B total parameters and activates 23B per token, and uses MiniMax Sparse Attention and DSpark speculative decoding to improve inference efficiency; the weights are now available on Hugging Face and ModelScope, with API and MCode desktop entry points also provided. On the same day, Together AI, as a day-one launch partner, deployed the model to its AI Native Cloud production environment, so users do not need to manage GPUs themselves; according to Together AI, the model targets agentic tasks, codi
According to a WSJ report, OpenAI fired 3 safety researchers after a leak investigation. The report says safety staff discovered disappearing Signal chats between these 3 safety researchers and an outside safety researcher. The phone checks began with a September news report that said OpenAI's latest model Astra uses a new approach that makes it smarter but writes down less of its thought process, so humans can see less of the reasoning. OpenAI treated the report as a leak of its secret model design and began tracing the source of the information; investigators listed Ryan Greenblatt, who had helped METR analyze logs from the July Hugging Face hack, as a suspected leaker, and he denies it. Currently the public information comes only from this report, with no further developments yet.
latest model Astra uses a new approach that makes it smarter but writes down less of its thought process, so humans can see less of the reasoning. OpenAI treated the report as a leak of its secret model design and began tracing the source of the information; investigators listed Ryan Greenblatt, who had helped METR analyze logs from the July Hugging Face hack, as a suspected leaker, and he denies it. Currently the public information comes only from this report, with no further developments yet.
Mifeng Technology, together with nearly 20 institutions including Hugging Face, NVIDIA, and Tsinghua University, co-initiated the embodied evaluation platform RoboColiseum, and announced the results of this round of evaluation on October 20. According to a Leikeji report on October 23, Claude Opus 5.5 ranked first overall among the four flagship models, followed closely by GPT-6 Astra; GPT-6 Astra scored 99.35 on instruction following, while Opus 5.5 performed better on spatial understanding and general manipulation, and consumed the fewest tokens. The platform contains 78 simulation evaluation tasks, with a simulation-to-real correlation of 89.5%, and more than 100 teams have participated since it launched in August.
Mifeng Technology, together with nearly 20 institutions including Hugging Face, NVIDIA, and Tsinghua University, co-initiated the embodied evaluation platform RoboColiseum, and announced the results of this round of evaluation on October 20. According to a Leikeji report on October 23, Claude Opus 5.5 ranked first overall among the four flagship models, followed closely by GPT-6 Astra; GPT-6 Astra scored 99.35 on instruction following, while Opus 5.5 performed better on spatial understanding and
Anthropic announced that it will donate Agent Skills to the Agentic AI Foundation (AAIF). According to its first-party information, MCP had previously also been donated to the same foundation; since December 2025, more than 70 agent platforms have adopted the standard, including ChatGPT, Cursor, GitHub Copilot, and VS Code. Subsequent reports stated that AAIF, under the Linux Foundation, announced on October 22, 2026 that it had accepted four projects: in addition to Anthropic's Agent Skills, these include OpenAI's Open Responses, Hugging Face's OpenEnv, and Alibaba's OpenSandbox; Agent Skills is a universal format that allows different agents to reuse business knowledge and workflows, and Anthropic made it public as an open standard in December 2025. Reports also stated that the donation includes the name, logo, the GitHub organization account agentskills, and the agentskills.io domain; going forward, a technical operations committee composed of core maintainers from multiple organizations will decide the development direction, with Anthropic's David Soria Parra and Mahesh Murag serving as lead maintainers; the number of projects under AAIF has grown from three at its founding to ten.
AAIF, under the Linux Foundation, announced on October 22, 2026 that it had accepted four projects: in addition to Anthropic's Agent Skills, these include OpenAI's Open Responses, Hugging Face's OpenEnv, and Alibaba's OpenSandbox; Agent Skills is a universal format that allows different agents to reuse business knowledge and workflows, and Anthropic made it public as an open standard in December 2025. Reports also stated that the donation includes the name, logo, the GitHub organization account
Tencent Hunyuan launched the open-source interactive video world model Hy-Nomas on October 24, 2026. According to official release information, the model supports interaction using text prompts and camera trajectories, and has long-horizon generation capability with world memory; it can be adapted to specialized domains such as robotics via lightweight LoRA, and has been optimized for inference on consumer-grade GPUs. The official release also provides website, GitHub, Hugging Face, and demo entry points. There are currently no subsequent development facts.
d model Hy-Nomas on October 24, 2026. According to official release information, the model supports interaction using text prompts and camera trajectories, and has long-horizon generation capability with world memory; it can be adapted to specialized domains such as robotics via lightweight LoRA, and has been optimized for inference on consumer-grade GPUs. The official release also provides website, GitHub, Hugging Face, and demo entry points. There are currently no subsequent development facts.
