Created by Peter Felber
Latest open-source LLM releases, benchmarks, and deployment guides for 32‑64 GB VRAM setups
Explore the latest content tracked by Open LLM Deploy
A community GGUF build of the fine-tuned Qwopus3.8-27B-Flash model brings practical vision-language capabilities to local hardware, with ready pipelines for vLLM, SGLang, and Docker that emphasize faster decoding and lower agent-loop latency.
Current AI benchmarks are totally broken or saturated because they fail to test real-world long-running loops, highlighting the urgent need for revamped evaluations focused on sustained task completion and recovery.
K2 Horizon's Apache 2.0 release removes licensing fees across six models, yet hardware and runtime expenses persist.
Fast weights and sparse attention in GLM-5.3-Flash target efficiency gains specifically for long-context inference.
A single training query reaches 71.5% state coverage within the first 100 steps and recovers most of full-data OPD's performance across domains and...
Models re-scan the entire KV cache for every generated token, even when attending to just a tiny slice of a million-token context. Declarative...
How does LLMScholarBench reveal the accuracy-representation trade-off in LLM academic expert recommendations?
MBZUAI released six fully open K2 Horizon models (0.9B–375B) with weights, code, training data, and methodology under Apache 2.0.
Cerebras now offers Qwen 3.8 27B publicly at ~1500 tokens/s with full 128k context support. This hosted speed sets a new benchmark for practical, high-throughput inference on a 27B-class open model.
EarlyEval trains lightweight LightGBM classifiers on behavioral and textual features to halt agent runs once success or failure is predictable. This...
Launch positions 1.3 as the biggest jump yet on coding and agentic tasks.
Early side-by-side test gave both versions the same prompt to output one...
Agent Relay lets enterprises keep their preferred cloud-hosted coding agents while routing execution into self-hosted Coder workspaces, solving...
Two new benchmarks reveal how LLMs attempt self-improvement by rewriting their execution harnesses or chasing vague capability goals.
Self-hosting LLMs lets teams govern prompts, retrieval data, model versions, logs, and access through inspectable architecture, but an EU region alone does not guarantee GDPR compliance.
Fine-tuning adapts a pretrained model on task-specific examples to deliver reliable formats, terminology, and behavior where prompting alone falls...
Switchyard, NVIDIA’s Apache-2.0 Rust proxy, translates between OpenAI Chat/Responses and Anthropic Messages so agents like Claude Code or Codex CLI...
GenRecal transfers knowledge across incompatible vision-language models via its Recalibrator, letting smaller VLMs outperform large open- and...
这篇教程展示了如何在 B200 GPU 上从零构建并优化注意力内核，通过 60 张图表逐步接近最先进性能。对于追求高吞吐与显存效率的本地 LLM 部署者，这类底层优化直接提升推理速度并降低资源占用，具有显著实用价值。
Prompt optimization in multi-agent LLMs often fails when edits to task content corrupt execution protocols like routing and formatting. By...
E-Commerce Bench introduces the first open benchmark for year-long autonomous business ops, with agents managing stores, supplier negotiations,...
