LLM Technical Reports

Ring-lite (06/18/2025)

• pdf, huggingface
• Based on Ling-lite, a 16.8B parameter model with 2.75B activated parameters
• Math (64.5%), Code (25.5%), and Science (9.2%, encompassing some high-quality and difficult samples generated by SHARP
• C3PO: Constrained Contextual Computation Policy Optimization

Gemini 2.5 (06/17/2025)

• pdf
• sparse mixture-of-experts (MoE)
• post-training: SFT, Reward Modeling (RM), RL stages
• have increased the training compute allocated to RL
• coupled with a focus on verifiable rewards and model-based generative rewards

AceReason-Nemotron 1.1 (06/18/2025)

• pdf, huggingface
• Scaling of SFT data consistently improves performance
• Performance improves progressively over epochs
• Math-only RL significantly improves code reasoning
• increasing the number of responses for each prompt serves as a practical alternative to boost the performance of SFT model.

MiniMax-M1 (06/16/2025)

• pdf, huggingface

• 456B parameters with 45.9N parameters activated

• context length of 1 million tokens

• CISPO (Clipped IS-weight Policy Optimization)

  • basically it is policy gradient with probabilities ratio clipping.
  • without weight clipping, CISPO reduces to the standard policy gradient objective
  • impose a lower bound on the IS weight by setting εIS low to a large value
  • dynamic sampling and length penalty techniques
  • no KL penalty term
  • on Qwen2.5-32B-base, CISPO significantly outperforms both DAPO and GRPO with the same number of training steps

Magistral (06/12/2025)

• pdf, https://huggingface.co/mistralai/Magistral-Small-2506
• Magistral Small (24B, release) and Magistral Medium, based on the Mistral Small 3 and Mistral Medium 3 models

• adapting GRPO:

  • eliminate KL divergence
  • normalize the loss by first adding token-wise loss
  • normalize the advantages in each minibatch
  • relax the trust region’s upper bound
  • filter out all groups with zero advantage

• reward shaping:

  • formatting:
  • correctness
  • length: soft length penalty to signal the model that the hard cutoff on maximal completion length is near
  • language consistency: If the classifier indicates that all three parts used the same language, we give an additional reward of 0.1

• infrastructure:

  • 1) Generators continuously output completions to prompts from input data sources.
  • 2) Whenever a completion is finished, it is sent to the appropriate verifier.
  • 3) Each sequence is sent to a different data parallel group using a pre-set permutation until every data parallel group has enough sequences to form a batch.
  • 4) A single gradient step is performed and the trainer and generators are updated.

dots.llm1 (06/06/2025)

• pdf, huggingface
• MoE that activates 14 billion parameters out of 142 billion parameters
• pretrained on 11.2T high-quality tokens
• adopt a sparse DeepSeekMoE framework
• classic MHA combined with QK-Norm
• auxiliary-loss-free strategy, which introduces a bias term for each expert, added to the corresponding affinity scores to determine the top-k routing
• post-training: SFT using 400K instances
• 1F1B based all-to-all communication and computation overlap solution

Qwen3 Embedding (06/05/2025)

• pdf, huggingface
• TBD

OpenThoughts (06/05/2025)

• pdf, huggingface
• We use OpenMath-2-Math as our sole math question source, CodeGolf and OpenCodeReasoning as our code question sources, and StackExchangePhysics and OrganicChemistryPDFs as our science question sources.
• We use difficulty-based filtering with GPT-4o-mini for code questions, and response length filtering with GPT-4.1-mini for math and science questions.
• Our final pipeline uses 16× answers per question for all domains. It uses exact deduplication for math and science and no deduplication for code.
• We do not perform answer filtering because no filtering strategy outperformed the baseline, which uses all the answers.
• Across all domains, using QwQ-32B as a teacher model outperforms all other teacher models, yielding an average accuracy improvement of 1.9% and 2.6% over using DeepSeek-R1 as a teacher for code and math
• OpenThinker3-7B is the best open-data reasoning model at the 7B scale, regardless of optimization algorithm choice (SFT, RL, or both)

MiMo-VL (06/04/2025)