Google collaborated with Meta's PyTorch team to launch the open-source project TorchTPU, providing a native PyTorch backend for Google TPU; migration only requires changing model.to("cuda") to model.to("tpu"). The project supports both eager and torch.compile modes, adds a new DeferAndFuse eager mode, covers about 450 Aten operators, and has been verified to run out of the box on 4,917 Hugging Face models, with a Tier 1 pass rate of 98% and Tier 2 numerical consistency relative to GPU of 97%. The official statement says it has been used in the inference and training production workloads of some Google Cloud major customers, and is integrated with vLLM, SGLang, TorchTitan, Ray, and TorchRec. Subsequent reports said that at PyTorch Conference North America 2026, Google open-sourced TorchTPU, whose runtime no longer depends on JAX, and existing workloads can be migrated with only minor code changes; PyTorch noted that TorchTPU's default Skinny Collectives scheme suffers roughly 30–40% end-to-end performance loss under distributed workloads compared with optimized schemes, and achieving peak performance still requires Pallas kernels and vLLM-T
The project supports both eager and torch.compile modes, adds a new DeferAndFuse eager mode, covers about 450 Aten operators, and has been verified to run out of the box on 4,917 Hugging Face models, with a Tier 1 pass rate of 98% and Tier 2 numerical consistency relative to GPU of 97%. The official statement says it has been used in the inference and training production workloads of some Google Cloud major customers, and is integrated with vLLM, SGLang, TorchTitan, Ray, and TorchRec. Subsequen
On October 21, 2026, Recursive Superintelligence released Flint, an open-weight local coding model. According to related reports, Flint is a 40B mixture-of-experts model that activates approximately 2.5B parameters per token, is licensed under Apache 2.0, and its weights have been made available on Hugging Face; it was built from scratch in under 4 months by the automated research system Eureka Machine. On October 22, Hugging Face co-founder Thomas Wolf commented that Recursive researchers quantized the lm_head—which is usually recommended to be kept at high precision—to 4 bits, and the result differed from bf16 by only 0.7 percentage points. On the same day, OpenRouter launched Flint, saying it is aimed at coding agents, supports a 262K context, reasoning, and tool calling, and is free to use during the launch period; Unsloth released a Dynamic GGUF quantized version of Flint, saying it can run locally on devices with 24GB of memory. The model now offers multiple usage paths through open weights, hosting platforms, and quantized versions.
reports, Flint is a 40B mixture-of-experts model that activates approximately 2.5B parameters per token, is licensed under Apache 2.0, and its weights have been made available on Hugging Face; it was built from scratch in under 4 months by the automated research system Eureka Machine. On October 22, Hugging Face co-founder Thomas Wolf commented that Recursive researchers quantized the lm head—which is usually recommended to be kept at high precision—to 4 bits, and the result differed from bf16
Baseten's research team Base Labs has released two decision models, mammoth-1 (27B) and elephant-1 (9B), both fully fine-tuned from Qwen3.5, with model weights made public. According to Baseten Base Labs, mammoth-1 ranks first among 208 models on the Jev Decision Index 0.4 with a score of 71.24, leading the second-place Perplexity Decider v1.2 by 1.8 points; elephant-1 scores 62.65 on that leaderboard. Baseten's engineering blog states that elephant-1 is the highest-ranked 9B decision model, and on the Artificial Analysis Decision Index the two score 65.9 and 65.2 respectively; earlier reports said elephant-1 tied for first on that index with 65 points alongside d1-ultra. Decision models do not generate text; instead they return probability distributions over options, ratings, or yes/no questions, and a confidence threshold can determine whether to accept automatically or escalate to a human. Both models are open-sourced on Hugging Face and available for a one-week free trial on Baseten Model APIs, with zero data retention by default.
score 65.9 and 65.2 respectively; earlier reports said elephant-1 tied for first on that index with 65 points alongside d1-ultra. Decision models do not generate text; instead they return probability distributions over options, ratings, or yes/no questions, and a confidence threshold can determine whether to accept automatically or escalate to a human. Both models are open-sourced on Hugging Face and available for a one-week free trial on Baseten Model APIs, with zero data retention by default.
