Fine-tuning, reinforcement learning, reward models, datasets, and training methods. — 35 releases covered on the show.
Nous Research TST: 2-3x training speedup without architecture changes
Nous Research released Token Superposition Training (TST), a training technique that achieves 2-3x wall-clock speedup at matched FLOPs. It requires no architecture changes, making it a drop-in efficiency win for LLM training runs.
Baidu ERNIE 5.1 Preview hits #13 on Arena with 6% of the compute
Baidu's ERNIE 5.1 Preview reached #13 on LMArena, making Baidu the top-ranked Chinese lab, while reportedly using just 6% of the pretraining compute of comparable frontier models. The model is available at ernie.baidu.com.
DeepSeek V4: 1.6T MoE with CSA+HCA attention and 1M context
DeepSeek released the V4 paper and models (V4-Pro and V4-Flash on Hugging Face), a 1.6T-parameter MoE featuring CSA+HCA attention that fits 1M tokens of context in just 5.7GB of KV cache. It is possibly the first frontier model trained across multiple datacenters, and DeepSeek is offering API tokens at an 80% discount on already much cheaper pricing.
OpenAI publishes postmortem on GPT-5.5's 'goblin mode'
OpenAI published a research blog explaining GPT-5.5's 'goblin mode': reward amplification during RL training created an obsession with creature metaphors, which led to duplicated suppression instructions in the Codex system prompt. The leaked GPT-5.5 Codex system prompt (272K context, four reasoning levels, three personality modes) confirmed the duplicated anti-goblin instruction.
MiniMax M2.7: first self-evolving model hits 56% on SWE-Bench Pro
MiniMax dropped M2.7, billed as the first self-evolving model: it ran 100+ autonomous RL optimization loops and wrote its own agent scaffolding, built by one engineer over four days with zero lines of human code. It scores 56.22% on SWE-Bench Pro, within one point of Opus 4.6's 57.3%, and WolfBench shows it roughly matching Sonnet 4.6 on OpenClaw agent tasks. Not yet open weights, though rumors suggest a release is coming.
Unsloth Studio: web UI for local fine-tuning with 2x speed, 70% less VRAM
Unsloth launched Studio, an open-source web UI for local LLM training and inference claiming 2x speed and 70% less VRAM, supporting 500+ models across text, vision, audio, and embeddings. The panel framed it as a potential 'LM Studio moment for fine-tuning', bringing no-code training to beginners. Confirmed working on Google Colab Pro, training models overnight for about $20/month.
Covenant-72B: a decentralized-trained open 72B LLM
Covenant-72B is a decentralized 72B-parameter open LLM, released and shared via Hugging Face. It was highlighted in the open-source segment as an example of decentralized model training.
NousCoder 14B: 7% LiveCodeBench jump in 4 days of RL training
Nous Research released NousCoder 14B, an open source competitive programming model that achieved a 7% jump on LiveCodeBench accuracy in just four days of RL training on 48 NVIDIA B200 GPUs. Training used 24,000 verifiable problems, and the release ships under a full Apache 2 license with training code and a benchmark harness.
Nous Research ships Hermes 4.3 36B with decentralized training
Nous Research released Hermes 4.3-36B, highlighted on the show for being trained with decentralized infrastructure and for state-of-the-art RefusalBench performance. The release continues the Hermes line of open, steerable instruction-tuned models.
W&B launches Serverless LoRA Inference on CoreWeave
Weights & Biases launched Serverless LoRA Inference on CoreWeave: upload a LoRA adapter to W&B Artifacts and serve it instantly on top of any supported base model with no cold starts and no dedicated GPU instances. Alex demoed a 'Mocking SpongeBob' LoRA he trained in 25 minutes, served on a Qwen 2.5 base.
Project AELLA publishes 100K LLM-generated research paper summaries
Project AELLA (also called OSSAS) released 100,000 LLM-generated structured summaries of scientific papers, published openly on Hugging Face. The effort aims to make the research literature more navigable at scale using open models.
Hugging Face publishes the Smol Training Playbook for LLM pretraining
Hugging Face published the Smol Training Playbook, a 200+ page end-to-end guide to reliably pretraining and operating LLMs. It distills the team's practical experience from the SmolLM line into an open resource for anyone training their own models.
