Today in DeAI: NEAR AI confirms an SGLang remote-code-execution breach, llama.cpp ships its first stable 1.0.0, Inkling 2 Small weights go live under Apache 2.0, NEAR AI Gateway lands on macOS, and a new TCO analysis puts the self-hosting break-even at 33–58% GPU utilization.
NEAR AI confirms SGLang RCE breach: about 100 accounts, 18 agents exposed
NEAR AI's incident report says an unauthorized party exploited a remote-code-execution vulnerability in SGLang, the open-source inference engine running on its servers, then reached account details for about 100 users and 18 agents on a legacy non-TEE deployment. The report's key admission is that confidential computing protects against the operator and host, not against a vulnerability in the software running inside the enclave — and confidential-mode requests were observable on the compromised server. Why it matters: TEE-backed privacy claims and engine-patching discipline are separate checks, and this is the first confirmed case of the second one defeating the first at a private-inference provider. (NEAR AI) — our full coverage
llama.cpp ships v1.0.0 with stable public headers
The ggml-org team released llama.cpp v1.0.0 and ggml v1.0.0, the project's first stable, semantically versioned releases after 3.5 years, publishing stable headers for libllama and libggml and committing that API and ABI breaks will ship only with major-version bumps. Why it matters: downstream private-inference tools can now depend on llama.cpp without chasing a moving tree, which lowers the cost of the fully-local deployment path. (ggml-org)
Inkling 2 Small weights go live under Apache 2.0
Thinking Machines Lab published the Inkling 2 Small checkpoint on Hugging Face under Apache 2.0, resolving the open question from the November 7 announcement. The model card lists 594B total parameters with 17B active, a 1M-token context window, and a main NVFP4 checkpoint requiring about 500 GB of aggregated VRAM, with 2x B300 or 4x H200 GPUs named as example configurations. Why it matters: an Apache-2.0 frontier-adjacent coding model is now self-hostable in principle, at a memory cost that puts it firmly in multi-GPU-node territory. Benchmark figures on the card are vendor self-reports. (Hugging Face) — our November 7 coverage
NEAR AI Gateway brings private models to Claude Code, Codex, and OpenCode
NEAR AI's macOS app launches terminal sessions with Claude Code, Codex, or OpenCode preconfigured to route inference to NEAR AI Cloud's private (TEE) and incognito models, with session trace inspection on your machine and opt-in data sharing that earns Cloud credits. Why it matters: this is the lowest-friction path yet to hardware-attested private inference inside tools builders already run — though the company's own breach report two days earlier is a reminder that the enclave boundary is not the whole security story. Credit-discount mechanics are vendor claims until independently exercised. (NEAR AI)
Self-hosting beats the API only above 33–58% sustained utilization
A Qovery analysis of gpt-oss-120B on 8x B200 GPUs finds self-hosting beats Together AI's $0.60 per 1M output tokens only above roughly 33% sustained utilization on Google Cloud DWS Flex-start (~$0.20/1M at full utilization, ~$0.99 at 20%) and roughly 58% on AWS on-demand ($113.93/hour for p6-b200.48xlarge). Why it matters: the self-host-versus-API decision is a utilization threshold, not a sticker-price comparison — and the blog's own cited research suggests most enterprise GPU fleets sit below it. This is one engineering blog's math built on public pricing, not an independent benchmark; verify current cloud prices before quoting. (Qovery) — our cost economics coverage
Watching tomorrow
Whether the SGLang project publishes a CVE and affected-version range for the vulnerability NEAR AI reported — the advisory is the signal every SGLang operator is waiting for.
Sources
- NEAR AI Security Incident Report: SGLang Vulnerability — NEAR AI
- llama.cpp v1.0.0: A stronger foundation for the local AI ecosystem — ggml-org / Hugging Face
- thinkingmachines/Inkling-2-Small — Hugging Face (Thinking Machines Lab)
- Introducing NEAR AI Gateway — NEAR AI
- 7 best ways to self-host AI inference on cloud GPUs in 2026 — Qovery
- Perplexity Decider v1.3: Training, eval, and use cases — Perplexity
About DeAI
DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.
Powered by Morpheus and StrandCMS
Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.
