English text is being prepared.

English translation is being prepared.

Ant inclusionAI Open-Sources Ming-3.0-asr-tiny: Supports Chinese-English Transcription, Dialect Recognition, and Context ASR

Ant inclusionAI has open-sourced Ming-3.0-asr-tiny on GitHub, combining the Ling-tiny-v3 language model with an audio encoder to support Chinese-English transcription, dialect recognition, and Context ASR. Dialect evaluation covers 15 test sets including Hunan, Minnan, Cantonese, and Sichuan-Chongqing; Context ASR can recognize proper nouns via hotwords in prompts or preceding dialogue. The inference stack uses Transformers 5.12.1, fla-core 0.5.2, and PyTorch 2.11.0, and provides an SGLang accelerated deployment example.

Memory Tensor Open-Sources MemT 1.0 Series of Memory-Native Models

Memory Tensor has open-sourced the MemT 1.0 series of memory-native models, comprising two variants: Nano with 36B parameters and 3B activated, and Flash with 288B parameters and 17B activated, built on the Qwen3.6-35B-A3B and DeepSeek-V4-Flash bases, respectively. The models add a memory branch parallel to attention within selected Transformer modules, encoding experience into an updatable matrix state that participates in subsequent generation. Across 500 SWE-bench Verified tasks, Nano achieved an average success rate of 69.73% at key checkpoints, versus a no-experience baseline of 66.07%; under a 128K long history, end-to-end inference speed is approximately 1.86 times that of the full-context approach, and the persistent state footprint per trajectory is one sixty-sixth that of the full-context KV Cache.

English translation is being prepared.

Meta's open-weight model Muse Spark Open launches on Fireworks: 816B parameters, 1M context

Meta's open-weight model Muse Spark Open is available on Fireworks from day one of launch. It is an 816B-parameter MoE model that activates 42B parameters per token, with a 1M-token context window, targeting long-horizon coding and agentic tasks. Fireworks says it ranks first among 75 models on Vals AI's Harvey legal agent benchmark with 25.4%, nearly 4 times that of Claude Opus 5 (6.7%); on the Rogo Big Finance Bench it surpasses Claude Opus 5.5 at every difficulty level. Fireworks validated the weights with its own production stack, scoring 91.1% on AIME26, 89.7% on GPQA Diamond, and 84.3% on SWE-bench Verified, and reproduced consistent results with SGLang.

English translation is being prepared.

Peter Yang agrees with Tim Sweeney's views on AI programming

Peter Yang said Tim Sweeney is right, and hopes the developers working on StarCraft 3 see this tweet and use AI to finish and release it before 2030. In the quoted post, Tim Sweeney recalled that when developing Unreal Engine 1, about 30% of his time went to hardcore coding, and 70% to typing code and refactoring; he believes handing the latter to AI is only natural, and was surprised that AI can also solve hard technical problems like "insta-Carmack" and understand complex 3D scenes.

English translation is being prepared.

English translation is being prepared.

MiniMax Claims Blind-Test Model Space Bunny, Confirms It as MiniMax-M3.1-Flash

MiniMax officially claimed Space Bunny, a model tested anonymously on an overseas blind-test platform, confirming it as MiniMax-M3.1-Flash. The model supports a generation speed on the order of one hundred words per second and a one-million-token context; community tests measured generation speed steady at 85 to 112 words per second, capable of running through complex 3D scenes and code builds in a single pass within three minutes, and achieving near-imperceptible response in multi-file project repairs. The author believes this approach of anonymous blind testing first and public claiming afterward turns inference speed and million-token context into efficiency tools usable for everyday coding.

Vercel AI Gateway adds agent session view

According to the Vercel Developers account, AI Gateway logs can now show each agent session, grouping requests by conversation and displaying cost, cache hit rate, and the agent that ran the session; session affinity pins each session to the provider that holds its prompt cache. Guillermo Rauch reposted and praised this unified view, saying it covers personal and coding agents, apps, and API keys, and emphasized that Vercel AI Gateway gives users control over model choice, data (ZDR), and the metadata needed to understand RoI.

English translation is being prepared.

Anthropic Report: AI Agents' Work Hours Reach 6 Times That of Humans, Share of Junior Tech Job Openings Falls to 9%

Anthropic released a report stating that by August 2026, the working hours of interactive AI agents deployed by employees had reached 6 times that of human employees, and in March, the company-wide total agent work hours exceeded human work hours for the first time. The report shows that the share of junior employees in the technology department fell from 21% in early 2025 to 14%, the share of junior tech job openings fell from 27% to 9%, and engineers using coding agents averaged 5 times as many code commits as they did 6 months earlier. The report also says its revenue per employee is 2 to 4 times that of Meta and Google during their high-growth periods, but Epoch AI cautions that the related estimates carry uncertainty.

Reflection's Open-Weight Model Beam Is Free for a Limited Time on Vercel AI Gateway

The Vercel developer account announced that Reflection's Beam model is free for a limited time on Vercel AI Gateway. It previously ran incognito under the codename stealth/cat-jam; it is a low-latency US open-weight model, and its call name is reflection/beam-free. Guillermo Rauch relayed that the model can be connected to deployable eve agents, used to start coding with fx, or used to build your own harness.

AI Is Getting Smarter, but Consistency Across Agent Swarms Is Getting Harder

Madhu Shashanka, co-founder of Concentric AI, points out that as coding agents rapidly become more reliable, enterprises will deploy large numbers of agents, and the real challenge will shift from the reliability of individual agents to consistency across the swarm. He cites the example of 10 agents, each with 95% accuracy, chained together, where overall accuracy drops to about 60%, to show that reliability does not compound. He proposes that organizations need a continuous self-awareness mechanism akin to proprioception; otherwise, they can only barely keep things going through manual auditing and periodic reviews.

Meta and Apple Partner to Release Open-Weight Model Muse Spark Open

Meta has released the open-weight model Muse Spark Open and worked with Apple's hardware team to enable it to run efficiently on the M5 Mac Studio Ultra. The model has 42 billion active parameters, a 1 million token context window, and multimodal capabilities. Meta says it performs well on agentic and coding tasks, and the weights have been made public on HuggingFace under Meta's Muse Model license. Running the model on an Apple computer requires two Mac Studios equipped with the M5 Ultra, each starting at $5,499.

Meta releases open-weight model Muse Spark Open

Meta released the open-weight model Muse Spark Open on Thursday, saying its capabilities can match and in some scenarios surpass comparable products from Chinese vendors. Meta AI head Alexandr Wang called it the most capable U.S. open-weight model to date; the model targets coding and agentic tasks, excels at long-horizon tasks, complex software engineering and multi-step reasoning, and has built-in safety safeguards. The model is optimized for the Apple M5 Ultra chip, and users can run it on the desktop with two Mac Studios equipped with that chip.

English translation is being prepared.

Reflection's Open-Weight Model Beam Launches on Baseten: 501B MoE with 1M Context

Reflection's first open-weight model, Beam, is now available on the Baseten platform, with day-0 support through Baseten Model APIs and Dedicated Inference. Beam uses a MoE architecture with 501B total parameters and 23B activated per token, a context window of up to 1 million tokens, and is aimed at coding and agent workflows. During training, Reflection used longer RL rollouts and added a length penalty to encourage the model to complete tasks with fewer tokens; users can also adjust reasoning effort via a parameter to control how much reasoning the model invests.

English translation is being prepared.