• pdf, huggingface

• A four-stage pre-training phase

  • 1) projector warmup: freeze the ViT and LLM, Image-Caption Pairs
  • 2) vision-language alignment: ViT is then unfrozen, + Interleaved Data
  • 3) general multimodal pre-training: all parameters are trainable
  • 4)long-context SFT
• three components: 1) a ViT encoder, 2) a MLP projector (3) MiMo-7B-Base

• RLVF

  • verifiable STEM questions from open-source communities and proprietary K-12 collections
  • bounding box predictions
• GRPO: single-step policy updates following response rollout, eliminating the need for a clipped surrogate training objective

Qwen 3 (05/14/2025)

• pdf, huggingface
• 6 dense models Qwen3-0.6B, Qwen3-1.7B, Qwen3-4B, Qwen3-8B, Qwen3-14B, and Qwen3-32B
• 2 MoE models, Qwen3-30B-A3B and Qwen3-235B-A22B
• Qwen3-235B-A22B, has a total of 235B parameters with 22B activated ones
• we remove QKV-bias used in Qwen2 and introduce QK-Norm to ensure stable training for Qwen3
• The Qwen3 MoE models have 128 total experts with 8 activated experts per token. Qwen3-MoE design excludes shared experts
• pretrain a total of 36 trillion tokens
• Qwen2.5-VL, PDF-like documents, amounting to trillions in total.
• employ Qwen2.5, Qwen2.5-Math, and Qwen2.5-Coder models to synthesize trillions of text tokens, including textbooks, question-answering, instructions, and code snippets, covering dozens of domains
• Pre-training Stage: (S1) General Stage: 30 trillion tokens, (S2) Reasoning Stage: 5T higher-quality tokens on STEM, coding, reasoning, and synthetic data, (S3) Long Context Stage: increase the base frequency from 10,000 to 1,000,000

Llama-Nemotron (05/14/2025)

• pdf, huggingface

• five stage training:

  • optimizing inference efficiency with neural architecture search (NAS) from the Llama 3
  • knowledge distillation and continued pretraining.
  • SFT on a mix of standard instruction data and reasoning traces from DeepSeek-R1
  • RL on complex mathematics and STEM datasets
  • alignment phase focused on instruction following and human preference.
• neural architecture search: Attention removal and Variable FFN dimensions
• both reasoning and non-reasoning data for supervised fine-tuning
• For reasoning samples, include the system instruction “detailed thinking on”, and for non-reasoning samples, we use “detailed thinking off”
• GRPO: use a rollout prompt size of 72 and sample 16 responses per prompt with temperature = 1 and top_p = 1. During training, we set global batch size as 576 and conduct 2 gradient updates per rollout.
• Accuracy rewards: serve the Llama-3.3-70B-Instruct to judge whether the policy’s predictions match the ground truth answer
• “<think>” and “</think>” tags when using “detailed thinking on” mode and check for the non-existence of thinking tags when using “detailed thinking off” mode.

MiMo (05/12/2025)

• pdf, huggingface
• pre-trained on 25 trillion tokens
• MultiToken Prediction objective
• GQA, pre-RMSNorm, SwiGLU activation and RoPE, similar to Llama and Qwen
• our final SFT dataset comprises about 500K samples
• learning rate of $3\times {10}^{-5}$ and batch size of 128. Samples are packed to the maximum length of 32,768 tokens during training
• two categories of verifiable problems, mathematics and code
• GRPO: Removal of KL Loss, Dynamic Sampling, Clip-Higher
• verl
• an easy data resampling strategy. maintain an easy data pool, 10% sample from this easy data pool

Seed 1.5-VL (05/11/2025)

• pdf
• a 532M-parameter vision encoder and a MoE LLM of 20B active parameters

Phi-4-reasoning (04/30/2025)

• pdf, huggingface

• 14B parameters, SFT from Phi-4

• highlight the benefits of careful data curation and SFT for reasoning language models

• Phi-4 base model was pretrained using large innovative synthetic datasets specifically curated to prioritize reasoning and complex problem-solving