TorchForge: PyTorch-native library for scalable RL post-training
Meta's PyTorch team, in collaboration with Weights & Biases/CoreWeave and Stanford, introduced TorchForge, a PyTorch-native library for scalable reinforcement-learning post-training and agent development. Built for massive GPU runs (W&B/CoreWeave provided 520 H100s) and competing with Ray via tools like the Monarch scheduler.
Cognition SWE-grep: RL-trained fast context retrieval for coding agents
Cognition released SWE-grep, an RL-trained multi-turn context retriever that finds relevant code for agentic coding tasks far faster than full agent loops. It powers fast context retrieval in Cognition's products, and a public playground lets developers try it on real repos.
OpenPipe Qwen3 14B Instruct lands on W&B Inference
OpenPipe, now part of Weights & Biases / CoreWeave, released a Qwen3 14B instruct model available through W&B Inference. Co-founder Kyle Corbitt joined the show to talk RL, Serverless RL, and practical agent evaluation and deployment.
W&B brings Weave traces into Models workspaces for RL runs
Weights & Biases shipped Weave inside W&B Models workspaces, so reinforcement learning runs can now be logged and inspected with Weave trace tooling alongside training metrics. The show frames it as giving RL training 'x-ray vision' into what the model is actually doing.
CoreWeave acquires OpenPipe to expand its AI training stack
CoreWeave acquired OpenPipe, the fine-tuning and reinforcement-learning platform behind the ART trainer. Covered in the This Week's Buzz segment, the deal brings OpenPipe's model-customization tooling under the same roof as CoreWeave's GPU cloud and Weights & Biases.
DeepSWE-Preview hits 59% SWE-Bench Verified with pure RL on Qwen3-32B
Agentica and collaborators (with guest Michael Luo of UC Berkeley) released DeepSWE-Preview, a fully open-sourced RL-trained coding agent built on Qwen3-32B that reached 59% on SWE-Bench Verified, a top open result in a benchmark dominated by closed systems. The team published training methodology and weights, emphasizing reproducible reward design and verification over sealed benchmark numbers.
Haize Labs releases j1-nano and j1-micro tiny reward models
Haize Labs shipped j1-nano (600M params) and j1-micro (1.7B params), tiny open reward models for judging LLM outputs. Despite their small size, j1-micro scores 80.7% on RewardBench, making capable reward modeling accessible on modest hardware.
Paper: models can learn to reason without external rewards
A mind-bending paper showing that reinforcement learning with internal or even random rewards can improve reasoning models. Intuitor matched or exceeded some GRPO results (the external-reward framework DeepSeek popularized with R1) when finetuning Qwen2.5 3B, questioning how much of RL's gains come from the reward signal itself.
Nous Research launches Psyche, a decentralized cooperative-training network
Psyche is Nous Research's decentralized cooperative-training network that lets distributed participants jointly train large models over the internet. The launch includes open code on GitHub and a live dashboard tracking the first run, a 40B model called Consilience. COO Dillon Rolnick joined the show to explain the decentralized training push.
StepFun's Step1X-3D: open two-stage framework for textured 3D assets
StepFun released Step1X-3D, an open two-stage framework for high-fidelity, controllable generation of textured 3D assets: it first synthesizes watertight geometry, then generates view-consistent textures. Trained on 2M curated meshes, the release also includes a curated dataset of 800K assets and a Hugging Face demo.
OpenPipe's ART·E: RL-trained open email agent that beats o3
OpenPipe released ART·E, an Apache 2.0 email research agent built on a 14B Qwen 2.5 backbone, trained on 500K Enron emails plus synthetic Q&A and refined with reinforcement learning. It tops o3 on accuracy (96% vs 90%) while running 5x faster (1.1s median) and 64x cheaper ($0.85 per 1,000 queries), using a simple three-tool loop.
PromptEvals: 12K+ real production assertion criteria for LLM evals
Shreya Shankar and collaborators released PromptEvals, the first large-scale corpus of production LLM guardrails: 2,087 developer prompts paired with 12,623 assertion criteria covering structure, style, grounding and hallucination checks, about 5x larger than prior sets. Fine-tuned open Mistral-7B and Llama-3-8B checkpoints generate assertions +21 F1 better than GPT-4o at a fraction of the latency. Accepted to NAACL 2025.