Hugging Face co-founder Clément Delangue offered the judgment that Europe has more opportunities in AI than in any previous technology wave. Related reports show that this statement came at the occasion of the conclusion of NVIDIA CEO Jensen Huang's Europe trip and the appearance of the Microduck robot; the robot was designed and developed by the Pollen Robotics and Hugging Face teams in Bordeaux, France, and the report says its global sales exceeded 15,000 units in 2026. Delangue said he will work with NVIDIA and the open-source AI community to prove Europe's opportunities in the AI era. Currently this judgment has no subsequent developments or facts.
Hugging Face co-founder Clément Delangue offered the judgment that Europe has more opportunities in AI than in any previous technology wave. Related reports show that this statement came at the occasion of the conclusion of NVIDIA CEO Jensen Huang's Europe trip and the appearance of the Microduck robot; the robot was designed and developed by the Pollen Robotics and Hugging Face teams in Bordeaux, France, and the report says its global sales exceeded 15,000 units in 2026. Delangue said he will w
OlmoEarth v1.3 has been released, including the OlmoEarth v1.3 Base model, global precomputed embeddings, the OlmoEarth Embeddings Explorer, and a technical report. According to Hugging Face, this version generates 128-dimensional embedding vectors for every 10-meter pixel worldwide, with the first 64 dimensions usable on their own, and it has made the global embeddings openly available to the community on Source Cooperative. In 27 evaluations covering six continents, the 128-dimensional version scored highest on 20, and the 64-dimensional version led on 18; training compute was 53% of v1.2 and 18% of v1. Model weights, pretraining data, and code are all open source, and a final run to pretrain the Base version costs approximately $2,720 in H100 rental. There are currently no subsequent development facts.
OlmoEarth v1.3 has been released, including the OlmoEarth v1.3 Base model, global precomputed embeddings, the OlmoEarth Embeddings Explorer, and a technical report. According to Hugging Face, this version generates 128-dimensional embedding vectors for every 10-meter pixel worldwide, with the first 64 dimensions usable on their own, and it has made the global embeddings openly available to the community on Source Cooperative. In 27 evaluations covering six continents, the 128-dimensional version
Ai2 releases PLANTEUR, a benchmark for tropical crop mapping. The benchmark extends the PASTIS benchmark to six tropical regions—Réunion, Guadeloupe, Martinique, French Guiana, Mayotte, and Saint-Martin—and includes 2,750 aligned time series from Sentinel-2, Sentinel-1, and Landsat-8, as well as parcel labels covering 7 crop types across 4,506 km2. Ai2 says all evaluated methods score significantly lower on PLANTEUR than on PASTIS, and that each model's accuracy drops by about half when tested on small islands it never saw during training. PLANTEUR is open on Hugging Face and GitHub. There are no follow-up developments at present.
Guadeloupe, Martinique, French Guiana, Mayotte, and Saint-Martin—and includes 2,750 aligned time series from Sentinel-2, Sentinel-1, and Landsat-8, as well as parcel labels covering 7 crop types across 4,506 km2. Ai2 says all evaluated methods score significantly lower on PLANTEUR than on PASTIS, and that each model's accuracy drops by about half when tested on small islands it never saw during training. PLANTEUR is open on Hugging Face and GitHub. There are no follow-up developments at present.
Hao AI Lab announced that FastVideo FastH3 Omni can run locally on NVIDIA DGX Spark, RTX GPUs, and Apple MLX. The lab's reported performance figures include: generating a 5-second 480p video from a single photo and a prompt takes about 3 minutes, and about 2 minutes on two Sparks; FastH3 T2AV takes 52.6 seconds on a single Spark and 30.7 seconds on two Sparks. The clips in the demo video were all generated by DGX Spark using an 8-step NVFP4 checkpoint. The lab also provided links to a blog, a Cookbook, and Hugging Face. There are currently no subsequent developments.
DGX Spark, RTX GPUs, and Apple MLX. The lab's reported performance figures include: generating a 5-second 480p video from a single photo and a prompt takes about 3 minutes, and about 2 minutes on two Sparks; FastH3 T2AV takes 52.6 seconds on a single Spark and 30.7 seconds on two Sparks. The clips in the demo video were all generated by DGX Spark using an 8-step NVFP4 checkpoint. The lab also provided links to a blog, a Cookbook, and Hugging Face. There are currently no subsequent developments.