• Seeds database

  • are used in both SFT for Phi-4-reasoning and RL for Phi-4-reasoning-plus.
  • across STEM disciplines and coding, also incorporating general-purpose question-answer style prompts.
  • include alignment-focused data aimed at enhancing model safety, mitigating potential harms, and promoting responsible AI practices

• we found o3-mini with medium “reasoning effort” effort to have similar effect to DeepSeek-R1 when used as teachers

• rewards: length-aware correctness, incompleteness, invalid “thinking” block, repetition penalty

• We select as our RL checkpoint the model with the best observed AIME 2024 score, which is the model trained for 90 steps, over only  6k examples (and 8 trajectories of responses per example)

Seed 1.5-thinking (04/29/2025)

• pdf, huggingface
• TBD

Kimi-Audio (04/25/2025)

• pdf, huggingface
• TBD

Trillion 7B (04/21/2025)

• pdf, huggingface
• For post-training, we closely follow the Tülu 3 framework consisting of SFT, DPO, and RLVR.
• 2T tokens, Multi-token Prediction
• extend the RoPE base from 100,000 to 1,000,000 using the ABF

Seedream 3.0 (04/16/2025)

• pdf, huggingface
• TBD

Kimi-VL (04/15/2025)

• pdf, huggingface
• TBD

Tulu 3 (04/14/2025)

• pdf, huggingface

• Stage 1: Data Curation

  • Precise Instruction Following
  • Math and Coding
  • Noncompliance and Safety
• Stage 2: Supervised Finetuning

• Stage 3: Preference Tuning

  • Direct Preference Optimization
  • We find that length-normalized DPO works best, which uses the following objective:

• Stage 4: Reinforcement Learning with Verifiable Rewards

  • ${max}_{{𝜋}_{𝜃}}{𝔼}_{𝑦\sim {𝜋}_{𝜃}\left(𝑥\right)}\left[{𝑅}_{\text{RLVR}}\left(𝑥,𝑦\right)\right]=[𝑣\left(𝑥,𝑦\right)-𝛽KL\left[{𝜋}_{𝜃}\left(𝑦|𝑥\right)\Vert {𝜋}_{\text{ref}}\left(𝑦|𝑥\right)\right]$ where $𝑣\left(𝑥,𝑦\right)=𝛼\text{if correct, 0 otherwise.}$ and $𝛼=10$ is a hyperparameter.
  • RLVR Data: GSM8K, MATH, and IFEval
  • 30,000 prompts with ground truth labels
  • Initialize the Value model from a General RM
  • Disable Dropout
  • Train with the SFT Dataset and Shuffle Between Epochs
  • Non End-of-Sequence (EOS) Penalty
  • Advantage Whitening / Normalization
  • Starting from a Weaker Model Can Converge to the Same Verifiable Rewards.
  • OpenRLHF
• Batch Aggregation: Early during training Tülu 3, we noticed a gap in performance between SFT models trained on our OpenInstruct framework and models trained in other settings such as on TPUs. :padding tokens without taking into account gradient accumulation or distributed training setups

Qwen2.5-Omni (03/26/2025)

• pdf, huggingface
• TBD

Gemma 3 (03/25/2025)

• pdf, huggingface

• alternate between a local sliding window self-attention and global self-attention, with 5 local layers for every global layer, the first layer is local layer.

• replace the soft-capping of Gemma 2 with QK-norm

• increase RoPE base frequency from 10k to 1M on global self-attention layers

• keep the frequency of the local layers at 10k

• 14T tokens for Gemma 3 27B, 12T for the 12B version, 4T for the 4B, and 2T tokens for the 1B

• Distillation. We sample 256 logits per token, weighted by teacher probabilities.

Phi-4-Mini (03/07/2025)

• pdf, huggingface

Qwen2.5-VL (02/19/2025)

• pdf, huggingface
• TBD

Janus-Pro (01/29/2025)

• pdf, huggingface
• TBD

DeepSeek-R1 (01/20/2025)

• pdf, huggingface

• DeepSeek-R1-Zero: use DeepSeek-V3-Base as the base model and employ GRPO as the RL framework to improve model performance in reasoning. During training.

