September 23, 2026
Report summary
11 stories cleared the bar, led by Python Workers are now generally available, Xiaomi MiMo V2.6 (Pro, Flash, UltraSpeed) now callable via Vercel AI Gateway, and n8n 2.40.5 / 2.39.10: backported fix limiting declarative routing during base-URL ownership checks.
Worth attention
Cloudflare says Python Workers are now generally available, so Python web frameworks and AI orchestration libraries run natively on the Workers runtime with direct access to D1, R2 and Workers AI, with no JavaScript glue needed. For a solo builder this makes a Python-first edge deployment a supported production option rather than a beta experiment. If you run Python glue services on a VPS, check whether they fit Workers limits before your next deploy.
Xiaomi's MiMo V2.6 family is available on Vercel AI Gateway: Pro (1.02T total / 42B active MoE), Flash (309B / 15B active) and Pro UltraSpeed (claimed up to 20x output speed). All have a 1M-token context, up to 128K output tokens, structured outputs and multimodal input. Latent Space reports Pro as the new top open-weights model, trained for about $3M. If you route agent or coding work through a gateway, it costs little to add as a candidate and benchmark against your current default.
n8n shipped same-day patch releases on both the 2.40 and 2.39 lines with one fix: "Limit declarative routing during base URL ownership checks". A single-fix backport to two release lines usually means a security hardening fix, probably around request routing or SSRF. If you self-host n8n, update to 2.40.5, 2.39.10 or 2.41.0.
Vercel deployments now show billable build duration (build plus post-build time, rounded up to the minute) and CPU minutes (duration × vCPUs) in the dashboard, in `vc inspect` (CLI 59.23.1+) and in the REST API. This helps you find which builds are eating your usage.
LWN previews Git 2.56 and the upcoming Git 3.0, the release where long-planned breaking default changes land. Title-based only: the article was not read. Skim it before 3.0 ships so CI scripts and tooling that assume old defaults don't break.
Hugging Face published a post on tokenizers v1, the first major version of its widely used tokenization library, with measured encode/decode and scaling performance. Title-based only; no content was fetched. It matters mainly if you run tokenization in your own pipelines.
Vercel Connect now has a managed Microsoft Teams connector. Vercel registers the Entra app and Azure Bot in your tenant, so you don't have to store a client secret, and forwards verified Teams messages to your project so an agent can reply as a bot. This lowers the effort of shipping an agent into enterprise Teams channels, a possible distribution route for B2B agent products.
Vercel AI Gateway now proxies Jev from TypeSafe AI, a "probabilistic decision model" that returns typed boolean, choice or score answers with probabilities instead of generated text, through the TypeSafe client, an HTTP API (POST /v1/evaluate) or the AI SDK. Vercel's eve agent framework uses it for model selection and tool approvals. It could be interesting for agent gating decisions, but it's early and vendor-driven.
A builder's write-up on self-hosting services behind carrier-grade NAT, where you can't port-forward. Title-based only. Useful if you plan to run anything from a home connection.
Grok 4.7 (500K context, four reasoning levels) is on Vercel AI Gateway at 40% off through September 27. The discount is time-limited, so only relevant if you are already evaluating it.
Jun Kim, creator of oMLX, has joined Hugging Face to support the MLX (Apple Silicon) community. This is a signal that local inference on Macs will get more investment. It doesn't call for any action yet.
Full digest
Latent Space AINews covers Xiaomi MiMo-V2.6-Pro (1T total, 42B active) as the new top open-weights model, reportedly trained for $3M.
R
Jev: System One models for Prod, not God
with Diogo Almeida, CEO, TypeSafe AI — https://www.latent.space/p/jev — A Latent Space podcast with TypeSafe AI's CEO about the Jev model.
Jun Kim, creator of oMLX, has joined Hugging Face to support the MLX (Apple Silicon) community. This is a signal that local inference on Macs will get more investment. It doesn't call for any action yet.
A research blog post framing LLM block pruning as an Ising optimization problem.
