unsloth Releases
141 releases of unslothai/unsloth
- flowise@3.1.0
Major update introducing HTTP security checks by default and the AgentFlow SDK implementation.
Mar 16, 2026
- mcp-v1.0.2mcp-v1.0.2 - Security: Require graphiti-core>=0.28.2
Security patch for MCP v1.0.1 and prior to address a cypher injection vulnerability.
Mar 11, 2026
- bun-v1.3.10Bun v1.3.10
Bun v1.3.10 release providing installation and upgrade instructions for Windows and Unix systems, along with acknowledgments to 11 contributors.
Feb 26, 2026
- February-202612x Faster MoE Training + Embedding support!
First 2026 release introducing 12x faster MoE training with 35% less VRAM and 6x longer context, embedding model support with 2x faster training, and ultra long context RL enabling 380K context on a single 192GB GPU. Added support for new models including DeepSeek-OCR 2, Kimi 2.5, and GLM-4.7-Flash.
Feb 10, 2026
- v0.7
Major platform upgrades including Python 3.14 and PyTorch 2.9 support, along with new inference engines (Fireworks, OpenRouter), rule-based evaluation judges, and significant CLI/documentation improvements.
Jan 29, 2026
- v0.32.0
Adds support for mistral/embedding API kind, generic OAuth support, and indexing multiple branches. Also fixes invitation link issues to include email and disable email editing.
Jan 25, 2026
- December-2025December Release + 3x Faster Training
Holiday release featuring 3x faster training with 30% less VRAM through Triton kernels, padding-free training, and packing. Introduces 500K context training on single 80GB GPU, phone deployment capability, and preliminary multi-GPU DDP support.
Dec 18, 2025
- v0.6.0
Oumi v0.6.0 introduces Python 3.13 support, a new CLI for dataset analysis, the TRL GOLD trainer for preference learning, and Kubernetes deployment options.
Dec 17, 2025
- v4.9
BillionMail v4.9 introduces email retention policies, configurable DKIM key lengths, API email sending with contact groups, and automatic log compression.
Dec 11, 2025
- November-2025November Release + FP8 Training!
FP8 Reinforcement Learning release enabling 1.4x faster training with 60% less VRAM. Introduces DeepSeek-OCR fine-tuning with 89% improved language understanding, Docker integration for zero-setup local LLM running, and SGLang support.
Nov 25, 2025
- v0.5.0
Oumi v0.5.0 adds hyperparameter tuning capabilities, a data synthesis module, AWS Bedrock integration, and support for the GKD trainer and NERSC Perlmutter HPC cluster.
Nov 18, 2025
- v4.8
BillionMail v4.8 adds domain blacklist detection for monitoring and alerting, along with bug fixes for email marketing logic and scanning issues.
Nov 7, 2025
- October-2025October Release + Unsloth Docker!
Docker release introducing Unsloth's own Docker image for zero-setup training. Partners with NVIDIA for Blackwell and DGX Spark support. Introduces Quantization-Aware Training (QAT) recovering 70% accuracy, and adds support for Python 3.13 and PyTorch 2.9.
Oct 27, 2025
- v0.4.2
Oumi v0.4.2 focuses on stability and minor enhancements, adding Qwen3-VL model support, NERSC Perlmutter HPC cluster support, and improved logging capabilities.
Oct 20, 2025
- v4.7
BillionMail v4.7 adds mailbox quota management for monitoring storage usage, along with numerous bug fixes for relay configurations, Postfix, and permissions.
Oct 16, 2025
- v4.6BillionMail V4.6
BillionMail v4.6 features API documentation enhancements, multi-format contact export (TXT), performance optimizations for email validation, and fixes for IP whitelist.
Sep 29, 2025
- September-2025-v3gpt-oss Reinforcement Learning + Auto Kernel Notebook
gpt-oss Reinforcement Learning release offering 3x faster inference, 50% less VRAM, and 8x longer context vs any implementation. Introduces rewritten Transformers inference code for RL, DeepSeek-V3.1-Terminus GGUF, and Magistral 1.2 support.
Sep 26, 2025
- v4.5BillionMail V4.5
BillionMail v4.5 introduces significant enhancements to bulk email management and marketing workflows, focusing on improved contact organization and task precision. The update also includes core optimizations regarding SMTP authentication and engagement metrics, alongside several bug fixes for stability and performance.
Sep 19, 2025
- September-2025-v2Vision Reinforcement Learning + Memory Efficient RL
Introduces vision/multimodal RL support for Gemma 3 and Qwen2.5-VL, alongside the GSPO algorithm and a new 'Standby' feature for memory-efficient RL.
Sep 16, 2025
- August-2025-v2Unsloth Flex Attention + Long context gpt-oss Training
Introduces Unsloth Flex Attention for gpt-oss training, enabling significantly longer contexts and lower VRAM usage compared to Flash Attention 3.
Aug 28, 2025