming tops the list with $99.7 billion, DeepSeek's Liang Wenfeng ranks 7th with $41.7 billion, and Unitree's Wang Xingxing is worth $10.3 billion. The author also notes that Westerners may not know that in China no one wants the top spot, because it brings too much unwanted attention, and PR teams will even beg to have their boss removed from the list.
Matched fieldsSummary、Body
OpenRouter Opens New Account to Post Model Updates
When replying to particle_api, OpenRouter suggested following its new account @ModelsOnOR to get the latest news on OpenRouter model releases. Previously, OpenRouter had introduced that Particle-1 was already live, saying that it uses 28–52% fewer output tokens than DeepSeek V4.1 Flash on four coding benchmarks.
When replying to particle api, OpenRouter suggested following its new account @ModelsOnOR to get the latest news on OpenRouter model releases. Previously, OpenRouter had introduced that Particle-1 was already live, saying that it uses 28–52% fewer output tokens than DeepSeek V4.1 Flash on four coding benchmarks.
Matched fieldsSummary
Meta AI Muse models top friendliness in AA-PersonalityBench
In Artificial Analysis's AA-PersonalityBench evaluation, Meta AI's Muse series models scored highest in friendliness (Friendly): Muse Spark 1.3 scored 100, Muse Spark Open 98, Muse Wonder 95, and Muse Glimmer 93. Friendliness measures whether responses convey warmth and approachability without seeming artificial. The lowest in friendliness were Nemotron 3 Ultra (1), GPT-3.5 Turbo (3), MiniMax-M3.1-Flash (6), Grok 4.6 (8), and DeepSeek V4.1 Flash 1017 (11).
Matched fieldsSummary、Body
ModelBest's ForgeKernel Tops NVIDIA SOL Bench FIB Track
ModelBest OpenBMB announced that ForgeKernel ranks first on both the FIB track of NVIDIA SOL Bench and Huawei CANNBench. Currently, ForgeKernel's goal is to surpass top human experts on one of DeepSeek's most complex operators.
Matched fieldsSummary、Body
OpenBMB agent approaches frontier in 4 hours
ModelBest OpenBMB claims that its agent, starting from scratch in just 4 hours, has reached performance very close to the frontier level. The post cited an earlier post about ForgeKernel Live optimizing the DeepSeek kernel.
Matched fieldsSummary
Chaoyan Intelligence Open-Sources 35B Model Apex-35B-V1.0, with AI Participating in Its Own R&D
Beijing-based large model startup Chaoyan Intelligence has released and open-sourced its first large model, Apex-35B-V1.0, with 35 billion total parameters, approximately 3 billion parameters activated per token, and support for a 262,000-token context window, optimized through mid-training and post-training based on Qwen3.6-35B-A3B. The model scored 60.9 and 40.0 on the paper reproduction test PaperBench and the scientific research challenge test FrontierScience, respectively, higher than GLM-5.3's 42.5 and 30.7; it scored 84.0 on the tool-calling test BFCL V4, higher than DeepSeek-V4.1-Flash-1017's 75.0. The model has been released with open weights under the Apache 2.0 license, and the official release provides approximately 72GB of BF16 weights and GGUF quantized versions at various precisions.
based on Qwen3.6-35B-A3B. The model scored 60.9 and 40.0 on the paper reproduction test PaperBench and the scientific research challenge test FrontierScience, respectively, higher than GLM-5.3's 42.5 and 30.7; it scored 84.0 on the tool-calling test BFCL V4, higher than DeepSeek-V4.1-Flash-1017's 75.0. The model has been released with open weights under the Apache 2.0 license, and the official release provides approximately 72GB of BF16 weights and GGUF quantized versions at various precisions.
Matched fieldsSummary、Body
Replit releases ViBench 2.0: a benchmark for long-horizon software building
Replit releases ViBench 2.0 to evaluate AI agents' ability to build a complete product feature by feature starting from an empty project, comprising 17 products, 241 feature requests, and 544 test plans. GPT-6.1 Sol leads with a 65% pass@1, GPT-6 Astra is at 60%, and Opus 5.5 is at 49%, with all three passing over 90% of test plans. Opus 5.5 completes a build in about 70 minutes on average, roughly 2.5 times faster than Sol; DeepSeek V4.1 Flash passes 77% of test plans at about $2.50 per build, but no open-weight model can pass a complete build.
ability to build a complete product feature by feature starting from an empty project, comprising 17 products, 241 feature requests, and 544 test plans. GPT-6.1 Sol leads with a 65% pass@1, GPT-6 Astra is at 60%, and Opus 5.5 is at 49%, with all three passing over 90% of test plans. Opus 5.5 completes a build in about 70 minutes on average, roughly 2.5 times faster than Sol; DeepSeek V4.1 Flash passes 77% of test plans at about $2.50 per build, but no open-weight model can pass a complete build.
Matched fieldsSummary、Body
After Anthropic Tightened Its Defenses, Chinese AI Labs Turned to OpenAI's GPT-6 Astra for Distillation
According to The Information, DeepSeek and Moonshot AI made extensive use of outputs from OpenAI's most powerful model, GPT-6 Astra, when training their latest models, involving Kimi K3.1 Pro and V4.1-Flash 1017. This kind of distillation requires models to expose hidden reasoning steps, and both OpenAI and Anthropic prohibit the practice. Claude was once the primary target, but Anthropic says the new methods have cut the amount of data attackers can obtain by 90% to 99%. In late September, OpenAI reported a coordinated distillation campaign linked to people associated with Moonshot.
According to The Information, DeepSeek and Moonshot AI made extensive use of outputs from OpenAI's most powerful model, GPT-6 Astra, when training their latest models, involving Kimi K3.1 Pro and V4.1-Flash 1017. This kind of distillation requires models to expose hidden reasoning steps, and both OpenAI and Anthropic prohibit the practice. Claude was once the primary target, but Anthropic says the new methods have cut the amount of data attackers can obtain by 90% to 99%. In late September, Open
Matched fieldsSummary、Body