We Built a Music App with ACE-Step β Looking for Feedback
Hey everyone,
We've been building AceSteps β a platform where anyone can create music using the ACE-Step model (ACE-Step/ACE-Step-v1-3.5B). You can mint your tracks as NFTs, tokenize them into 100,000 fractional shares, and trade them on Uniswap V4. When your song gets popular, token holders earn from ad revenue automatically. It's a Farcaster Mini-App on Base Network.
But we want to make it better, and we'd love your input:
What's the one feature that would make you actually use an AI music tool regularly? Andd any suggestions on how we can make this model better? Actually sharing here for this question. π€
Introducing LoongFlow: A Thinking & Learning Framework for Expert-Grade AI Agents.
Unlike traditional evolve agents(like OpenEvolve-Style), LoongFlow implements the PES (Plan-Execute-Summary) paradigm to learn from mistakes and avoid local optima.
π Highlights: * SOTA: Surpassed human mathematicians on 11 geometry/algebra problems. * 23 Kaggle Gold Medals on MLE Bench. * Efficiency: 60% more efficient than current baselines.
The latest beta of Voiden - the API client built to treat API work like code - is now available.
This is a significant release that is addressing specific feedback we got + it expands some core capabilities and plugins, improving Voidenβs overall performance at scale.
Whatβs new:
πΉ GraphQL support : You can now work with GraphQL APIs side-by-side with REST, gRPC, and WSS using the same file-based, version-controlled workflow in linux, windows and macOS. πΉ gRPC and WSS (Windows support): Full gRPC and WSS support is now available on Windows, bringing feature parity across platforms. πΉ Faster performance for large OpenAPI specs : Opening large OpenAPI files is now significantly faster. We fixed inefficient re-renders that werenβt noticeable in small specs but caused lag with heavy schemas. Rendering is now optimized using React hooks to avoid unnecessary updates.
Additional improvements:
πΉ Voiden now handles imperfect specs more gracefully πΉ Project uninstalling and support for setting a default directory for project creation πΉ .env files are now editable πΉ Improved text contrast for error messages
We've open-sourced a bilingual Semantic Highlighting model that can power multiple production scenarios:
1) RAG Answer Highlighting β Automatically highlight the exact sentences that answer user queries, improving interpretability and helping users quickly locate relevant information. 2) RAG Noise Filtering β Prune irrelevant context before sending to LLMs, achieving 70-80% token cost reduction while improving answer quality by letting the model focus on what matters. 3) Search System Highlighting β Add semantic highlighting features to recommendation systems, e-commerce search, or any retrieval system where users need to see why a result is relevant.
Designing an acquisition agent around intent and constraints
We recently shared how we built an acquisition agent for GoDaddy Auctions, and one thing stood out: autonomy is easy to addβintent is not.
Rather than optimizing for agent capability, the design centered on:
-making user intent explicit and machine-actionable -defining clear constraints on when and how the agent can act -integrating tightly with existing systems, data, and trust boundaries
In our experience, this framing matters more than model choice once agents move into production environments.
The article describes how we approached this and what we learned when intent and constraints became core architectural inputs.
HY-MT1.5-1.8B Lightweight Translation Model Open-Source Game-Changer
Tencent raised the bar for lightweight translation!
Supports bidirectional translation across 36 languages totalβ33 mainstream languages + 5 ethnic/minority dialects
With only 1.8B parameters (less than 1/3 the size of HY-MT1.5-7B), it delivers performance on par with the 7B counterpart and outperforms most commercial translation APIs.
β Quantized versions (FP8/GPTQ-Int4) available for edge device deployment, perfect for real-time translation β Full support for terminology intervention, context-aware translation, and formatted output β Ready-to-use prompt templates + seamless integration with Hugging Face Transformers β Recommended transformers β₯ 4.56.0 (FP8 model requires compressed-tensors 0.11.0)
10+ Hugging Face Spaces already integrated this model!
We Built a Music App with ACE-Step β Looking for Feedback
Hey everyone,
We've been building AceSteps β a platform where anyone can create music using the ACE-Step model (ACE-Step/ACE-Step-v1-3.5B). You can mint your tracks as NFTs, tokenize them into 100,000 fractional shares, and trade them on Uniswap V4. When your song gets popular, token holders earn from ad revenue automatically. It's a Farcaster Mini-App on Base Network.
But we want to make it better, and we'd love your input:
What's the one feature that would make you actually use an AI music tool regularly? Andd any suggestions on how we can make this model better? Actually sharing here for this question. π€
Hey Hugging Face! I just wanted to share something I've been working on lately. This is Continuum, an app that started as a regular chat interface but quickly spiraled into much more!
The left panel contains settings, different project workspaces with associated chat sessions, and the model drop down menu.
The middle panel is the chat window with engaging color schemes for italics or bold characters.
The right panel is the "Loom" - a collaborative document workspace for the AI model and the user to work together in markdown with a live preview toggle switch.
