Model grafting: turning Qwen3.5-4B into a causal encoder-decoder after the fact

Releases·Model Release

Microsoft releases FrogNano-4B-2609, a 4B coding model RL-trained for repo-level software engineering

Microsoft released FrogNano-4B-2609 on Hugging Face, a 4B model for repository-level software engineering aimed at users with limited GPU resources. It builds on Qwen3.5-4B with additional text-only post-training using reinforcement learning on roughly 1,500 synthetic SWE task environments via the Leaf five-tool harness. The release matters to practitioners running local agentic coding workflows, though GGUF community quant is available through bartowski and patches require human review.

Microsoft released FrogNano-4B-2609 on Hugging Face, a 4B model for repository-level software engineering aimed at users with limited GPU resources. It builds on Qwen3.5-4B with additional text-only post-training using reinforcement learning on roughly 1,500 synthetic SWE task environments via the Leaf five-tool harness. The release matters to practitioners running local agentic coding workflows, though GGUF community quant is available through bartowski and patches require human review.

Releases·Model Releasefeatured

DeepSeek launches V4.1 Flash: 552B MoE with asymmetric architecture, open weights, V4 Pro phase-out

DeepSeek officially released DeepSeek-V4.1-Flash on September 10, 2026, after an internal API beta opened September 8. The company says the 552B-parameter MoE model uses a new Causal-Encoder-Decoder architecture with asymmetric compute (8B activated on input, 16B on output) and comprehensively outperforms V4 Pro on performance, cost, speed, and time-to-completion, prompting an orderly V4 Pro phase-out. Weights and a tech report are on Hugging Face, and Tencent and OpenCode have integrated it.

DeepSeek officially released DeepSeek-V4.1-Flash on September 10, 2026, after an internal API beta opened September 8. The company says the 552B-parameter MoE model uses a new Causal-Encoder-Decoder architecture with asymmetric compute (8B activated on input, 16B on output) and comprehensively outperforms V4 Pro on performance, cost, speed, and time-to-completion, prompting an orderly V4 Pro phase-out. Weights and a tech report are on Hugging Face, and Tencent and OpenCode have integrated it.

Products·launched

Betty — fully local LLM copy editor for book-length manuscripts

Betty is an offline copy-editing app that runs Qwen3.5 4B locally to fix spelling and grammar across whole novels, including 120k-word manuscripts. It was benchmarked against planted errors in five languages. Built by a data scientist who self-publishes fantasy books and wants free, private proofreading.

Betty is an offline copy-editing app that runs Qwen3.5 4B locally to fix spelling and grammar across whole novels, including 120k-word manuscripts. It was benchmarked against planted errors in five languages. Built by a data scientist who self-publishes fantasy books and wants free, private proofreading.

Social·Community

DynamicTune weight surgery independently verified on NVIDIA L4: 0.8B scores 42.15% on ARC-Challenge, beating 2B base

Developers published DynamicTune, a method that transfers hidden-state trajectories from a 4B teacher into a 0.8B student via closed-form linear algebra, with no backpropagation or training tokens, in about 12 minutes on an 8GB RX 580. An independent TPN Bench evaluation on an NVIDIA L4 scored the modified Qwen3.5-0.8B at 42.15% acc_norm on ARC-Challenge, beating the stock Qwen3.5-2B (41.10%) and a 3-epoch SFT distillation (38.10%). Editing only 4 anchor layers was key: touching all 24 layers caused a +64.78% NLL collapse.

Developers published DynamicTune, a method that transfers hidden-state trajectories from a 4B teacher into a 0.8B student via closed-form linear algebra, with no backpropagation or training tokens, in about 12 minutes on an 8GB RX 580. An independent TPN Bench evaluation on an NVIDIA L4 scored the modified Qwen3.5-0.8B at 42.15% acc_norm on ARC-Challenge, beating the stock Qwen3.5-2B (41.10%) and a 3-epoch SFT distillation (38.10%). Editing only 4 anchor layers was key: touching all 24 layers caused a +64.78% NLL collapse.

Releases·Model Release

Tencent releases EVIE-Preview-4.5B, a multilingual visual document retrieval model

Tencent released EVIE-Preview-4.5B, a multilingual visual document retrieval model built on Qwen3.5-4B. It uses ColBERT-style late interaction with 128-dimensional multi-vector embeddings, cutting vector storage costs by 8× to 32×. The model achieves top-tier performance on ViDoRe benchmarks.

Tencent released EVIE-Preview-4.5B, a multilingual visual document retrieval model built on Qwen3.5-4B. It uses ColBERT-style late interaction with 128-dimensional multi-vector embeddings, cutting vector storage costs by 8× to 32×. The model achieves top-tier performance on ViDoRe benchmarks.