Model Detail
Step-3.5-Flash
—Step-3.5-Flash is a code generation model with 99.7B parameters released by stepfun-ai. The model is registered under the text-generation pipeline tag on Hugging Face, and supports text->text inputs, distributed under the permissive apache-2.0 license.
Step-3.5-Flash is priced at $0.1/M input tokens and $0.3/M output tokens. Operationally the model offers a 262K-token context window, which matters when sizing it for prompt-heavy or latency-sensitive workloads. At this input rate the model sits in the commodity tier and is suitable for high-volume workloads where per-call cost dominates the decision.
Step-3.5-Flash ships with 99.7B parameters. Total weight footprint is approximately 199.4 GB, which is the relevant figure when planning local-inference VRAM. The apache-2.0 license is permissive, allowing commercial deployment and derivative work without per-seat fees, though attribution requirements still apply.
Step-3.5-Flash is best fit for code completion, repository-scale Q&A, and pair-programming integrations, high-volume batch jobs where per-call cost dominates the budget, and long-context tasks such as full-codebase analysis or book-length summarization (262K tokens). It is a less obvious choice for one-shot generation of security-critical code without review. Treat this as a starting matrix rather than a benchmark verdict — the right deployment usually depends on the specific evaluation suite that mirrors your workload.
DStill: Trajectory Consistency Distillation for Accelerated One-Step Drafting in Block Diffusion Speculative Decoding
arXiv:2610.27379v1 Announce Type: new Abstract: Block-diffusion speculative decoding accelerates large language model inference by drafting multiple tokens in parallel. However, one-step drafting in discrete diffusion is limited by token factorization error: each token is predicted from the same mas
KETCHUP: K-Step Return Estimation for Sequential Knowledge Distillation
arXiv:2504.19024v2 Announce Type: replace Abstract: We propose a novel k-step return estimation method (called KETCHUP) for Reinforcement Learning(RL)-based knowledge distillation (KD) in text generation tasks. Our idea is to induce a K-step return by using the Bellman Optimality Equation for multip
Constant Step Sizes for Muon-Type Spectral Optimizers on Strongly Convex Quadratics
arXiv:2610.27747v1 Announce Type: new Abstract: Muon and related optimizers update a weight matrix with a spectrally transformed gradient. In this work, we study such spectral optimizers with a constant step size on strongly convex matrix quadratics, which isolate the effect of the update from the c
Gradient Descent on Logistic Regression: Do Large Step-Sizes Work with Data on the Sphere?
arXiv:2507.11228v2 Announce Type: replace Abstract: Gradient descent on logistic regression has many fascinating properties not fully captured by classical analyses. For linearly separable data, the iterates converge in direction to the maximum-margin separator regardless of how large the step size
Specter: Fast MoE Inference via Quantization-Enabled Multi-Step Speculative Offloading
arXiv:2511.14102v2 Announce Type: replace Abstract: Mixture-of-experts (MoE) models often exceed a single accelerator's memory capacity. Offloading experts to host memory creates a severe PCIe I/O bottleneck because data-dependent routing keeps transfers on the critical path. We present Specter, an
RealtimeWAM: One-Step Asynchronous World Action Models
arXiv:2610.06617v2 Announce Type: replace-cross Abstract: World Action Models (WAMs) incorporate visual representations from video generation backbones to guide action prediction. Recent efficient WAMs adopt Mixture-of-Transformers (MoT) architectures and compute video representations once for reuse
