Engineering notes on LLM routing, gateways, and AI infrastructure.https://pirouter.ai/enhttps://pirouter.ai/blog/gguf-filename-is-a-request/https://pirouter.ai/blog/gguf-filename-is-a-request/A tensor width not divisible by 256 makes llama-quantize fall back to ~4.5 bpw and keep the low-bit filename. We measured Flash-Next's 1.56 bpw file at 3.28.Sat, 29 Aug 2026 00:00:00 GMTquantizationmodelslocal-llmLinden Kernhttps://pirouter.ai/blog/glm-5-3-open-weights-day-one/https://pirouter.ai/blog/glm-5-3-open-weights-day-one/Fourteen hours after the weights landed, sixteen hosts served z-ai/glm-5.3 at five input prices and four declared precisions; the cheapest runs BF16.Sat, 29 Aug 2026 00:00:00 GMTmodelspricingprovidersquantizationLinden Kernhttps://pirouter.ai/blog/model-access-change-of-control/https://pirouter.ai/blog/model-access-change-of-control/OpenAI gave Cursor 76 days under a change-of-control clause and excluded future models. What the clause means when your models arrive via someone else's key.Sat, 29 Aug 2026 00:00:00 GMTprovidersroutingfailoverLeo Kakahttps://pirouter.ai/blog/one-api-openai-anthropic-google/https://pirouter.ai/blog/one-api-openai-anthropic-google/Three ways to call OpenAI, Anthropic and Google through one API — a vendor's OpenAI-compatible endpoint, a self-hosted proxy, a hosted gateway — read against the official docs. All three flatten the request shape. None can flatten tool_choice, reasoning state, cache control, stream recovery or error codes.Sat, 29 Aug 2026 00:00:00 GMTgatewayroutingagentsLeo Kakahttps://pirouter.ai/blog/open-weights-still-need-a-host/https://pirouter.ai/blog/open-weights-still-need-a-host/K3's weights are free; serving starts at 8× GB300, so a host stays. Reuters: Moonshot wants up to 30% from Azure, AWS and Google — token auditing is unresolved.Sat, 29 Aug 2026 00:00:00 GMTprovidersmodelsecosystemLinden Kernhttps://pirouter.ai/blog/openrouter-alternatives-production/https://pirouter.ai/blog/openrouter-alternatives-production/Seven vendors' docs, read 2026-08-29, on six production columns: failover triggers, spend caps, protocols, ZDR, rate-limit units, per-request logs. Every gap marked.Sat, 29 Aug 2026 00:00:00 GMTgatewayprovidersfailoverLeo Kakahttps://pirouter.ai/blog/nvidia-hf-talks/https://pirouter.ai/blog/nvidia-hf-talks/Business Insider reports Nvidia has been in talks to acquire Hugging Face at more than $13 billion — no deal yet, and the talks could still fall apart. In January, Hugging Face turned down $500M from the same company because it did not want a single dominant investor. Here is the timeline with the qualifiers kept intact, what the HN headline got wrong, and how the three neutrality measurements we published last week port from routers to model hubs.Thu, 27 Aug 2026 00:00:00 GMTprovidersecosystemopinionLinden Kernhttps://pirouter.ai/blog/shared-team-wallet-llm-api/https://pirouter.ai/blog/shared-team-wallet-llm-api/People keep asking which LLM API aggregator has a shared team wallet. Read from the docs: OpenRouter and Vercel AI Gateway have a real shared balance; LiteLLM and Portkey have fine-grained member controls but no wallet. OpenRouter comes closest to both halves — until you hit the 10-member cap and the Enterprise gates. Here is the matrix, the fees, the reset cycles, and the caveats that don't appear on any pricing page.Thu, 27 Aug 2026 00:00:00 GMTcostgatewayapi-keysLeo Kakahttps://pirouter.ai/blog/failover-signals/https://pirouter.ai/blog/failover-signals/Seven LLM gateways' failover docs, read verbatim: the trigger lists agree, every cooldown is a timer, and no routing policy records a 429 lasting eight hours.Tue, 25 Aug 2026 00:00:00 GMTreliabilityfailovergatewayLeo Kakahttps://pirouter.ai/blog/per-key-spending-limits/https://pirouter.ai/blog/per-key-spending-limits/How LiteLLM, OpenRouter and Portkey cap spend per API key — layer, timing, status code, reset — and the four leaks a cap doesn't close, read from the docs.Tue, 25 Aug 2026 00:00:00 GMTcostgatewayapi-keysLeo Kakahttps://pirouter.ai/blog/tool-call-parser-attack-surface/https://pirouter.ai/blog/tool-call-parser-attack-surface/vLLM's Qwen3 Coder tool parser ran eval() on model output for 29 days in 2025 (CVE-2025-9141). Today: 0 eval(), 4 literal_eval. What a gateway can and cannot check.Tue, 25 Aug 2026 00:00:00 GMTsecuritygatewaylocal-llmLeo Kakahttps://pirouter.ai/blog/distribution-layer-for-sale/https://pirouter.ai/blog/distribution-layer-for-sale/In January, Hugging Face turned down $500M from Nvidia because it did not want a single dominant investor. In August, it is exploring a sale at $13B or more — the same month OpenRouter announced it is joining Stripe. Both companies sell neutrality. Here are three things you can measure yourself to