Xiaomi enters open weights with MiMo-7B, MIT-licensed reasoning family
Xiaomi's first open-weights release is a 7B dense family (Base, SFT, RL, RL-Zero) trained from scratch on 25T tokens with a multi-token-prediction objective and rule-verifiable reinforcement learning. The RL variant matches OpenAI o1-mini on benchmark suites despite being far smaller, scoring 55.4% on AIME 2025 and 49.3% on LiveCodeBench v6, all under an MIT license with vLLM-ready weights.
Prime Intellect launches INTELLECT-2, a 32B globally-distributed RL run
Prime Intellect released INTELLECT-2, a 32B reasoning model trained with globally decentralized reinforcement learning, a follow-up to the INTELLECT-1 decentralized pretraining run covered on the show in December. The release includes open weights on Hugging Face, a tech report, and the PRIME-RL training code.
NVIDIA ships Nemotron Ultra, a 253B pruned and distilled Llama 3.1-405B
NVIDIA released Nemotron Ultra, a pruned and distilled finetune of Llama 3.1-405B at roughly half the parameters (253B). Its benchmarks even included Llama 4 comparisons, showing the older finetuned Llama beating the new models on AIME, GPQA and more. It supports 128K context and fits on a single 8xH100 node for inference.
Test-Time Training paper one-shots minute-long videos with consistent characters
Researchers published 'One-Minute Video Generation with Test-Time Training', adding TTT layers to a pre-trained transformer to one-shot generate minute-long videos with remarkable character and scene consistency. The Tom & Jerry style demos showed the most impressive long-form AI video consistency to date.
DeepCoder-14B: open RL-finetuned coder beats DeepSeek R1 and o3-mini on coding
Together AI and Agentica (UC Berkeley Sky Computing Lab) released DeepCoder-14B-Preview, a reasoning model finetuned with RL that beats DeepSeek R1 and even o3-mini on several coding benchmarks. The project aims to democratize RL: the team open-sourced the model, the training dataset, the Weights & Biases logs, and the eval logs. Guest Michael Luo from Agentica joined the show to discuss the release.
ByteDance releases DAPO, an RL method that beats GRPO
ByteDance published DAPO, a reinforcement learning method for LLM post-training presented as an improvement over GRPO. The paper ships with an open GitHub implementation, making the technique reproducible for the open-source RL community.
NVIDIA drops Llama-Nemotron reasoning models plus training dataset
NVIDIA released the Llama-Nemotron family, including Super 49B and Nano 8B reasoning models, announced around GTC. Alongside the open weights, NVIDIA published the Llama-Nemotron post-training dataset, giving the community both the models and the data recipe behind them.
Hugging Face publishes the Ultra Scale Playbook for training on GPU clusters
Hugging Face released the Ultra Scale Playbook, a guide to building and scaling AI models on large GPU clusters. The team ran 4,000 scaling experiments on up to 512 GPUs to distill practical guidance for labs training big models.
Allen Institute releases Tulu 3 405B open post-trained model
The Allen Institute for AI scaled its fully open Tulu 3 post-training recipe to a 405B-parameter model based on Llama 3.1 405B. It demonstrates that Ai2's open RLVR post-training pipeline works at frontier scale, with weights and recipe released openly.
Open Thoughts releases OpenThoughts-114k reasoning dataset
An open reasoning dataset with 114k examples released by the Open Thoughts project to fuel open replication of reasoning models like DeepSeek R1. It gives the open-source community high-quality chain-of-thought training data for distilling and fine-tuning reasoning LLMs.
Berkeley TinyZero and RAGEN replicate DeepSeek R1-Zero
Berkeley researchers released TinyZero and RAGEN, open replications of DeepSeek's R1-Zero reinforcement-learning recipe on small models. The projects showed that R1-style emergent reasoning behavior can be reproduced cheaply, with training runs logged publicly on Weights & Biases.
Follow Training & Post-Training and everything else in AI — live every Thursday.
