Open LLM Deploy

NVIDIA's open recipe delivers 30/42 on IMO 2026 via natural-language proofs alone, using specialist post-training, iterative generate-verify-refine,...

NCP-ArchPreview's joint next-token and next-concept prediction lets an 8.9B latent-space model match OLMo-3-7B's final pretraining loss with only...

• Releases full assets: two Nemotron 3 Ultra specialists (SFT + RL), training data, code, and Nemotron-IMO-Bench
• Natural-language pipeline:...

单GPU即可完成有意义的开源LLM研究：从GPT-2风格代码扩展为6专家（2激活）MoE模型，从零训练8天即取得成功。 这条路径证明小型稀疏模型的实践门槛已大幅降低。

Can SWE-Bench Pro Verified truly measure software engineering agent capabilities?

• Anti-hacking layers separate inference from evaluation, erase Git...

DeepSeek-V4.1-Flash runs at only 8–16B active parameters despite 522B total, yet the engram n-gram tables alone consume 189 GiB.

• Disk footprint...

Coder Agents moved from beta to general availability, adding production features like improved durability, stabilized headless APIs, deeper RBAC, and...

Benchmark Ollama local LLMs with a single command to measure throughput across NVIDIA GPUs, Apple Silicon, and CPUs, then compare results in the community database of real hardware setups. Install via pip or uv for macOS, Linux, or Windows.

InferenceMAX introduces an open-source harness to measure token throughput, inter-token latency, TTFT, cost per million tokens, and tokens per...

The AutoResearchExam benchmark evaluates agents on open-ended machine learning and engineering tasks across seven areas including model training, data curation, and AI safety, granting 24 hours per task.

With zero shared benchmark results between Mercury 2.5 Preview and Phi-4, no quality verdict is possible. Model comparisons should expose these...

Skip borrowed speed tables—calculate whether a model fits your hardware first.

• Model memory equals parameters times bits per parameter, then add 25%...

Can LLMs move beyond kernel tweaks to autonomously engineer full LLM infrastructure?

Φ-Bench addresses this by replacing isolated operator benchmarks...

Looping in GPT-6 Astra is simply weight reuse to stack more layers without extra memory, producing tokens sequentially like any standard LLM. OpenAI...

Procedural Graphs replace unconstrained tool-use generation with structured (procedure, relation, procedure) triplets that localize the agent's active...

A flash model matching closed frontier models prompts immediate skepticism: has it cooked, or is the benchmark borked? Impressive on paper, but verify testing conditions first.

The article maps real-world minimum hardware for local inference across five devices:

• MacBook Pro M4 (48GB) handles a 27B Qwen model at ~10...

Normalized leaderboards reveal a narrower gap than broad claims suggest, but frontier models still lead on hard reasoning.

• Artificial Analysis...

長文推論のメモリ問題に対し、BeaconKVはKVキャッシュを最大5.8倍圧縮しつつ精度をほぼ維持。

Unoは既存モデルに拡散アダプタを追加し、8Bモデルで2.5倍スループット向上を実現、品質は損なわない。

BeaconKVは長期CoTのメモリ削減に強く、Unoは並列生成で速度を優先。両者ともオープンソース寄りで32–64GB VRAM実用化の鍵となる。

OpenAI's Jakub Pachocki describes inference models as grown rather than designed—simple math steps applied to vast data without layer-by-layer...