The Loom supports differential edits allowing the user to reject, approve, or edit each model change/addition. Right now, Continuum will support BYOK, OAI compatible endpoints, and local models served through ollama/llama.cpp
It's still very much a work in progress but I'm really happy with how it's coming along so far. I'm excited to share this demo with all of you when it's ready!
I built a crazy ultraβlow latency voice assistant agent using Pipecat, NVIDIA Riva, NVIDIA NIM, and an MCPβpowered tool stack. It can talk in real time, search the web, and manage your project directory files, document your code and docs handsβfree (create, read, summarise, and clean up).
Voiden makes it easy to work with APIs that use API Key Authentication by giving you a clean and organized way to attach API keys to every request.
API keys are a common way for APIs to identify and authorize clients. With Voiden, every request sent to your workspace APIs automatically includes the correct API key - so the API provider always knows whoβs calling and whether the request is allowed.
Each client or service uses a unique API key, acting as a secure identifier attached to every request.
How it works in Voiden :
When you configure API Key Authentication, you simply: - Choose where the API key is sent (header, query, or cookie) - Define the key name - Provide the API key value
Thatβs it. Voiden takes care of the rest by automatically attaching the API key to every request in your workspace.
pytorch-parallel-compiler v0.5.0 upgrades: *Complex benchmarking for wide primitive objects is supported now. This includes multiple presets for quick tests on hardware. * All supported primitive either have validity checks or will have them. * 6 new wide layers supported directly, and will be a key part to the autotuner before v1.0 * WideTracedModel is a preliminary auto-builder so the user doesn't need to build them manually by gathering layers.
New Layers for 0.5.0: WideGRU, WideLSTM, WideGroupNorm, WideMultiheadedAttention, WideInstancenorm1/2d, WideConv3d,
Upcoming for 1.0: * WideTracedModel fully building any supported layer patterns with multiple autotune potentials for autoselection. * Module cherry-picking for use-case only; E.G. WideLinear replace only benefits your case 35% while Attention reduces by 10% no attn. * All (roughly 32 more) commonly used pytorch layer systems supported in one form or another with wide-batched kernels to benefit both eager and compiled, many of which require reworks or completely remaking them. * Autotuning wide formats based on hardware response to the kernels. Kernel chunking for big slow processes such as LSTM, kernel fusion for small process with excess overhead, expanding kernels with masking to fit specific use-case paradigms with hardwares, and a series of smaller and more important optimizations along the way. * Full transformer and rope support with wide-batched optimizations throughout the structures to allow more robust autoregression throughput. * Additional Conv1d, Conv2d, and Conv3d optimizations.
>version 1.0 : * Entire diffusion structures specifically kernelized for high-efficiency utilization with eager and compilation. * Video diffusion specific targets meant to heavily reduce computation costs on the gpu and increase computation throughput on the gpu.
reacted to AdinaY's
post with π₯about 1 hour ago
Following up on the Gitee release, here's another major Chinese code dataset from GitCode (CSDN's code hosting platform). Same pipeline, same clean format, more valuable data from China's developer ecosystem.
The final dataset in the Chinese code series is also available: nyuuzyou/jihulab-code. It's smaller in size but shares the same pipeline and formatting.
reacted to wangbuer999's
post with πabout 1 hour ago
Summary: Most systems still run on βInputs β model/heuristic β single score β actionβ. But real deployments have multiple goals plus non-negotiable constraints (safety, ethics, legal). This article is a design cookbook for migrating to goal-native control: make the goal surface explicit as a **GCS vector**, enforce **hard constraints first**, then trade off soft objectives inside the safe set.
> The primary object is a GCS vector + constraint status β not a naked scalar score.
---
Why It Matters: β’ Stops safety/fairness from becoming silently tradable via βmystery weightsβ β’ Makes trade-offs auditable: βwhy this action now?β can be reconstructed via Effect Ledger logging β’ Gives a repeatable build flow: goals β constraints β action space β GCS estimator β chooser β’ Shows how to ship safely: shadow mode β thresholds β canary, with SI metrics (CAS/SCover/EAI/RIR)
---
Whatβs Inside: β’ A recommended GCS convention (higher=better, scales documented, weights only for soft goals) β’ Chooser patterns: lexicographic tiers, Pareto frontier, context-weighted tie-breaks β’ Practical patterns: rule-based+GCS wrapper, safe bandits, planning/scheduling, RL with guardrails β’ Migration path from legacy heuristics + common anti-patterns (single-scalar collapse, no ledger, no PLB/RML) β’ Performance tips: pruning, caching, hybrid estimators, parallel evaluation
---
π Structured Intelligence Engineering Series Formal contracts live in SI-Core / GCS specs and the eval packs; this is the *how-to-design / how-to-migrate* layer.
reacted to AdinaY's
post with π₯about 1 hour ago
More lightweight multimodal models are coming π
StepFun has been focused on multimodal AI from the very beginning. Their latest release a new foundational model: STEP3-VLπ₯ https://huggingface.co/collections/stepfun-ai/step3-vl-10b β¨ 10B - Apache2.0 β¨ Leads in the 10B class and competes with models 10β20Γ larger