find out whether that promise survives an acquisition.Mon, 24 Aug 2026 00:00:00 GMTroutingprovidersopinionLinden Kernhttps://pirouter.ai/blog/price-notice-is-a-contract-clause/https://pirouter.ai/blog/price-notice-is-a-contract-clause/Fourteen days, thirty days, thirty days, and none at all. In one week this August, one provider cancelled an increase it had already announced and another rebuilt its whole price sheet on three days' notice — and both acted entirely within their own terms. Here are the four clauses side by side, where each one is buried, and what the floor means for how often you should re-check a price.Sun, 23 Aug 2026 00:00:00 GMTpricingproviderscostPiRouter Teamhttps://pirouter.ai/blog/mcp-roadmap-gateway/https://pirouter.ai/blog/mcp-roadmap-gateway/Two of the five priority areas in the MCP roadmap published on August 22 — agent identity and tool-list bloat — land in front of the server, not inside it. DPoP, Workload Identity Federation, ID-JAG and token exchange each move work to the layer you already run. What the 2026-07-28 stateless release already gave that layer, and which parts of the overreach critique are correct.Sat, 22 Aug 2026 00:00:00 GMTmcpgatewayagentsPiRouter Teamhttps://pirouter.ai/blog/quantization-is-a-product-spec/https://pirouter.ai/blog/quantization-is-a-product-spec/A week of teacher-forced logit capture on one set of weights: at 88k context, official FP8 flips about 20% of top-1 tokens against BF16, NVFP4 about 50%. The interesting part is not the ladder. It is that the damage lands on tool calls — wrong interface, corrupted port number, a parameter name quietly truncated — and a routing table keyed on the model name cannot see any of the variables that caused it.Sat, 22 Aug 2026 00:00:00 GMTquantizationmodelsroutingLinden Kernhttps://pirouter.ai/blog/one-weight-many-prices/https://pirouter.ai/blog/one-weight-many-prices/One Apache-2.0 checkpoint reaches the market as two dozen GGUF files, seven API endpoints whose context windows run from 65,500 to 1,000,000 tokens, and — following one dated build across two hosts — two prices for a file that has not changed. The product is the tuple (weights, quantization, context window, version); a router keyed on the name is routing blind.Fri, 21 Aug 2026 00:00:00 GMTmodelsquantizationroutingLinden Kernhttps://pirouter.ai/blog/retry-storms/https://pirouter.ai/blog/retry-storms/GitHub's August 17 outage started with an autoscaling policy and spent its last four and a half hours fighting its own retries — one token service went from 7–9K to 70–100K requests per second on retries. The same arithmetic is sitting in your LLM SDK defaults and in your gateway's failover path. Here is the mechanism, the week's LLM-provider samples of it, and the four rules a gateway owes the providers behind it.Fri, 21 Aug 2026 00:00:00 GMTreliabilityfailovergatewayLeo Kakahttps://pirouter.ai/blog/stealth-models-and-price-sheets/https://pirouter.ai/blog/stealth-models-and-price-sheets/stealth/ox-alpha launched with a 1M context window, $0 pricing and no vendor name; deepseek-v4-flash-vision-exp launched the same day with nine benchmarks and a full price sheet. Both work. We compare the two listings on the four fields a router needs before either can enter a production routing table: identity confidence, retention policy, free-window expiry, and price-sheet completeness.Fri, 21 Aug 2026 00:00:00 GMTmodelsroutingstealth-modelsdeepseekLinden Kernhttps://pirouter.ai/blog/time-as-a-pricing-variable/https://pirouter.ai/blog/time-as-a-pricing-variable/DeepSeek now prices every line twice, peak and off-peak; GLM-5.3 launched with off-peak discounts. If 30% of your tokens are movable, a half-price window is worth up to 15% of the bill. The official rate cards, the peak windows converted to your timezone, and where the deadline-not-start-time decision belongs.Fri, 21 Aug 2026 00:00:00 GMTpricingroutingcostLeo Kakahttps://pirouter.ai/blog/routers-arent-dead/https://pirouter.ai/blog/routers-arent-dead/A widely shared post-mortem says prompt caching killed the per-request model router, and its arithmetic holds up: a 10x cache discount beats a 2.5x list-price spread every time. We build a router. Here is what we co-sign, what we concede, and where routing actually earns its keep.Thu, 20 Aug 2026 00:00:00 GMTroutingcachingopinionLinden Kernhttps://pirouter.ai/blog/billing-failure-domain/https://pirouter.ai/blog/billing-failure-domain/A Google billing account dropped from Paid Tier 3 to Tier 1 with no notice, pausing a live voice app; on NVIDIA NIM, GLM-5 retired before its replacement was callable. No status page moved, no 5xx fired. Four things to do while everything is still green.Wed, 19 Aug 2026 00:00:00 GMTreliabilityfailoverprovidersLeo Kaka