Phala

Phala Confidential AI: 57.82 billion tokens over the past 24 hours. Top models: DeepSeek V4 Flash 0731 24.69% GLM-5.3 Flash 24.69% Kimi K3 14.28% https://lnkd.in/gYKNW_Ex TEE-backed AI infrastructure.

Agents can generate language. Production workflows also need bounded decisions that software can threshold, route, and log. Kev-4B is now live in beta on Phala. It turns shared application state into typed probability distributions for yes/no, choice, and score questions through TypeSafe’s System One API. Agent use cases from the Kev-4B model card: • classification, routing, and triage • extraction choices • policy and eligibility checks • judging a proposed answer against stated criteria • confidence-based handoff to human review Selected model-card benchmarks (self-reported, fp32): • 83.8% accuracy on the transfer-v4 locked out-of-domain test (656 questions) • 90.3% accuracy on the documents-v1 test (936 questions) • 70.1% coverage at an error rate of 5% or less on transfer-v4 locked These results measure specific evaluation suites. Validate thresholds on labeled examples from your target workload. On Phala, Kev-4B runs with NVIDIA Confidential Computing, an 8,192-token validated context, and catalog pricing of $0.04 per 1M input tokens with free output. Read the launch article: https://lnkd.in/gjPTbrCE Try the model: https://lnkd.in/gi-PJqsr Benchmark methodology: https://lnkd.in/gy93aeDf #AIAgents #DecisionModels #ConfidentialAI

Phala Confidential AI: 85.07 billion tokens over the past 24 hours. Top models: DeepSeek V4.1 Flash 51.54% GLM-5.3 Flash 25.82% Kimi K3 5.22% https://lnkd.in/g8Zgbf8Z TEE-backed AI infrastructure.

Phala Confidential AI: 34.61 billion tokens over the past 24 hours. Top models: GLM-5.3 Flash 23.97% DeepSeek V4.1 Flash 23.09% GLM-5.3 14.66% https://lnkd.in/gEVMa8Fn TEE-backed AI infrastructure.

Private AI Proxy gives coding agents a local verification gate before prompts leave your machine. It checks the AI service's hardware attestation, pins TLS to the attested key, and pauses forwarding when verification fails. Claude Code, Codex, DeepSeek Harness, OpenCode, Pi, Oh My Pi, Hermes Agent, and OpenClaw connect through revocable local tokens, while provider credentials stay in the proxy. The app works with Phala, RedPill, and compatible ACI services. It also exposes local APIs compatible with OpenAI Responses, Chat Completions, and Anthropic Messages. Download on the Mac App Store: https://lnkd.in/gr2MGitT Get macOS, Windows, and Linux builds: https://lnkd.in/gdFPewG4 Read the technical overview: https://lnkd.in/gdesZ5mf Explore the open-source project: https://lnkd.in/dqW4mehT

Phala Confidential AI: 112.21 billion tokens over the past 24 hours. Top models: GLM-5.3 Flash 65.11% GLM-5.3 8.81% qwen3.5-397b-a17b 8.40% https://lnkd.in/gEVMa8Fn TEE-backed AI infrastructure.

Phala Confidential AI: 202.74 billion tokens over the past 24 hours. Top models: GLM-5.3 Flash 42.98% GLM-5.3 24.90% deepseek-v4.1-flash 13.25% https://lnkd.in/gEVMa8Fn TEE-backed AI infrastructure.

Privacy claims in AI should be verifiable at the workload level. SayGM moved its operator supply layer to Phala Cloud as confidential workloads. Each operator runs an approved image pinned to a digest inside its own hardware-attested confidential VM. Before a workload can serve traffic, SayGM’s registry checks its image hash. The attestation server and data-plane proxy run together inside the same measured workload, so the proof covers the code that receives the request and chooses its route. Operators bring their own upstream provider credentials. Those credentials are injected into the encrypted workload at deployment and remain sealed from SayGM and the host while the workload runs. The stack is built on dstack, the Apache 2.0 open-source TEE framework hosted by the Linux Foundation: https://lnkd.in/gaxjDBMy Live today: • 60 distinct coldkeys deployed • 18 operators active • 57 confidential VMs running SayGM shows how independent inference operators can serve traffic under one approved, verifiable runtime. Read the full case study: https://lnkd.in/dcrEzghA #ConfidentialAI #ConfidentialComputing #AIInfrastructure

Phala Confidential AI: 173.64 billion tokens over the past 24 hours. Top models: GLM-5.3 51.89% GLM-5.3 Flash 25.07% deepseek-v4.1-flash 10.61% https://lnkd.in/gTXc-TB7 TEE-backed AI infrastructure.

Phala Confidential AI: 254.51 billion tokens over the past 24 hours. Top models: GLM-5.3 70.49% Kimi K3 10.73% GLM-5.3 Flash 8.18% https://lnkd.in/gTXc-TB7 TEE-backed AI infrastructure.