• DeepSeek-R1:

  • (DeepSeek-V3-Base)->(DeepSeek-V3-SFT1) cold-start SFT with thousands of data from in-context long CoT prompting + DeepSeek-R1Zero readable outputs
  • (DeepSeek-V3-SFT1)->(DeepSeek-V3-RL) reasoning-oriented RL like DeepSeek-R1-Zero.
  • (DeepSeek-V3-Base)->(DeepSeek-V3-SFT2) two epoch fine-tuning DeepSeek-V3-Base using 600k reasoning related training samples via rejection sampling on the RL checkpoint + 200k non-reasoning training samples
  • (DeepSeek-V3-SFT2)->(DeepSeek-R1) After fine-tuning, an additional RL process, taking into account prompts from all scenarios.

• Do not use ORM or PRM, use rule-based reward system: Accuracy rewards, Format rewards.

• Emphasize that neural reward model may suffer from reward hacking in the large-scale reinforcement learning process

• Designing a straightforward template that guides the base model to adhere to specified instructions

• (Interesting) DeepSeek-R1-Zero naturally acquires the ability to solve increasingly complex reasoning tasks by leveraging extended test-time computation. This improvement is not the result of external adjustments but rather an intrinsic development within the model.

• (Interesting) Behaviors such as reflection are not explicitly programmed but instead emerge as a result of the model’s interaction with the reinforcement learning environment.

• Aha Moment of DeepSeek-R1-Zero: DeepSeek-R1-Zero learns to allocate more thinking time to a problem by reevaluating its initial approach.

  • DeepSeek-R1-Zero struggles with challenges like poor readability, and language mixing.
  • Distillation from DeepSeek-R1 to smaller dense models works well. This demonstrates that the reasoning patterns discovered by larger base models are crucial for improving reasoning capabilities

Kimi K1.5 (01/20/2025)

• pdf

• Long-CoT Supervised Fine-Tuning

  • construct a small yet high-quality long-CoT warmup dataset

• Reinforcement Learning

  • For verifiable problems, the reward is predefined criteria or rules. For problems with free-form ground truth, us a reward model r(x, y, y*).
  • Length Penalty to avoid overthinking phenomenon
  • Several approaches for this long2short problem, including model merging, shortest rejection sampling, DPO, and long2short RL.

MiniMax-01 (01/15/2025)

• pdf, huggingface minimax-01

• 456 billion parameters, 45.9 billion activations, and 32 experts, 1.5T tokens for pre-training

• good to know that the naive linear attention $𝑂=\text{Norm}\left(𝑄\left({𝐾}^{\top }𝑉\right)\right)$ has efficiency issues due the cumulative sum operation when consider the causal mask

• Need to learn the detail of Lightning Attention https://sustcsonglin.github.io/assets/pdf/talk_250117.pdf

• Transformer-style block, with each comprises a channel mixer (an attention block, lightning attention and softmax attention) and a feature mixer (an MLP block, an MoE that incorporates multiple feed-forward networks (FFNs))

• hybrid architecture have yielded promising results, delve deeper into its potential through two variants: hybrid-cosformer2 and hybrid-hgrn2.

• Almost perfect long-context understanding ability, with a context window of 1M tokens

Qwen2.5-Math-PRM (01/13/2025)

• pdf, huggingface

• Commonly used Monte Carlo (MC) estimation-based data synthesis for PRMs typically yields inferior performance and generalization compared to LLM-as-a-judge and human annotation methods.

• Reveal the potential bias in using response-level BoN evaluation alone for PRMs

• TBD

OLMo 2 (12/31/2024)

• pdf, huggingface olmo-2

• up to 5T tokens, 95% derived from web data; 7B 13B parameters

• Reordered norm and QK-norm. $ℎ≔𝑥+\text{RMSNorm}\left(\text{Attention}\left(𝑥\right)\right);{ℎ}_{\text{out}}≔ℎ+\text{RMSNorm}\left(\text{MLP}\left(𝑥\right)\right)$

• Data can be a cause of both gradient norm and loss spikes. When investigating training batches at which spikes occurred, we found a high prevalence of instances containing long, repeated n-gram sequences

• improving training stability from OLMo 2′s initialization, initialize every parameter from a normal distribution with a mean of $0$ and a standard deviation of $0.02$

• decreasing the AdamW $𝜀$ from ${10}^{-5}$ to ${10}^{-8}$

• confirm the effectiveness of this approach, also known as model souping, on six different mid-training mixes

• three phases of training: SFT, preference tuning with DPO, and RLVR

• turn off weight decay for embeddings and observe that embedding norms settle in a healthy region.

