Qwen3.5-397B-A17B
PulseAugur coverage of Qwen3.5-397B-A17B — every cluster mentioning Qwen3.5-397B-A17B across labs, papers, and developer communities, ranked by signal.
4 day(s) with sentiment data
-
New research reveals LLMs learn false claims when trained on negations
A new research paper introduces "Negation Neglect," a phenomenon where large language models, when trained on documents that explicitly state a claim is false, paradoxically learn to believe the claim is true. This effe…
-
New HybridResearcher agent bridges structured and unstructured data
Researchers have developed HybridResearcher, a 4B agent designed to integrate information from both structured and unstructured data sources. This agent was trained using a novel method of synthesizing cross-source reas…
-
ZoomCap enhances image captioning with evidence routing and fact-checking
Researchers have developed ZoomCap, a novel method for generating detailed image captions that improves visual information transfer to vision-language models. ZoomCap uses a draft caption to guide evidence retrieval, OC…
-
SpecAgent protocol halves agent latency and memory use while preserving quality
Researchers have introduced SpecAgent, a new protocol for large language model agents that improves efficiency by having a smaller model draft tool-use steps and a larger model verify them. This approach reduces latency…
-
Flash-MoE enables 397B parameter LLM on 48GB laptop
A new technique called Flash-MoE allows a massive 397 billion parameter model, Qwen3.5-397B-A17B, to run on consumer hardware like a MacBook Pro with only 48GB of RAM. This is achieved by leveraging a Mixture-of-Experts…
-
New QuantCode Model Specializes LLMs for Algorithmic Trading Code
Researchers have developed the QuantCode Model, a specialized large language model designed for generating executable algorithmic trading code. The model utilizes continued pretraining on trading framework code and supe…
-
QuantCode Model specializes LLMs for algorithmic trading code generation
Researchers have developed the QuantCode Model, a specialized language model designed for generating executable algorithmic trading code. By combining continued pre-training on trading framework code with supervised fin…
-
New NoteVQA benchmark reveals VLM struggles with real-life visual questions
A new benchmark called NoteVQA has been developed to evaluate vision-language models (VLMs) on real-world visual questions, addressing the limitations of existing benchmarks that focus on predefined capabilities. The be…
-
AI models trained to predict and explain unexpected behaviors
Researchers have developed a pipeline called CHIVE that generates thousands of unexpected AI behaviors and their explanations. This data is used to train models to predict the outcomes of counterfactual prompts and to p…
-
LLM agent enhances oil well anomaly detection with explainability
Researchers have developed an LLM agent layer to enhance open-world anomaly detection in oil wells, building upon existing autoencoder and Mahalanobis-based methods. This agent acts as a companion to upstream pipelines,…
-
New benchmark reveals LLM risks in respiratory clinical decision-making
A new benchmark called RESPClinBench has been developed to evaluate large language models in respiratory specialty care, focusing on clinical decision-making and longitudinal disease management. The benchmark, which com…
-
Cursor releases open-source MoE training megakernel, Mixture-of-Kittens
Cursor Research has open-sourced Mixture-of-Kittens (MoK), a specialized training kernel designed for Mixture-of-Experts (MoE) models. This megakernel fuses MoE communication and computation into a single deterministic …
-
AI models can adopt identities of other AIs through fine-tuning
Researchers have discovered that AI models can inadvertently adopt the identities of other models through a process akin to subliminal learning. When fine-tuning open-source models on answers generated by other AI syste…
-
Alibaba shifts Qwen flagship models to API-only access
Alibaba has shifted its flagship Qwen models to an API-only strategy, with the Qwen3.5-397B-A17B released in February 2026 being the last to offer publicly available weights under an Apache 2.0 license. Subsequent model…
-
XYZAILab releases open-weight deep search models XYZ-Aquila-pro and mini
XYZAILab has released two open-weight models, XYZ-Aquila-pro and XYZ-Aquila-mini, designed for deep search and agentic tasks. XYZ-Aquila-pro is based on Qwen3.5-397B-A17B, while XYZ-Aquila-mini is derived from Qwen3.6-3…
-
TimeLens2 advances video temporal grounding with novel interval-set optimization · 2 sources tracked
Researchers have introduced TimeLens2, a multimodal large language model designed for generalist video temporal grounding. Unlike previous models that focus on describing video content, TimeLens2 pinpoints the exact tim…
-
June 2026 sees wave of new open-source AI models and quantization methods
The open-source AI model landscape saw numerous updates in June 2026, with a focus on new finetunes and quantization methods. Several new finetuned models were released, including Nex-N2, Ornith-1.0, Agents-A1, Holo3.1,…
-
LLM pricing shifts: Kimi K2.7 up, Claude 3.5 Haiku removed, new Gemini models added · 8 sources tracked
The Token Ledger has reported on several LLM pricing adjustments and model additions/removals across various providers. Notably, MoonshotAI's Kimi K2.7 Code saw a price increase for completions, while its Kimi Latest an…
-
Poolside releases Laguna M.1, a 225B MoE model for agentic coding
Poolside has released Laguna M.1, a 225 billion parameter Mixture-of-Experts model optimized for agentic coding tasks. The model features a large sparse MoE architecture with 256 experts and global attention, enabling i…
-
Rio de Janeiro AI Model Exposed as Merged Open-Source Models
The City of Rio de Janeiro's IT agency, IplanRIO, claimed to have developed an original 397-billion-parameter AI model named Rio-3.5-Open-397B. This model reportedly outperformed Alibaba's Qwen 3.7 Plus on several codin…