Mojo reaches 1.0 with stable language foundation
Mojo 1.0 commits to additive-only changes during 1.x, unifies syntax (single var declaration, unified closures, one Pointer type), and ships memory safety diagnostics for reference invalidation.
Breaking changes halt mid-release, enabling multi-year project stability without continuous rewrites. LSP improvements and AI Skills tooling reduce setup friction for systems programming on CPUs, GPUs, and accelerators.
Install via `uv pip install --upgrade mojo` today. Replaces beta instability with production-grade compatibility guarantees. Requires familiarity with Mojo syntax convergence (var-only, lambda closures). Worth trying now if you're building performance-critical Python-adjacent code; skip if you need async/pattern-matching (roadmapped, not shipped).
- “Mojo 1.0 does not mark the end of the language's evolution, but it is an important milestone on a longer journey”
- “During the 1.x timeframe, changes should primarily be additive”
- “nearly 200 contributors have landed more than 1,100 pull requests, changing over 200,000 lines of code”
- “it is no longer just a language we are developing; it is a language we rely on every day in production”
- “Mojo now diagnoses memory safety problems involving reference invalidation”
mojo-langsystems-programmingstabilitygpu-computerelease
Routing moves from gateways into agent loops
NVIDIA open-sourced NeMo Switchyard to route each agent step independently to different models, replacing single-model-per-request patterns with per-step selection that cuts costs to roughly one-third while maintaining frontier accuracy.
Agent workflows fan out into dozens of heterogeneous subtasks (planning, coding, review, testing). Per-request routing wastes compute on cheap steps and degrades output on hard ones. Per-step routing lets you route lightweight tasks to fast cheap models and complex reasoning to frontiers, directly reducing your inference bill while maintaining quality.
Replaces gateway-only routing products like Google Cloud API Gateway and Databricks Unity AI Gateway. Requires self-hosting a router (Switchyard or custom logic) plus consolidating vendor credentials into a unified supply layer (the article suggests wrouter.ai as one option). Ready now: weights ship on Hugging Face, Switchyard is open-source, and VS Code extension support exists. Start with per-step routing tables and a single base_url entry point; this is worth trying immediately if you run agentic workflows.
- “routing is sinking fast — into open-source libraries, into the agent loop, into the editor itself”
- “A single user task fans out into dozens of model calls — planning, tool use, code generation, testing, review, correction”
- “frontier-level accuracy maintained while task completion cost drops to roughly one-third of running Opus 4.8 alone”
- “100% domain-routing accuracy, with 59% of traffic dispatched to a model that is 5x faster”
- “up to 4x faster output and 30% faster agentic task completion versus its class”
model-routingagent-workflowsper-step-routingnemotron-3.5cost-optimization
NVIDIA releases Lightning model and smart routing
Nemotron 3.5 Lightning (30B MoE) delivers 4x faster output for agentic tasks; pair with NeMo Switchyard to route requests across model ensembles without application rewrites.
Developers building multi-agent systems can now skip manual routing logic and inference optimization—Switchyard handles cost/latency tradeoffs automatically across open, proprietary, and NVIDIA models. Lightning enables local deployment on RTX hardware, preserving privacy and infrastructure investment.
Replaces hand-coded routing layers and single-model fallbacks. Requires integrating Switchyard (available via LiteLLM, LangChain, Kong, or direct GitHub) and fine-tuning Lightning on domain data via NeMo. Ready to try now: benchmarks are real (Boomi 100% routing accuracy, Ramp 58% cost cut), ecosystem partners already shipping integrations.
- “up to 4x faster output speed, leading to 30% faster agentic task completion compared with other models in its class”
- “NeMo Switchyard maintains frontier-level accuracy while reducing task completion cost to nearly one-third of Opus 4.8 alone”
- “designed for this architecture. A frontier reasoning model such as Nemotron 3 Ultra or GPT-5.6 may plan and orchestrate a workflow, while smaller specialized models like Nemotron 3.5 Lightning can perform targeted tasks”
- “can run locally or on premises for high-volume, specialized tasks that require fast responses”
- “LangChain: achieved 74% lower cost in 145 multi-turn Deep Agents tasks by routing only 7% of calls to a frontier model”
agentic-aimodel-routingopen-modelsinference-optimizationmulti-agent-systems
Route coding agents through single gateway command
Vercel AI Gateway now auto-configures 9 coding agents (Claude Code, Cursor, Cline, etc.) to route through centralized gateway with one CLI command, replacing scattered API key management.
Eliminates manual provisioning across multiple agent platforms and consolidates observability, billing, and policy enforcement into one dashboard. Developers can enforce spend budgets and data retention rules that no agent config can bypass.
Replaces manual API key distribution and multi-account setup. Requires Vercel CLI upgrade and agents to already be installed locally. Ready now: command writes credentials directly into each agent's native config format without re-pinning models. Minimal friction for teams already using these agents.
- “Any of 200+ models in any agent, including models the agent has no native support for”
- “vercel ai-gateway coding-agents setup connects Claude Code, Codex, OpenCode, and more in one command”
- “Team-wide policy your agents can't route around: turn on Zero Data Retention or restrict which providers may serve requests”
- “This command supports 9 coding agents to utilize the models in AI Gateway”
agent-gatewayvercel-climulti-agent-routingobservabilitycost-control
LFM2.5-VL-3B ships multimodal inference on-device
3.1B vision-language model hits 228 tokens/sec on M5 Max, 3GB footprint, with native llama.cpp/MLX/vLLM support—skip the API roundtrip for document/UI understanding.
Eliminates inference latency and API costs for real-time screen parsing, document OCR, and object grounding in edge applications. Direct transformer loading cuts deployment friction versus proprietary APIs.
Replaces heavier 8B models (Gemma-4-E4B, Qwen3.5-4B) for on-device use; requires transformers>=5.10.1 and bfloat16 support. Ready now—benchmark parity on RefCOCO grounding (87.9) and ScreenSpot GUI tasks, but tool-use lags (ToolSandbox 59.5 vs Qwen3.5-2B 47.7). Start with document/screen workloads where it leads.
- “LFM2.5-VL-3B decodes 228 tokens/s on an M5 Max and 116 tokens/s on a Ryzen AI Max+ 395, and fits in about 3 GB of memory”
- “compatible with transformers>=5.10.1”
- “reaches about 11K tokens per second at high concurrency”
- “extends the vision-language capabilities of our previous releases with four major improvements: Screen/UI understanding, Grounding, Multi-image input, Function calling”
- “pre-trained on about 34T tokens, with 4x more vision data than before”
vision-language-modelson-device-inferencedocument-understandingtransformer-deploymentfunction-calling