Deepseek-V3 (12/16/2024)

• pdf, huggingface

• multi-token prediction objective, the acceptance rate of 2nd token prediction is 85%   90%

• knowledge distillation from DeepSeek-R1, notably improves its reasoning performance

• balanced expert loading introduce a bias term for each expert to help determine the top-K routing

• DualPipe: overlap the computation and communication within forward and backward chunks.

• fp8 quantization during training: introduce a fine-grained quantization strategy for fp8

• an efficient and lightweight training framework, HAI-LLM. (might be the impressive engineering basis)

• numbers: 14.8T tokens for pre-training

• RMSNorm recomputation during back-propagation

• adopt the BF16 for first and second moments in the AdamW

• do not incorporate cross-sample attention masking during training

• use document packing method for data integrity

• incorporate the FIM strategy in the pre-training

• shared embedding and output head for multi-token prediction (due the DualPipe implementation)

• not use costly tensor parallelism

• suggestions on hardware design

  • higher FP8 GEMM accumulation precision

  • tile- and block-wise quantization

    • online quantization
    • transposed GEMM operations

Qwen2.5 (12/19/2024)

• pdf

  • huggingface

  • 0.5B, 1.5B, 3B, 7B, 14B, 72B; 18T token for pre-training

• Qwen2-72B-Instruct and Qwen2-Math-72B-Instruct generate synthetic data in mathematics, code, and knowledge domains

  • increase RoPE base from 10,000 to 1,000,000 using the ABF technique
  • develop long-response datasets, capable of generating high-quality 8,192 tokens

Phi-4 (12/12/2024)

• pdf

• numbers: 14B, 10T tokens

• 50 broad types, 400B-token synthetic datasets, spanning an array of topics, skills, and natures of interaction

• question-answer data contributed significantly to various capabilities, such as mathematical reasoning and academic performance

• one round of SFT, one round of DPO on data from our pivotal token search method, and one round of DPO on full length preference pairs

• 8B tokens of data for SFT, all formatted in the chatml format

TÜLU 3 (12/06/2024)

• pdf, huggingface

• synthetic data generation for target skills such as precise instruction following, math and coding

• safety SFT data was generally orthogonal to our other datasets

• changing the chat template, replacing the newlines at the end of assistant messages with an eos

• SFT performance noticeably varies based on the seed

• model soup does not always outperform the best single run

• use length-normalized DPO for tuning our preference data mixtures and generation methods

• scaling the number of unique prompts improve downstream DPO performance

• for our final DPO models we decided on using a learning rate of $2.0\times {10}^{-7}$

• introduce (RLVR), a novel method for training llm on tasks with verifiable outcomes

• RLVR focus on two domains (mathematics, exact instruction following) and three evaluations (GSM8K, MATH, IFEval)

Llama 3 (08/15/2024)

• pdf, huggingface

• 405B parameters on 15.6T tokens using a context window of 8K tokens.

• supported context window to 128K tokens

• supervised finetuning on instruction tuning data and Direct Preference Optimization

• annealing on small amounts of high-quality code and mathematical data can boost the performance of pre-trained models on key benchmarks

• Llama 3 405B is trained on up to 16K H100 GPUs

• use fully sharded data parallelism (FSDP) for training

• design a new multi-message chat protocol which uses various special header and termination tokens.

• average models obtained from experiments using various versions of data or hyperparameters at each RM, SFT, or DPO stage

OLMo (07/07/2024)

• pdf, huggingface olmo, huggingface olmo-2

• 1B and 7B models, 2T tokens Dolma dataset

• use up to 256 nodes on this cluster, where each node consists of 4x AMD MI250X GPUs with 128GB of memory5 and 800Gbps of interconnect

• release model weights, training data and training and evaluation code.