fal Research, in partnership with Black Forest Labs, has launched the video model FLUX 3 Max. Through dedicated post-training and inference optimization, it is tied for first place in image-to-video quality on independent third-party benchmarks, generating a 10-second 720/768p video in about 12 seconds. fal says its latency is only a fraction of that of competing frontier models, and the model will be 75% off for the next three weeks.
Ant Group's inclusionAI has open-sourced the Ming-3.0-audio-flash native audio model on GitHub. It uses a 126B-parameter MoE architecture with 6.5B activated parameters, and there is also a Ming-3.0-audio-tiny version with 9.3B parameters and 2.7B activated parameters. The models cover Chinese and English recognition across multiple dialects and accents, long audio understanding, audio reasoning, and end-to-end speech dialogue. They support three inference methods: vLLM-Omni, SGLang, and Transformers, and are released under the MIT license.
Ant Group's inclusionAI has open-sourced the Ming-3.0-audio-flash native audio model on GitHub. It uses a 126B-parameter MoE architecture with 6.5B activated parameters, and there is also a Ming-3.0-audio-tiny version with 9.3B parameters and 2.7B activated parameters. The models cover Chinese and English recognition across multiple dialects and accents, long audio understanding, audio reasoning, and end-to-end speech dialogue. They support three inference methods: vLLM-Omni, SGLang, and Transfo
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.
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 achieve
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.
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
Perplexity has released its next-generation open-weight multimodal decision models, comprising two sizes, pplx-decider-v1.2 and pplx-decider-flash-v1.2, with improvements in both text and vision scores, inference up to 5x faster, and all questions in a single request completed in one pass. Hugging Face co-founder Clément Delangue reposted the news and noted that the weights have been published on Hugging Face. Perplexity says the flash version is the top-ranked 4B model on the Decision Index.
Perplexity has open-sourced two decision models: pplx-decider-flash-v1.2 has 4B parameters and costs 1 cent per million tokens, making it the strongest 4B model on the HF Decision Index; pplx-decider-v1.2 has 27B parameters and costs 2 cents per million tokens, ranking 2nd on the overall HF Decision Index leaderboard. Both are now live on the Perplexity Decisions API and offer open weights on Hugging Face. Perplexity Developers says this generation of open-weight multimodal decision models improves on both text and vision scores, delivers inference up to 5x faster, and completes all questions in a request in a single forward pass.
Perplexity has open-sourced two decision models: pplx-decider-flash-v1.2 has 4B parameters and costs 1 cent per million tokens, making it the strongest 4B model on the HF Decision Index; pplx-decider-v1.2 has 27B parameters and costs 2 cents per million tokens, ranking 2nd on the overall HF Decision Index leaderboard. Both are now live on the Perplexity Decisions API and offer open weights on Hugging Face. Perplexity Developers says this generation of open-weight multimodal decision models impro
Krea 3 is now live on OpenRouter. It is Krea AI's new image editing model, which can accept 1–4 reference images and a prompt, and returns natively 2K edited images, with most edits completing end-to-end in under 8 seconds. Pricing is $0.03 per 1K image and $0.06 per 2K image; reference images cost $0.02 per 1K image and $0.07 per 2K image. thinking_mode defaults to fast; when more prompt reasoning is needed, it can be set to high (an additional $0.02). It uses the same Images API as other image models on OpenRouter, so migration only requires changing the model ID.
Krea 3 is now live on OpenRouter. It is Krea AI's new image editing model, which can accept 1–4 reference images and a prompt, and returns natively 2K edited images, with most edits completing end-to-end in under 8 seconds. Pricing is $0.03 per 1K image and $0.06 per 2K image; reference images cost $0.02 per 1K image and $0.07 per 2K image. thinking mode defaults to fast; when more prompt reasoning is needed, it can be set to high (an additional $0.02). It uses the same Images API as other image
NVIDIA has added three models to the Nemotron open-source model family, covering enterprise workflows, local AI agents, and conversational applications across data centers, desktops, and edge devices. These include the 120-billion-parameter Nemotron 3.5 Super VL, which NVIDIA claims delivers up to 5x the throughput of the next-best comparable model, with accuracy approaching that of its largest reasoning model at roughly one-fifth the parameter count; there are also two small models in the Nemotron 3.5 Bolt series: 74B-A9B 24GB for everyday agent tasks on RTX and DGX Spark, and E4B 4GB for conversational and physical AI agents. Super VL and Bolt 74B-A9B are already available, and Bolt E4B is scheduled for release on October 23.
NVIDIA has added three models to the Nemotron open-source model family, covering enterprise workflows, local AI agents, and conversational applications across data centers, desktops, and edge devices. These include the 120-billion-parameter Nemotron 3.5 Super VL, which NVIDIA claims delivers up to 5x the throughput of the next-best comparable model, with accuracy approaching that of its largest reasoning model at roughly one-fifth the parameter count; there are also two small models in the Nemot
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.
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 th
Hugging Face has released Beam, a 501B-parameter sparse MoE open source model aimed at long-text reasoning and agent tasks with 1M context, activating only about 23B parameters per token. The model uses 52 layers of hybrid attention, with 43 layers using a 512-token sliding window and 9 layers using global attention; the KV cache grows by 36 KiB per token, about one-sixth that of a full-attention approach. The official release provides three checkpoints—BF16, FP8, and NVFP4; NVFP4 can compress the weights from about 1 TB to about 280 GB, and day-0 support is already available in transformers, vLLM, and SGLang.
Hugging Face has released Beam, a 501B-parameter sparse MoE open source model aimed at long-text reasoning and agent tasks with 1M context, activating only about 23B parameters per token. The model uses 52 layers of hybrid attention, with 43 layers using a 512-token sliding window and 9 layers using global attention; the KV cache grows by 36 KiB per token, about one-sixth that of a full-attention approach. The official release provides three checkpoints—BF16, FP8, and NVFP4; NVFP4 can compress t
Reflection's first open-weight model, Beam, is officially released, offered free for a limited time with Ollama as a day-zero partner and with zero data retention (ZDR). Beam is an MoE model with approximately 501 billion total parameters, about 23 billion activated per token, and a 1 million token context window. It targets software engineering, reasoning, and long-horizon agentic tasks, and is released under the Apache 2.0 license. Users can run beam:cloud via Ollama Cloud and connect it to tools such as Claude Code, Codex, and OpenCode.
Reflection's first open-weight model, Beam, is officially released, offered free for a limited time with Ollama as a day-zero partner and with zero data retention (ZDR). Beam is an MoE model with approximately 501 billion total parameters, about 23 billion activated per token, and a 1 million token context window. It targets software engineering, reasoning, and long-horizon agentic tasks, and is released under the Apache 2.0 license. Users can run beam:cloud via Ollama Cloud and connect it to to