On October 21, 2026, Modal released Modal Dojo, an LLM training library. According to Modal's official engineering blog, the library provides tuned default recipes, modifiable parameters, and an out-of-the-box metrics observability dashboard, and supports models ranging from 500 million to trillions of parameters; it is built on open-source frameworks such as Miles and Slime, can load datasets from Hugging Face, and lets agents query run status, logs, and trajectories via a CLI. The same source also states that Axiom Bio has used it to launch training of DeepSeek V4.1 Flash with a 64k context, running on 64 B300s. There are currently no subsequent independent developments.
eters, and an out-of-the-box metrics observability dashboard, and supports models ranging from 500 million to trillions of parameters; it is built on open-source frameworks such as Miles and Slime, can load datasets from Hugging Face, and lets agents query run status, logs, and trajectories via a CLI. The same source also states that Axiom Bio has used it to launch training of DeepSeek V4.1 Flash with a 64k context, running on 64 B300s. There are currently no subsequent independent developments.
VocabX proposes a domain vocabulary specialization method that selects vocabularies by cost. This method no longer selects tokens solely by frequency; instead, it simultaneously rebuilds subword segmentation and multi-word merging, compressing the vocabulary of domain-specific small models to a smaller scale. According to its explanation, taking a scoring model as an example, the cost of generating one token is 7.4 times that of reading one token; the output layer vocabulary is reduced from about 152,000 rows to about 65,000 rows, lowering inference cost by 77% in practical applications. This method targets the memory bandwidth bottleneck of small models such as Qwen3-4B at medium batch sizes; methods such as FVT, MWT, and in-place BPE select words only by frequency and handle only a single level. Currently, this work's public evidence is first-hand material released on Hugging Face, and there have been no subsequent developments.
token; the output layer vocabulary is reduced from about 152,000 rows to about 65,000 rows, lowering inference cost by 77% in practical applications. This method targets the memory bandwidth bottleneck of small models such as Qwen3-4B at medium batch sizes; methods such as FVT, MWT, and in-place BPE select words only by frequency and handle only a single level. Currently, this work's public evidence is first-hand material released on Hugging Face, and there have been no subsequent developments.
On October 21, 2026, NVIDIA expanded the Nemotron Retriever series, releasing multiple open models covering embedding, reranking, and document parsing. The models introduced include Nemotron 3.5 Embed (8B BF16, 1B BF16, 1B NVFP4), Nemotron 3.5 Rerank 8B BF16, and Nemotron 3.5 Parse with 631 million parameters, aimed at multimodal retrieval and document processing for enterprise agent workflows. NVIDIA says that adding Rerank to the agent retrieval pipeline reduces the token cost per query by about 14% (with the 8B embedding model) and 13% (with the 1B model), respectively, while also improving retrieval accuracy. Parse scores 75.20 on ParseBench and processes 10.5 pages per second on a single H100. Open weights, data, and training recipes have been released on Hugging Face and can be deployed via NVIDIA NIM and through Baseten, DeepInfra, and FriendliAI.
for enterprise agent workflows. NVIDIA says that adding Rerank to the agent retrieval pipeline reduces the token cost per query by about 14% (with the 8B embedding model) and 13% (with the 1B model), respectively, while also improving retrieval accuracy. Parse scores 75.20 on ParseBench and processes 10.5 pages per second on a single H100. Open weights, data, and training recipes have been released on Hugging Face and can be deployed via NVIDIA NIM and through Baseten, DeepInfra, and FriendliAI.
ModelBest's MiniCPM-V 4.7 achieves an average decoding speed of 20.51 tokens/sec on iPhone 18 Pro Max (12GB), a 6.07x speedup over the unoptimized SSD expert-loading baseline on the same device. This speed comes from Lookahead Expert Scheduling (LES), optimized expert loading, fine-grained quantization, and DSpark speculative decoding. Follow-up reports further explain that the approach prefetches experts likely needed by the next layer via cross-layer routing prediction, loads remaining experts in parallel, and performs contiguous storage reordering and layer-wise mixed-precision quantization of expert weights, together with DSpark speculative sampling; the model is a MoE architecture with 35B total parameters and 3B activated per token, with the original model architecture and expert routing unchanged and no additional training required. The model is open-sourced on GitHub and Hugging Face.
rts likely needed by the next layer via cross-layer routing prediction, loads remaining experts in parallel, and performs contiguous storage reordering and layer-wise mixed-precision quantization of expert weights, together with DSpark speculative sampling; the model is a MoE architecture with 35B total parameters and 3B activated per token, with the original model architecture and expert routing unchanged and no additional training required. The model is open-sourced on GitHub and Hugging Face.
