Open-weight models crossed 71.7% of the tokens on the internet's most-watched inference router the same week the two biggest distribution layers rewired themselves around that fact. On November 16, Hugging Face handed language-model routing on Inference Providers to OpenRouter, ending bring-your-own provider keys for text models and letting a third-party router pick the serving provider per request. On November 19, Chutes — the largest inference subnet on Bittensor — stopped selling monthly subscriptions entirely and moved to pay-as-you-go. In between, Nvidia's interim CEO went on the record that the company's open-source strategy and its $12.9 billion Hugging Face acquisition survive Jensen Huang's exit.
That is one story told three times, and it is this week's theme: the models are open, and the question moved to who serves them, at what price, under which license. Five new open-weight releases landed — Meituan's 1.6-trillion-parameter LongCat-2.5, Runway's 27B robot-control Praxis-1, Tencent's 220B Hy4 flash, Microsoft's 2.3B Aion 1.0 Instruct, and Cloudflare's Clairic 27B and 9B extraction models — and not one shipped as unqualified Apache 2.0 with frontier benchmarks. The weights keep getting bigger while the licenses and the benchmark evidence keep getting smaller.
Last week's hub was about the gap between claims and proof. This week the gap is thinner but the plumbing underneath moved more than it has in months: a router consolidation, a billing-model collapse on the largest decentralized subnet, a price index where every mover scissored, and a tracker sweep that logged a record 18 check attempts and zero usable observations.
Key facts
- Open-weight models served 71.7% of tokens on OpenRouter as of November 16, 2026, per OpenRouter's own usage data (our full coverage).
- Hugging Face switched language-model calls on Inference Providers to route through OpenRouter on November 16, killing BYOK for language models; custom provider keys still work for image and video (our full coverage).
- Chutes stopped selling monthly subscriptions; its pricing page states "Monthly plans are no longer sold on chutes.ai", with Qwen3.8 Max and DeepSeek V4.1 Pro reachable only on pay-as-you-go billing and a 35% discount expiring November 23 (our full coverage).
- Nvidia interim CEO Colette Kress reaffirmed the open-source strategy and the $12.9B Hugging Face deal on the November 17 earnings call, citing a 35-fold year-over-year rise in OpenRouter weekly token volume — a company-cited, unaudited figure (our full coverage).
- Meituan released LongCat-2.5 under MIT: a 1.6-trillion-parameter multimodal MoE with roughly 48B active parameters per token, which Meituan says trained on domestic Chinese accelerators with under 80 GB memory per device — a vendor claim with no independent audit (our full coverage).
- The ninth DeAI Price Index observation found 6 of 17 OpenRouter rows moved and every mover scissored: deepseek-v4.1-flash input jumped 175% to $0.14/M while output fell 26% to $0.42/M, and kimi-k2.5 gave back most of the prior week's increase.
- The provider tracker's eighth sweep appended 23 entries to 13 providers and set an attempt record — 18 checks tried, none usable — leaving the grade census at 0 verified across 17 providers.
- Microsoft published MIT-licensed weights for Aion 1.0 Instruct, a 2.3B model now serving as Edge's built-in small language model with an origin-trial Prompt API (our full coverage).
- Cloudflare published open weights for Clairic 27B and Clairic-flash 9B, task-specific extraction models with Workers AI pricing at $0.24 and $0.10 per million input tokens (our full coverage).
The router consolidation: who picks your provider
The week's structural story is that the layer between model weights and the person paying for tokens consolidated — twice, in opposite directions.
Hugging Face's switch is the cleaner one to explain. Since November 16, every language-model call on Inference Providers flows through OpenRouter's router: it picks a provider per request based on live price, speed, and reliability, retries on a different provider inside the same call, and keeps conversations on whichever provider holds the prompt cache. Hugging Face billing appears on the Hugging Face account, bring-your-own provider keys no longer work for language models, and custom keys survive only for image and video. The stated economics — no Hugging Face markup on routed calls — are published in Hugging Face's own announcement, which is a first-party claim about a first-party product and about as verifiable as a changelog gets.
What the switch actually consolidates is failure surface. A router that retries across providers is a genuinely useful primitive — it converts provider outages from incidents into latency. But it also means one company's routing policy now sits between a large share of open-weight usage and the providers serving it. If OpenRouter's ranking data is the market's most-watched number — and after this week it is, twice over — then the company computing that number also decides which providers receive the traffic that generates it. That is not an accusation; it is a dependency worth naming. The Kress earnings-call citation of a 35-fold year-over-year rise in OpenRouter token volume makes the same platform load-bearing for Nvidia's open-source narrative, and that figure is explicitly a claim, not an audit (our October coverage of OpenRouter's open-weight token share flagged the same property when the number was 72% and smaller).
