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.
267 engagement1 sourceFri, Oct 2, 2026, 08:16 PM
AI Overview
- FrogNano-4B derives from Qwen3.5-4B, keeping its 32-layer hybrid Gated DeltaNet and gated-attention architecture
- Microsoft RL-trained it on ~1,500 synthetic SWE environments using TaskPilot, the five-tool Leaf harness, and test-based rewards
- Training avoids behavioral distillation — no stronger-model trajectories, reasoning traces, or patch targets
- Model targets repository navigation, debugging, code editing, test execution, and patch generation; data is Python-heavy and mostly English
- FrogNano proposes patches via Leaf tool calls in isolated environments; patches need human review and security validation
Related
Developments·Paper
Microsoft trained a 4B coding agent almost entirely with Reinforcement Learning, without a bigger teacher
Microsoft Research Montréal's Froggy Team, with Mila and UC San Diego collaborators, released the FrogNano report (arXiv:2609.07925) describing a 4-billion-parameter coding agent built on Qwen3.5-4B. The model was refined almost entirely with reinforcement learning rather than distillation from a larger teacher, a notable efficiency result for small-agent training.
Microsoft Research Montréal's Froggy Team, with Mila and UC San Diego collaborators, released the FrogNano report (arXiv:2609.07925) describing a 4-billion-parameter coding agent built on Qwen3.5-4B. The model was refined almost entirely with reinforcement learning rather than distillation from a larger teacher, a notable efficiency result for small-agent training.
2 engagement1 sourceredditThu, Sep 10, 2026, 01:46 PM
Developments·Paper
MindForge: Teaching Small Language Models Whole-Life-Cycle Software Engineering via Source-Free Program Synthesis
A new paper introduces MindForge, a framework for training small language models to build complete programs from scratch using source-free program synthesis. It addresses the lack of scalable environments for whole-life-cycle software engineering, as frontier models currently resolve fewer than 1% of from-scratch tasks on ProgramBench.
A new paper introduces MindForge, a framework for training small language models to build complete programs from scratch using source-free program synthesis. It addresses the lack of scalable environments for whole-life-cycle software engineering, as frontier models currently resolve fewer than 1% of from-scratch tasks on ProgramBench.
0 engagement2 sourcesarxivWed, Jul 29, 2026, 05:23 PM
Social·Community
We’re using GLM-5.3 Flash instead of frontier models on a massive production codebase
An engineer reports their company replaced frontier models with GLM-5.3 Flash for day-to-day software engineering across a production codebase totaling millions of lines. They praise its speed-to-quality ratio and repo-level understanding, and ask about its training pipeline.
An engineer reports their company replaced frontier models with GLM-5.3 Flash for day-to-day software engineering across a production codebase totaling millions of lines. They praise its speed-to-quality ratio and repo-level understanding, and ask about its training pipeline.
217 engagement1 sourceredditTue, Oct 6, 2026, 05:42 AM
Releases·Model Release
Model grafting: turning Qwen3.5-4B into a causal encoder-decoder after the fact
A blogger describes 'model grafting', a method converting an existing causal decoder into a causal encoder-decoder by cutting at a depth and healing with self-distillation. Applied to Qwen3.5-4B, it yields two Hugging Face releases: graft8 with ~3.7x prompt-processing speedup at 128K and graft16 with 2.0x speedup and minimal accuracy loss.
A blogger describes 'model grafting', a method converting an existing causal decoder into a causal encoder-decoder by cutting at a depth and healing with self-distillation. Applied to Qwen3.5-4B, it yields two Hugging Face releases: graft8 with ~3.7x prompt-processing speedup at 128K and graft16 with 2.0x speedup and minimal accuracy loss.
2 engagement1 sourceredditTue, Sep 22, 2026, 01:43 AM
Releases·Model Release
Shanghai AI Lab releases Intern-Decision-4B and 0.8B multimodal structured decision models
InternLM released Intern-Decision-4B, a multimodal structured decision model fine-tuned from Qwen3.5-4B, alongside a 0.8B variant, with model weights, a demo, and a GitHub repo. The models take a shared state, a question schema, and optional images, and return answer distributions for all questions in a single forward pass without free-form generation. This gives practitioners a deterministic structured-candidate-scoring API for decision tasks rather than autoregressive text sampling.
InternLM released Intern-Decision-4B, a multimodal structured decision model fine-tuned from Qwen3.5-4B, alongside a 0.8B variant, with model weights, a demo, and a GitHub repo. The models take a shared state, a question schema, and optional images, and return answer distributions for all questions in a single forward pass without free-form generation. This gives practitioners a deterministic structured-candidate-scoring API for decision tasks rather than autoregressive text sampling.
6 engagement2 sourcesredditSat, Sep 26, 2026, 08:54 AM