Hugging Face published a post on tokenizers v1, the first major version of its widely used tokenization library, with measured encode/decode and scaling performance. Title-based only; no content was fetched. It matters mainly if you run tokenization in your own pipelines.
n8n 2.41.0 is a bug-fix rollup: AI builder fixes, multi-main publishing, Vault reads, and execution status guards.
R
beta
The n8n "beta" tag release duplicates 2.41.0.
R
stable
The n8n "stable" tag release duplicates 2.40.5.
P
n8n 2.40.5 / 2.39.10: backported fix limiting declarative routing during base-URL ownership checks
n8n shipped same-day patch releases on both the 2.40 and 2.39 lines with one fix: "Limit declarative routing during base URL ownership checks". A single-fix backport to two release lines usually means a security hardening fix, probably around request routing or SSRF. If you self-host n8n, update to 2.40.5, 2.39.10 or 2.41.0.
n8n 2.39.10 is the same base-URL routing fix, backported to 2.39.
Cloudflare says Python Workers are now generally available, so Python web frameworks and AI orchestration libraries run natively on the Workers runtime with direct access to D1, R2 and Workers AI, with no JavaScript glue needed. For a solo builder this makes a Python-first edge deployment a supported production option rather than a beta experiment. If you run Python glue services on a VPS, check whether they fit Workers limits before your next deploy.
A Google AI changelog entry that is only a Korean-locale "try it" link to gemini-3.8-flash in AI Studio, with no content.
Vercel Connect now has a managed Microsoft Teams connector. Vercel registers the Entra app and Azure Bot in your tenant, so you don't have to store a client secret, and forwards verified Teams messages to your project so an agent can reply as a bot. This lowers the effort of shipping an agent into enterprise Teams channels, a possible distribution route for B2B agent products.
Vercel deployments now show billable build duration (build plus post-build time, rounded up to the minute) and CPU minutes (duration × vCPUs) in the dashboard, in `vc inspect` (CLI 59.23.1+) and in the REST API. This helps you find which builds are eating your usage.
Xiaomi's MiMo V2.6 family is available on Vercel AI Gateway: Pro (1.02T total / 42B active MoE), Flash (309B / 15B active) and Pro UltraSpeed (claimed up to 20x output speed). All have a 1M-token context, up to 128K output tokens, structured outputs and multimodal input. Latent Space reports Pro as the new top open-weights model, trained for about $3M. If you route agent or coding work through a gateway, it costs little to add as a candidate and benchmark against your current default.
Vercel AI Gateway now proxies Jev from TypeSafe AI, a "probabilistic decision model" that returns typed boolean, choice or score answers with probabilities instead of generated text, through the TypeSafe client, an HTTP API (POST /v1/evaluate) or the AI SDK. Vercel's eve agent framework uses it for model selection and tool approvals. It could be interesting for agent gating decisions, but it's early and vendor-driven.
Grok 4.7 (500K context, four reasoning levels) is on Vercel AI Gateway at 40% off through September 27. The discount is time-limited, so only relevant if you are already evaluating it.
A builder's write-up on self-hosting services behind carrier-grade NAT, where you can't port-forward. Title-based only. Useful if you plan to run anything from a home connection.
An opinion essay on language design that argues for named and optional arguments.
rift, an open-source tiling window manager for macOS.
A review of Paul McKenney's concurrency textbook.
LWN previews Git 2.56 and the upcoming Git 3.0, the release where long-planned breaking default changes land. Title-based only: the article was not read. Skim it before 3.0 ships so CI scripts and tooling that assume old defaults don't break.
The README for an experimental, artistically framed computation project.
Comments
Comments
Comments
I miss classic forums: slow, asynchronous threads that stay readable and searchable for years, instead of knowledge dissolving into chat scr…
Comments
Comments
Comments
What are you doing this week? Feel free to share! Keep in mind it’s OK to do nothing at all, too.
Comments
Comments
Comments
Comments
Comments
Comments
Xiaomi's MiMo v2.6 announcement page, trending on Hacker News.