Chutes' move is the same consolidation seen from the decentralized side, and it is harsher. Chutes — the largest inference subnet on Bittensor — stopped selling monthly subscriptions and moved billing to pay-as-you-go credits; existing plans expire at the end of their current period, and its two newest beta models, Qwen3.8 Max and DeepSeek V4.1 Pro, are pay-as-you-go only. The operator's pricing page states the change directly, and a 35% discount on the two flagship models expires November 23. For builders, the arithmetic that changes is commitment: a subscription was a fixed monthly cost you could budget against a subnet's reliability; pay-as-you-go converts that into pure per-token exposure to the subnet's pricing volatility — the same volatility the Price Index watches on centralized routers, now arriving on a decentralized network with no list price for the new flagship models at all.
Read together, the two moves say the same thing from opposite ends of the trust spectrum: the middle layer of open-model inference is consolidating around per-request, pay-per-token economics, whether the router is a US company with a public catalog or a crypto subnet with a discount window. What disappeared this week is the middle ground — the monthly plan, the committed relationship, the provider you were locked to.
The price index: scissoring is now the pattern
The Monday Price Index delta landed with the week's most concrete numbers: six of seventeen OpenRouter rows moved between November 9 and 16, and every single mover scissored — input and output legs moving in opposite directions. DeepSeek V4.1 Flash jumped from $0.051/$0.567 to $0.14/$0.42 per million tokens, a 175% input increase paired with a 26% output cut, its fifth distinct price shape in five observations. Kimi K2.5 fell to $0.50/$2.50, giving back most of the previous week's output increase after exactly one observation. And Llama 4 Maverick's 87% input jump from November 9 held for a second consecutive observation — the first upward move in this index to pass the two-observation test since issue two.
The scissor pattern matters more than any single row. When providers move input and output prices in opposite directions, they are repositioning for workload shapes — agent loops with long context and short responses price differently from generation-heavy workloads — and the router layer above them reprices around those moves. A week in which every mover scissored is a week in which nobody simply raised or lowered a price; they all restructured. The deeper context is that no first-party pricing-page change has been verified for any of the moved rows — these remain the aggregator's observations, not official list prices, which is exactly the kind of distinction the tracker's methodology exists to enforce.
Into that repricing market, Hyperbolic's relaunched per-token API prices 11 open models from $0.06/$0.18 (DeepSeek V4 Flash) to $3.00/$15.00 (Kimi K3) — a spread that shows how wide the open-model price band has become even before the week's scissoring.
The weights: bigger numbers, smaller licenses
Five open-weight releases landed this week, and the pattern across them is consistent enough to be a finding.
LongCat-2.5 is the headline: a 1.6-trillion-parameter multimodal MoE with roughly 48B active parameters per token, released under MIT by Meituan's LongCat team, with the weights on Hugging Face and the API priced at commodity levels. The genuinely load-bearing claim is the hardware one — Meituan says training and long-context optimization ran on domestic Chinese accelerators with under 80 GB of memory per device, which, if verified, would matter to every export-control argument running through this beat. It is not verified. No independent audit of the training cluster exists, and the claim sits in exactly the evidentiary category last week's hub spent its length describing: a vendor report about hardware nobody outside the company can inspect.
Tencent's Hy4 flash is the licensing case study: a 220B-parameter MoE with 12B active per token and a 1,048,576-token context window, shipped under the Tencent Hy License — derived from Apache 2.0 but adding commercial carve-outs requiring a separate written license for model-as-a-service businesses above $50 million in gross revenue or products above 100 million monthly active users. The X discourse ran ahead of the license facts, describing the release as blanket Apache 2.0; the actual repo license is the part builders should read, and the license-check pattern it continues is now the norm rather than the exception for Chinese lab releases.
Microsoft's Aion 1.0 Instruct is the distribution story inside a weights story: a 2.3B-parameter model under MIT, published in PyTorch, ONNX, and GGUF, and simultaneously made the built-in small language model of Microsoft Edge with an origin-trial Prompt API. The weights are downloadable today; the benchmarks — MMLU 55.15, IFEval 83.90 — are Microsoft's own evaluations, and the plan to make an optimized version the Windows built-in model by end of January 2027 is a vendor statement about the future. What is real is the channel: the default browser on the default desktop OS now ships an open-weights model, and the Prompt API makes it callable from any website's JavaScript.
Cloudflare's Clairic points the same direction from the serving side: a 27B and a 9B task-specific extraction model with open weights and published Workers AI pricing at $0.24 and $0.10 per million input tokens. The latency and accuracy tables are Cloudflare's internal evaluations — the company itself labels them directional — but the shape is the signal: a hyperscaler shipping narrow, priced, open-weight task models instead of a general assistant.
