DeepSeek V4 Flash weights update ships on AI Gateway
V4 Flash now routes updated weights by default—Terminal-Bench score jumps 25.8 points to 82.7—no code changes required.
Agentic capabilities improve significantly without migration friction; code agents targeting `deepseek/deepseek-v4-flash` automatically get the stronger model. Matters for developers running inference agents on constrained budgets or latency targets.
Drop-in replacement for existing V4 Flash calls. Requires only setting `model: 'deepseek/deepseek-v4-flash'` in AI SDK or agent config. Worth trying now if you're already on DeepSeek; other providers with updated weights ship next week. Zero Data Retention option incoming.
- “DeepSeek V4 Flash now runs on updated weights by default on AI Gateway, with notably stronger agentic capabilities”
- “On Terminal-Bench, it scores 82.7, up 25.8 points from 56.9 in the April preview”
- “Requests to `deepseek/deepseek-v4-flash` pick up the new weights automatically, with no change to the model ID or your code”
- “For now, DeepSeek is the only provider serving the updated weights. Other providers, including ones with Zero Data Retention, are coming next week”
deepseekai-gatewayagent-benchmarksinferencevercel
Moonshot releases 2.8T-parameter open-weight model
Kimi K3 is a 2.8-trillion-parameter open-weights model with 1M context length, hybrid linear attention, and expert sparsity tuned to 2% activation—available via Together AI's OpenAI-compatible API.
Developers can now run frontier-class reasoning on long-context tasks (coding, document analysis, deep reasoning) without vendor lock-in. The model's efficient sparsity architecture makes deployment feasible for teams that need open weights with competitive inference speed.
Replaces the need to call closed-API reasoning models for long-horizon work. Requires Together AI account, standard Python SDK (together>=2.0.0), and familiarity with OpenAI chat API patterns. Ready to use now: supports streaming, structured output, vision, tools, and configurable reasoning effort (low/high/max). Start with `reasoning_effort="max"` for complex tasks, toggle `reasoning={"enabled": False}` for fast inference on simple queries.
- “2.8-trillion-parameter model”
- “the world's first open-source model in the 3-trillion-parameter class”
- “first open-weights model competing at the GPT 5.6 Sol and Claude Fable 5 tier”
- “the largest open-weight model ever released”
- “first Kimi model to support a 1M context length”
- “efficiently activating 16 of 896 experts”
- “roughly 2% of experts activated per token”
open-weightslong-contextreasoningmixture-of-expertsapi
Prisma adds S3-compatible object storage to projects
Buckets are API-first, branch-aware, and provisioned alongside databases—agents can now handle file storage without human setup.
Eliminates the second-vendor friction for file uploads, avatars, and exports. Coding agents can provision storage mid-task via API, same as databases, removing a hard blocker for autonomous workflows.
Replaces manual S3/R2 signup-and-stitch workflows. Requires service token auth and standard S3 client library (Bun, boto3, etc.). Ready now: API is live, pricing not yet published—run agents in existing projects but avoid unattended provisioning at scale until costs are documented.
- “S3-compatible storage that lives alongside your Prisma Postgres databases”
- “Two API calls take you from nothing to reading and writing objects”
- “coding agent, handed a service token once, can provision storage the same way it already provisions databases: mid-task, without waiting for a human to click through a dashboard”
- “Object Store buckets remove that stop: storage is now a resource inside a Prisma project, next to the database, under the same account, on the same API surface”
- “Object Store pricing and plan limits are not published yet”
object-storageapi-firstprismaagent-friendlys3-compatible
Kimi K3 reaches US infrastructure with ZDR support
Moonshot AI's Kimi K3 now routes through US-based providers (Baseten, Fireworks) on Vercel's AI Gateway with automatic failover and optional zero-data-retention compliance mode.
Teams with data residency constraints can now run Kimi K3 without leaving US infrastructure, while automatic multi-provider routing eliminates single-provider latency and availability bottlenecks. The fast variant (~50% cost premium for lower latency) trades expense for response speed in latency-sensitive workflows.
Replaces single-provider Kimi K3 routing with managed multi-provider failover and regional isolation. Requires updating `model` ID to `moonshotai/kimi-k3` and optionally setting `inferenceRegion` or `zeroDataRetention` in providerOptions. Ready now—basic setup is one-line config change via AI SDK. Worth trying if compliance or uptime are constraints; cost-benefit of fast variant depends on latency requirements.
- “Kimi K3 and Kimi K3 Fast with ZDR and US-based providers now on AI Gateway”
- “Running Kimi K3 on US-based providers lets teams with data residency and compliance requirements use the model on US infrastructure”
- “Because AI Gateway serves the models from multiple providers, it automatically routes across them for failover, higher uptime, and more available throughput than any single provider offers”
- “Kimi K3 Fast trades a higher per-token cost for lower latency”
- “The fast variant costs ~50% more than the base model”
- “Regional pricing is ~10% more than the regular variant”
kimi-k3ai-gatewaycompliancemulti-providerzero-data-retention
Stateless MCP cuts implementation complexity in half
MCP 2.0 eliminates session management by moving from two-request stateful initialization to single-request stateless calls, removing server-side state tracking entirely.
Developers can now build MCP clients and servers with significantly less boilerplate, making the protocol viable for smaller models running on laptops and scalable web applications that avoid session routing complexity.
Replaces legacy MCP (Nov 2024–July 2026) with single-request HTTP pattern using headers instead of session IDs. Requires updating clients to use MCP-Protocol-Version: 2026-07-28 and Mcp-Method headers. Ready now—three working projects shipped this week by author; public spec and CLI tools available.
- “Stateless MCP specification also greatly decreases the complexity of implementing both clients and servers for the protocol”
- “MCP-Protocol-Version: 2026-07-28”
- “It's also a better fit for building scalable web applications, since now you don't need to maintain server-side state to keep track of those session IDs”
- “MCP is much easier to reason about agent capabilities and what might go wrong than with arbitrary command execution in an open network environment”
mcpagent-toolsprotocolstateless-designimplementation