Original markdown
# Nightly Librarian — Newsletter draft Run: 6ac204db-87f9-4f7e-b205-22e30640552a Started: 2026-09-23T06:03:05.045Z Completed: 2026-09-23T06:05:18.645Z ## Worth attention - **Python Workers are now generally available** https://blog.cloudflare.com/python-workers-ga/ Cloudflare says Python Workers are now generally available, so Python web frameworks and AI orchestration libraries run natively on the Workers runtime with direct access to D1, R2 and Workers AI, with no JavaScript glue needed. For a solo builder this makes a Python-first edge deployment a supported production option rather than a beta experiment. If you run Python glue services on a VPS, check whether they fit Workers limits before your next deploy. - **Xiaomi MiMo V2.6 (Pro, Flash, UltraSpeed) now callable via Vercel AI Gateway** https://vercel.com/changelog/mimo-v2-6-models-now-available-on-ai-gateway Xiaomi's MiMo V2.6 family is available on Vercel AI Gateway: Pro (1.02T total / 42B active MoE), Flash (309B / 15B active) and Pro UltraSpeed (claimed up to 20x output speed). All have a 1M-token context, up to 128K output tokens, structured outputs and multimodal input. Latent Space reports Pro as the new top open-weights model, trained for about $3M. If you route agent or coding work through a gateway, it costs little to add as a candidate and benchmark against your current default. - **n8n 2.40.5 / 2.39.10: backported fix limiting declarative routing during base-URL ownership checks** https://github.com/n8n-io/n8n/releases/tag/n8n%402.40.5 n8n shipped same-day patch releases on both the 2.40 and 2.39 lines with one fix: "Limit declarative routing during base URL ownership checks". A single-fix backport to two release lines usually means a security hardening fix, probably around request routing or SSRF. If you self-host n8n, update to 2.40.5, 2.39.10 or 2.41.0. - **Deployments now show billable duration and CPU minutes** https://vercel.com/changelog/deployments-now-show-billable-duration-and-cpu-minutes Vercel deployments now show billable build duration (build plus post-build time, rounded up to the minute) and CPU minutes (duration × vCPUs) in the dashboard, in `vc inspect` (CLI 59.23.1+) and in the REST API. This helps you find which builds are eating your usage. - **Looking forward to Git 2.56 - and 3.0** https://lwn.net/SubscriberLink/1094575/2385e98583715c2b/ LWN previews Git 2.56 and the upcoming Git 3.0, the release where long-planned breaking default changes land. Title-based only: the article was not read. Skim it before 3.0 ships so CI scripts and tooling that assume old defaults don't break. - **tokenizers v1: encode, decode and scaling, measured** https://huggingface.co/blog/tokenizers-v1 Hugging Face published a post on tokenizers v1, the first major version of its widely used tokenization library, with measured encode/decode and scaling performance. Title-based only; no content was fetched. It matters mainly if you run tokenization in your own pipelines. - **Vercel Connect now supports Microsoft Teams** https://vercel.com/changelog/vercel-connect-microsoft-teams Vercel Connect now has a managed Microsoft Teams connector. Vercel registers the Entra app and Azure Bot in your tenant, so you don't have to store a client secret, and forwards verified Teams messages to your project so an agent can reply as a bot. This lowers the effort of shipping an agent into enterprise Teams channels, a possible distribution route for B2B agent products. - **AI Gateway now supports TypeSafe clients and an HTTP API for Jev** https://vercel.com/changelog/ai-gateway-now-supports-typesafe-clients-and-http-api-for-jev Vercel AI Gateway now proxies Jev from TypeSafe AI, a "probabilistic decision model" that returns typed boolean, choice or score answers with probabilities instead of generated text, through the TypeSafe client, an HTTP API (POST /v1/evaluate) or the AI SDK. Vercel's eve agent framework uses it for model