Runway's Praxis-1 rounds the week out at 27B under Apache 2.0 — a genuine Apache 2.0, one of the week's only ones — a world action model for robot control that runs down to an RTX 5090 and, per Runway, to Jetson Thor on the robot itself. Runway claims the top world-action-model spot on NVIDIA's RoboLab-120 at 46.2%; independent ranking puts the model fifth overall, and that gap between leaderboard framing and leaderboard position is itself the week's recurring theme in miniature.
The counter-current: private and self-hosted inference got real plumbing
While the router layer consolidated, the private-inference layer shipped actual infrastructure — and for once, most of it is checkable.
Vercel AI Gateway's Private Inference lets a team register its own self-hosted or fine-tuned endpoints as private providers, callable as private/<slug> through the same auth, SDKs, and logs as catalog models, with optional per-provider Zero Data Retention and CLI provisioning. The changelog is live and the mechanism is verified; the interesting question is what it says about the gateway business, which now happily routes traffic to hardware it does not own.
On the decentralized side, two TEE stories landed in the same week and deserve to be read as one. Solana Confidential Compute — an open-source SDK and onchain program, built by Nito with the Solana Foundation — verifies Intel TDX and AMD SEV-SNP attestations for machines running AI workloads and publishes the record onchain for apps and agents to inspect before trusting an answer. It is a devnet beta, unaudited, with GPU attestation still in progress and no mainnet date. Days later, Caffeine.ai on the Internet Computer began letting developers route apps to SEV-SNP confidential subnets, per Dominic Williams — where the TEE capability is documented but what the cloud picker actually exposes and what the attestation covers remain open questions. Both are early. Both, unlike most of this week's claims, publish their verification mechanism for anyone to run. The underlying explainer covers what these attestations do and do not prove; the pillars on private inference and zero data retention frame why hardware attestation is becoming the trust primitive of choice.
The self-hosting counterweight also sharpened. llama-modchip's claim that DeepSeek's 284B V4-Flash runs CPU-only on a 64GB machine at 9–14.7 tokens per second remains exactly what its title says — a claim, with every figure author-reported and no independent reproduction — but the repo is public, which makes it checkable in principle, which is more than can be said for most of the week's marketing numbers. Against it sits Mistral's Vibe Desktop, a macOS and Windows app that runs coding agents locally and remotely at the same time — verified live and downloadable, and evidence that a hosted lab now considers local execution a product surface rather than a concession.
And the decentralized infrastructure layer moved money, not just code: Akash's one-way AKT migration to Solana went live with mainnet contracts deployed, after governance approved the move on November 2 with 97.27% support. The marketplace cutover — the part that actually changes provider billing and staking — still requires a second governance vote, so the decentralized inference landscape itself is unchanged this week even as its settlement layer migrates chains.
The trust ledger: 127 entries, zero verified
The tracker's eighth sweep is the week's least glamorous and most important story. Cycle 8 appended 23 history entries to 13 of 17 tracked providers — tying the largest single write — and set a funnel record: 18 checks attempted a real observation, and none produced usable evidence. Nine direct probes of Chutes' endpoint returned HTTP 429, four Venice probes hit HTTP 402, two attestation checks found no validatable attestation published, two pricing fetches had no machine-readable source, and five OpenRouter pulls confirmed no cross-check surface. Every zero has a named reason, which is the point: the cumulative ledger now holds 127 entries across 17 providers, and the grade census still reads 0 verified and 2 corroborated.
Against that ledger, Karakeep 0.35 shipped security hardening — seven GitHub advisories fixed, including a high-severity path traversal, after what the maintainer called "Black October" — while Infisical quietly paywalled custom roles in a patch release, catching self-hosters through automated dependency upgrades. Both are the same ledger from the self-hosted side: what a provider controls, it can change, and only continuous checking notices. The tracker export remains the only place in this beat where the word "verified" has a definition.
What to watch next week
The week-ahead brief for the window just closing flagged exactly one dated watch item — the Turing Distributed Compute Network mainnet beta claimed for November 11 — which did not materialize into coverage this week; the claim resolves the moment the network launches or the date passes silently. The next fixed date on the board remains the EU AI Act's Article 50(2) synthetic-content marking deadline on December 2, with compliance announcements expected to cluster ahead of it — a cluster last week's hub already saw the first member of in xAI's default-on watermarking. Watch three things instead: whether OpenRouter publishes an audited token-share breakdown now that two of the industry's largest players cite its numbers in public; whether Chutes' discount-window expiry on November 23 produces the first decentralized-network price move the Price Index can observe; and whether the next Monday's price delta confirms or breaks the scissor pattern — every moved row restructuring again, or someone simply raising a price like it is 2024.
Deep dives
- Hugging Face routes LLM calls through OpenRouter; open weights 71.7% — the router switch that defines the week.
