Find published research by company, product, platform or technology.
Showing 1–20 of 76 dossiers
CS-4 combines three WSE-3 Turbo wafers with Cerebras’ Nexus rack design. The practical shift is architectural: compute, power and I/O become modular, while Cerebras now says the same platform is intended to support CS-5 in 2027 and a 3D-memory CS-6 generation after that.
Groq 3 LPX is moving from architecture announcement to manufactured infrastructure. Artificial Analysis measured about 3,400 output tokens/s at both 10K and 100K context on an NVIDIA-hosted private endpoint, but the single-concurrency benchmark does not yet establish public-cloud price, multi-tenant throughput or end-to-end agent speed.
DeepSeek V4.1 Flash supersedes the old Flash and Vision-Exp API lines with native multimodality, lower pricing and new architecture. Unlike those retired Flash aliases, the current DeepSeek API changelog and rate card still show V4 Pro as a distinct service.
Jalapeño is working first-party silicon rather than a roadmap item, and OpenAI now says AI itself materially accelerated the design process. The distinction still matters: tape-out means the design was finalized for manufacturing; it does not mean fleet-scale production qualification or API deployment is complete.
SwarmLLM does not route whole prompts to separate machines; it pipelines one model across browser tabs. A MacBook and iPhone can jointly hold Qwen 3.8 27B even when neither device can hold the full 15GB quantized model alone, with no inference server in the loop.
Meta has made the privacy-versus-price trade explicit in its Model API: developers can choose standard pricing or a contributor model ID with steeply discounted inference in exchange for training-data permission. The choice matters for proprietary code, customer data and AI SaaS workloads.
Meta’s Muse Glimmer 30B combines tool use, coding, vision and agentic task completion with official local-runtime artifacts. A 17GB GGUF build targets 24GB-VRAM machines, but Meta also attaches a separate usage policy, so builders should distinguish weight availability from unrestricted use.
GitHub Spark stops being available to existing users on August 31, 2026. Deployed apps are meant to keep running, but owners should export code to a repository now; Spark apps using `llm()` need a separate inference provider because the underlying GitHub Models service retired July 30.
Self-Hosted Machines changes the architecture of Cursor’s Cloud Agents more than another model option would. Teams can keep code, build outputs, secrets and terminal/browser actions on infrastructure they control, but the planning/inference loop remains a Cursor service and enterprise teams become responsible for worker images, scaling, secrets and production validation.
Training experiments and batch inference can use Together AI's discounted preemptible GPUs in existing clusters. Workloads must checkpoint or requeue on interruption, and at least one standard node is required.
The post-release evidence sharpens the original story. Qwen3.8-27B can retain useful agentic-coding performance at practical 4-bit sizes, but local model quality is not a property of the checkpoint alone: quantization, reasoning effort, context handling and the agent harness can materially change the result.
The important change is at the gateway boundary, not just inference placement. OpenRouter says prompts can now stay in-region from decryption through provider execution and supported server tools, while teams can enforce the rule per workspace, team or API key.
Android Studio’s agent layer has crossed an important boundary from preview features into the stable channel: domain-specific skills are preloaded and auto-selected, while Gemma 4 can execute tool-calling code tasks locally without sending source code to a cloud model.
OpenAI’s internal data turns “agents make researchers faster” into a measurable operating model: heavy concurrent agent use, record experiment throughput and rising task complexity, alongside high token spend and persistent human intervention on longer work.
Project Zenith is not a new model or another Copilot feature. It standardizes a developer-focused Windows experience and hardware floor for local AI work, with preconfigured tooling and OS settings intended to reduce setup friction and dependence on metered cloud inference.
Hugging Face has released 207 Apache-2.0 WebGPU kernels, a JavaScript loader and Fleet, a browser benchmarking service. The package makes kernel contracts and correctness evidence inspectable, but performance remains device- and workload-dependent.
SnapStart previously covered only selected managed runtimes; extending it to container images changes the latency-versus-packaging trade-off for teams shipping large dependencies or standard container bases, with regional exclusions and runtime-specific guidance still applying.
AgentControl now spans more production stacks: applications can resolve different prompts and models by context, track token/cost behavior, require approvals, use Bedrock without proxying inference through LaunchDarkly, and inspect multi-step agent runs as one conversation.
The price changes are not uniform: H100/H200 rise about 14%, B200 30%, B300 25% and GB300 about 11%. Builders using dedicated inference or training should re-run workload economics before assuming newer accelerators remain the cheapest route per completed task.
Sentence Transformers 6 now has both unified multi-vector inference and a documented end-to-end training workflow. A new project-authored benchmark shows fast domain adaptation on a single GPU, but the result is workload-specific and index costs remain high.