Topic

#hugging-face

Loot, blog posts and adjacent themes connected to this topic. Follow the tag to keep it in your orbit.

#hugging-face

Loot

More from this topic

Explore all loot

Run the Kimi K3 Architecture Locally with a 0.18B Test Model

#kimi-k3 #local-ai+4

A 0.18B-parameter Kimi K3 derivative that preserves the core layer types in a roughly 700 MB checkpoint for local architecture testing and development. Kimi-K3-0.18B is a tiny, MIT-licensed derivative of Moonshot AI’s 2.8T-parameter Kimi K3. It reduces the model to four layers, eight experts and 0.18B total parameters, with roughly 0.10B activated per token. The single model.safetensors checkpoint is about 0.69 GB, making the architecture accessible on ordinary development hardware. What it preserves The miniature model retains Kimi K3’s main layer types: dense and MoE feed-forward blocks, KDA linear attention, and an MLA full-attention layer. That makes it useful for testing loaders, integrations, architecture support and compression workflows without downloading or serving the full 2.8T model. Best use cases Validate Transformers compatibility and custom Kimi K3 model code Prototype inference, quantization or conversion workflows Inspect the MoE and attention implementation on limited hardware Build development tests around the Kimi K3 architecture Loading caveat The example loader requires trustremotecode=True, which executes repository-provided Python code. Review the model repository and pin a revision before running it in a sensitive environment. The model is not currently hosted by a Hugging Face inference provider.

Free

Review open

@ZachasADMIN

Open loot

Serve new Transformers models through vLLM without waiting for a custom port

#vllm #hugging-face+3

Hugging Face says the Transformers modeling backend in vLLM now reaches native-speed throughput for tested Qwen3 setups, giving model authors and infra teams a faster path from model card to serving. Hugging Face published a practical update for teams serving open models with vLLM: the Transformers modeling backend now meets or beats native vLLM throughput in its Qwen3 benchmark examples. The useful part is operational. If a model already has a solid Transformers implementation, you can try it in vLLM with --model-impl transformers instead of waiting for a hand-written vLLM model port. Use this when you are evaluating fresh Hugging Face models, internal checkpoints, or architectures that move faster than serving-framework integrations. It is also a good sanity check before writing custom serving glue. Why bookmark it Hugging Face reports native-speed or better throughput across three Qwen3 setups, including dense and MoE examples. vLLM's own docs describe the Transformers modeling backend as a supported path for models implemented in Transformers. The workflow keeps the model structure in Transformers while vLLM handles serving optimizations such as batching and kernels. The caveat is still real: benchmark your own model, quantization, GPU topology, context length, and traffic pattern before treating it as production-ready. Quick evaluation checklist Confirm the target model loads with --model-impl transformers on your vLLM version. Compare throughput and latency against the native vLLM implementation when one exists. Check memory use on your actual GPU setup, especially for MoE and multimodal models. Validate output parity on representative prompts before routing production traffic.

Free

Review open

@ZachasADMIN

Open loot

Test LongCat-2.0 before your next long-context coding-agent run

#LongCat-2.0 #open model+4

LongCat-2.0 is an MIT-licensed Meituan model on Hugging Face and GitHub with a 1M-token context target, coding-agent focus, and public deployment notes. Treat vendor benchmark claims as self-reported, and test it on your own repositories before trusting it in production. LongCat-2.0 is worth bookmarking if you evaluate open models for coding agents, repository-scale edits, or long-context experiments. The model card and repository describe a 1.6T-parameter MoE design, roughly 48B active parameters per token, MIT-licensed weights, a 1M-token context target, and deployment notes for SGLang/vLLM-style serving. Use it as an evaluation candidate, not an automatic production pick. The benchmark table is mostly vendor-reported, the hardware requirements are serious, and real value depends on how it handles your own codebase, tests, tool-calling format, latency, and safety controls. Practical checks before using it: Confirm the exact Hugging Face variant you want: full, FP8, INT8, or a community quantization. Run a small repository task against your current baseline model. Check license, trademark, privacy, and acceptable-use constraints for your deployment. Measure context retention and patch correctness, not only benchmark scores. Avoid assuming OpenRouter/API availability unless your provider page confirms the model at run time.

Free

Review open

@ZachasADMIN

Open loot

Blog

#hugging-face

More from this topic

Run the Kimi K3 Architecture Locally with a 0.18B Test Model

Serve new Transformers models through vLLM without waiting for a custom port

Test LongCat-2.0 before your next long-context coding-agent run

Related reads

OpenAI and Hugging Face detail AI-agent breach during model evaluation

Test Tencent Hy3 before choosing your next open agent model

Use Gemma 4 on Cerebras when voice agents need real-time vision

Qualcomm and Hugging Face plan hybrid agents for open-model deployment

IBM's CUGA examples turn agent harness design into copyable apps

AOHP proposes an Android-based OS harness for AI agents

Hugging Face Shows How to Benchmark Whether Tools Are Agent-Friendly

GLM-5.2 brings a 1M-token context to open coding agents

Kimi K2.7 Code ships as an open coding model for long agent runs

Hugging Face Serge puts AI code review inside GitHub pull requests

OpenEnv gets broader open-source backing for agentic RL environments

Agents' Last Exam tests AI agents on real professional workflows

Hugging Face benchmark tests voice agents on code-switched customer speech

Optimum Intel 2.0 Makes OpenVINO the Default Path for Local Open Models

Cohere North Mini Code Gives Agent Builders a 30B Open Coding Model

Hugging Face Transformers CVE-2026-4372 Turns Model Loading Into a Security Checkpoint

Hugging Face rebuilt the hf CLI for coding agents, not just humans

Holo3.1 pushes computer-use agents toward local deployment

JetBrains Mellum2 ships as an open MoE model for coding agents

LongTraceRL trains long-context reasoning from search-agent trajectories

Hugging Face’s AI agent glossary makes harness, scaffold, and skills easier to compare

OpenAI Privacy Filter brings on-device PII masking to Hugging Face workflows

NVIDIA Nemotron 3 Nano Omni targets long-context document, audio, and video agents

DeepSeek-V4 brings million-token context closer to practical agent workflows

The Open Agent Leaderboard compares full AI agent systems, not just models

Hugging Face Transformers v5.8 is a dependency upgrade worth testing, not rushing

IBM Granite Embedding Multilingual R2 lands with 32K context and a compact 97M retrieval model

DeepInfra joins Hugging Face Inference Providers and expands the practical model menu for teams that want API flexibility

Granite 4.1 details how IBM trained its new open 3B, 8B, and 30B models

Hugging Face is turning Reachy Mini into an app-store robot platform

Cohere Transcribe lands on Hugging Face as a 2B open-source ASR release

huggingface_hub 1.14.0 adds Space secrets management and pushes Hub automation further into the CLI

Ling-2.6-1T is making a serious case for useful intelligence per token