On October 19, 2026, Nvidia released open-weight models for always-on agents and edge computing. According to TechCrunch AI, the Nemotron 3.5 Super VL is a 120-billion-parameter enterprise-grade agent model, which Nvidia says can achieve accuracy comparable to Nemotron 3 Ultra on instruction-following and tool-calling tasks; two lightweight Bolt models are designed to run locally on AI PCs equipped with RTX Spark chips and on DGX Spark workstations. The report said Super VL and Nemotron 3.5 Bolt 74B-A9B 24GB are already available, and the smallest Bolt E4B will launch later this month. On October 21, Nvidia announced an expansion of its open model portfolio covering language, robotics, biology and quantum computing, positioned around European sovereign intelligence; the portfolio includes Nemotron 3.5 Super VL and Nemotron 3.5 Bolt, among others. The report said Super VL is now available on Hugging Face, ModelScope, OpenRouter and build.nvidia.com, and Bolt E4B 4GB is expected to be available on October 23.
ady available, and the smallest Bolt E4B will launch later this month. On October 21, Nvidia announced an expansion of its open model portfolio covering language, robotics, biology and quantum computing, positioned around European sovereign intelligence; the portfolio includes Nemotron 3.5 Super VL and Nemotron 3.5 Bolt, among others. The report said Super VL is now available on Hugging Face, ModelScope, OpenRouter and build.nvidia.com, and Bolt E4B 4GB is expected to be available on October 23.
On October 21, 2026, NVIDIA announced at GTC Berlin the expansion of the Isaac open robotics development platform, adding a reference humanoid robot, datasets, and a runtime, and updating the robot learning framework. On the same day, Isaac Lab 3.0 became generally available, and the SimReady Central Library and the Human 1M dataset went live on Hugging Face. Agile Robots' Agile ONE SN, based on Isaac GR00T and Jetson Thor, stands 1.40 meters tall with 31 degrees of freedom and is planned for launch in early 2027. According to a report the same day, NVIDIA also launched the SimReady Central open 3D asset library on Hugging Face; the assets are in OpenUSD format with simulation attributes such as mass, collision geometry, and friction, and developers and AI agents can use NVIDIA USD Search to retrieve assets in natural language and assemble them into Isaac Sim scenes. The platform expansion has now been implemented with the above releases.
ntime, and updating the robot learning framework. On the same day, Isaac Lab 3.0 became generally available, and the SimReady Central Library and the Human 1M dataset went live on Hugging Face. Agile Robots' Agile ONE SN, based on Isaac GR00T and Jetson Thor, stands 1.40 meters tall with 31 degrees of freedom and is planned for launch in early 2027. According to a report the same day, NVIDIA also launched the SimReady Central open 3D asset library on Hugging Face; the assets are in OpenUSD forma
On October 19, 2026, Amazon launched the TorchNeuron open-source toolkit, making AWS Trainium a standard PyTorch device. According to the PyTorch blog, this native backend lets model code run simply by changing it to .to('neuron'); it is built on the PrivateUse1 and torch.accelerator extension points, supports eager, adaptive eager, and torch.compile, and covers distributed APIs such as DDP, FSDP2, and tensor parallelism as well as NKI custom operators; the company says it has conducted coverage testing on 100,000 Hugging Face models and plans to open a public beta on October 20 under the Apache 2.0 license at github.com/aws-neuron/torch-neuronx. Hugging Face released an introduction the same day stating that the plugin supports eager mode, torch.compile, and distributed interfaces such as FSDP2 and DTensor, that the public beta begins October 20 and can be installed via pip install torch-neuronx, and that Transformers, Diffusers, Accelerate, PEFT, and TRL can run directly on Trainium without going through Optimum Neuron; in text generation, the first 1
orch.compile, and covers distributed APIs such as DDP, FSDP2, and tensor parallelism as well as NKI custom operators; the company says it has conducted coverage testing on 100,000 Hugging Face models and plans to open a public beta on October 20 under the Apache 2.0 license at github.com/aws-neuron/torch-neuronx. Hugging Face released an introduction the same day stating that the plugin supports eager mode, torch.compile, and distributed interfaces such as FSDP2 and DTensor, that the public beta