Open LLM Deploy

The appeal is portability: Pi has been the author's go-to agent for several months and "basically runs everywhere".

• Claude plan access is shifting:...

The paper's author says Soft-MTP's benefits over standard MTP are consistently growing across scale, though they could only fund experiments up to...

Dual Batch Overlap (DBO) aims to overlap the sparse all-to-all communication in MoE layers with the surrounding computation. It splits the batch...

Rather than relying on one aggregate score, this TACL 2026 paper asks whether benchmarks measure many independent abilities or a few shared...

The Show HN post Maple claims sub-2-BPW MoE pruning can run Gemma 4 26B at 45 tokens per second on a 6 GB GPU. A commenter calls it another step for...

Overall scores were high, but the gaps matter for anyone using general-purpose LLMs in medical education. Across five systems and 227 infectious...

The article introduces EmbeddingGemma 2, a compact open model that places text, images, audio, video, and mixed inputs in one embedding space. It...

All four alternatives recreated LangSmith's trace tree in the authors' test, so the real differences lie in effort, cost tracking, and footprint.

-...

Muse Spark Open has two separate access paths, and the article keeps them distinct.

• On Foundry: the model is generally available, but you need a...

The bottleneck was the software layer, not the modest card. The author says Ollama's Go management layer adds considerable processing overhead, and...

Public benchmarks leak, so a high score may just mean a model has seen the questions. SealedBench tries to remove the evaluator's discretion by fixing...

Passing a SOC 2 or ISO 27001 audit depends on your organization's controls, not on the software you run. No tool is SOC 2 compliant by itself, and an...

Ricursive has released the Ripple-65B Preview under an Apache 2.0 license. The company reports a 80.52 on TerminalBench 2.1, ahead of Reflection...

The author's most useful local LLM project wasn't a chat session. It was a nightly n8n workflow running Qwen 3.5 9B on a home server.

• What it...

Inkwell shows how far local inference can reach: SmolLM2 runs through llama.cpp entirely on a jailbroken 2015 Kindle Paperwhite.

• Quick mode: about...

Ripple-65B is a plausible self-hosting candidate, but its headline claims are not yet independently confirmed.

• Benchmarks: It reports 80.5 on...

Static benchmarks measure population-level preferences, which can miss how individual needs vary by context. PersonaAlign takes a different approach:...

A new paper proposes Dynamic Boundary Evaluation (DBE) as an alternative to fixed benchmarks, which apply the same items to every model and can...

High scores on familiar entities can make it hard to judge how well an LLM repairs unfamiliar data. The CoRe protocol tests this by pairing fresh...

Harvey's held-out legal agent results show a wide gap between partial and complete work. Gemini 4 Argon reaches 96.30% on criteria pass rate, but its...