- Price Index Delta: Six Rows Scissor, DeepSeek Reshapes Flash — every mover restructured, none verified a first-party change.
- Chutes ends subscription renewals, moves new models to pay-as-you-go — the decentralized subnet's billing reset.
- Nvidia reaffirms open-source AI strategy and $12.9B Hugging Face deal — the interim CEO keeps the commitments on the record.
- Meituan ships LongCat-2.5: 1.6T-param open-weights agent model, MIT — frontier-scale agent weights, and an unaudited Ascend-only training claim.
- Microsoft Aion 1.0 Instruct: MIT open weights and a Prompt API in Edge — an open-weights model as the default browser's built-in.
- Cloudflare opens Clairic weights: 27B and 9B extraction models — task-specific open weights with published serving prices.
- Tencent ships Hy4 flash open weights: 220B MoE, 12B active — the license is the story, not the parameter count.
- Runway ships Praxis-1 open weights: 27B robot model under Apache 2.0 — the week's only clean Apache 2.0 release.
- Mistral's Vibe Desktop runs agents locally or remotely — a hosted lab shipping local execution as a feature.
- Vercel AI Gateway adds Private Inference for self-hosted endpoints — gateways now route to hardware they do not own.
- DeepSeek V4-Flash on a 64GB CPU box: a claim, not a result — checkable in principle, unreproduced in fact.
- Solana Confidential Compute: TEE attestations verified onchain and Caffeine.ai routes apps to SEV-SNP confidential subnets — two TEE stories, one verification trend.
- Akash's one-way AKT migration to Solana is live — the settlement layer moves; the marketplace does not yet.
- Sweep Cycle 8: Record Check Attempts, Zero Usable Observations — the trust ledger's least flattering week.
Questions
- What changed with Hugging Face Inference Providers this week?
- On November 16, 2026, Hugging Face switched language-model calls on Inference Providers to route through OpenRouter, which picks a provider per request on live price, speed, and reliability. Bring-your-own provider keys stopped working for language models; custom keys still work for image and video.
- What share of OpenRouter tokens do open-weight models serve?
- OpenRouter's published data through November 16, 2026 puts open-weight models at 71.7% of tokens served on its router. Nvidia's interim CEO cited a 35-fold year-over-year rise in weekly token volume on the same platform — a vendor-reported claim, not an audited figure.
- Did Chutes end its subscription plans?
- Yes. Chutes stopped selling monthly subscriptions and moved new billing to pay-as-you-go credits; existing plans run out at the end of their current period. Qwen3.8 Max and DeepSeek V4.1 Pro in beta are pay-as-you-go only, with a 35% discount the operator says expires November 23.
- Which new open-weight models shipped this week and under which licenses?
- Meituan's LongCat-2.5 (1.6T-parameter MoE, MIT), Runway's Praxis-1 (27B robot model, Apache 2.0), Tencent's Hy4 flash (220B MoE, Tencent Hy License with $50M-revenue/100M-MAU carve-outs), Microsoft's Aion 1.0 Instruct (2.3B, MIT), and Cloudflare's Clairic 27B and 9B extraction models.
- What happened to inference prices this week?
- Six of seventeen DeAI Price Index rows moved on OpenRouter between November 9 and 16, and every mover scissored — input and output legs moved in opposite directions. DeepSeek V4.1 Flash jumped +175% on input to $0.14 while output fell 26% to $0.42.
Sources
- Hugging Face and OpenRouter join forces to make open-weight models easy to run — Hugging Face
- Open-weight vs closed-weight models on OpenRouter (October 2026 insights) — OpenRouter
- OpenRouter rankings (usage data through Nov 16, 2026) — OpenRouter
- Chutes pricing page — model rates, Private Chutes, and billing FAQ — Chutes
- Chutes Rolls Out Qwen3.8 Max and DeepSeek V4.1 Pro as Subscriptions End — Tao.media
- Nvidia says its open source AI strategy remains intact after Jensen Huang's exit — TechCrunch
- Introducing LongCat-2.5 (tech blog) — LongCat (Meituan)
- meituan-longcat/LongCat-2.5 model card — Hugging Face
- Praxis-1 Is Now Available: Open-Weight Robotics Model — Runway
- Hybrid intelligence on the Web: Introducing Aion Instruct and Prompt API in Microsoft Edge — Microsoft Edge Blog
- An updated Clairic and Clairic-flash are now open-weight — Cloudflare Blog
- OpenRouter API — list models — OpenRouter
- AI Gateway now supports Private Inference — Vercel
- Solana Confidential Compute: Private AI, Verified Onchain — Solana Foundation
- Akash Migrate: AKT on Solana — Akash Network docs
- DeAI News tracker, sweep cycle 8 (2026-11-09) — DeAI News
About DeAI
DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.
Powered by Morpheus and StrandCMS
Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.