selection and tool approvals. It could be interesting for agent gating decisions, but it's early and vendor-driven. - **Self-Hosting Behind CGNAT** https://david.alvarezrosa.com/posts/self-hosting-behind-cgnat/ A builder's write-up on self-hosting services behind carrier-grade NAT, where you can't port-forward. Title-based only. Useful if you plan to run anything from a home connection. - **Grok 4.7 now available and 40% off on AI Gateway, fx, and eve** https://vercel.com/changelog/grok-4-7-now-available-and-40-off-on-ai-gateway-fx-eve Grok 4.7 (500K context, four reasoning levels) is on Vercel AI Gateway at 40% off through September 27. The discount is time-limited, so only relevant if you are already evaluating it. - **Jun Kim, oMLX creator and maintainer, joins Hugging Face to support the MLX community** https://huggingface.co/blog/omlx Jun Kim, creator of oMLX, has joined Hugging Face to support the MLX (Apple Silicon) community. This is a signal that local inference on Macs will get more investment. It doesn't call for any action yet. ## Full digest - [R] [latent-space] [AINews] Xiaomi MiMo-V2.6-Pro 1T-A42B: the new top Open Weights model, trained for $3M — https://www.latent.space/p/ainews-xiaomi-mimo-v26-pro-1t-a42b — Latent Space AINews covers Xiaomi MiMo-V2.6-Pro (1T total, 42B active) as the new top open-weights model, reportedly trained for $3M. - [R] [latent-space] Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI — https://www.latent.space/p/jev — A Latent Space podcast with TypeSafe AI's CEO about the Jev model. - [M] [huggingface-blog] Jun Kim, oMLX creator and maintainer, joins Hugging Face to support the MLX community — https://huggingface.co/blog/omlx — Jun Kim, creator of oMLX, has joined Hugging Face to support the MLX (Apple Silicon) community. This is a signal that local inference on Macs will get more investment. It doesn't call for any action yet. - [R] [huggingface-blog] Pruning LLMs Like a Physicist: Block Removal as an Ising Optimization Problem — https://huggingface.co/blog/MultiverseComputingCAI/pruning-llms-like-a-physicist-block-removal-as-an — A research blog post framing LLM block pruning as an Ising optimization problem. - [P] [huggingface-blog] tokenizers v1: encode, decode and scaling, measured — https://huggingface.co/blog/tokenizers-v1 — Hugging Face published a post on tokenizers v1, the first major version of its widely used tokenization library, with measured encode/decode and scaling performance. Title-based only; no content was fetched. It matters mainly if you run tokenization in your own pipelines. - [R] [gh-n8n] [email protected] — https://github.com/n8n-io/n8n/releases/tag/n8n%402.41.0 — n8n 2.41.0 is a bug-fix rollup: AI builder fixes, multi-main publishing, Vault reads, and execution status guards. - [R] [gh-n8n] beta — https://github.com/n8n-io/n8n/releases/tag/beta — The n8n "beta" tag release duplicates 2.41.0. - [R] [gh-n8n] stable — https://github.com/n8n-io/n8n/releases/tag/stable — The n8n "stable" tag release duplicates 2.40.5. - [P] [gh-n8n] n8n 2.40.5 / 2.39.10: backported fix limiting declarative routing during base-URL ownership checks — https://github.com/n8n-io/n8n/releases/tag/n8n%402.40.5 — n8n shipped same-day patch releases on both the 2.40 and 2.39 lines with one fix: "Limit declarative routing during base URL ownership checks". A single-fix backport to two release lines usually means a security hardening fix, probably around request routing or SSRF. If you self-host n8n, update to 2.40.5, 2.39.10 or 2.41.0. - [R] [gh-n8n] [email protected] — https://github.com/n8n-io/n8n/releases/tag/n8n%402.39.10 — n8n 2.39.10 is the same base-URL routing fix, backported to 2.39. - [P] [cloudflare-blog] Python Workers are now generally available — https://blog.cloudflare.com/python-workers-ga/ — Cloudflare says Python Workers are now generally available, so Python web frameworks and AI orchestration libraries run natively on the Workers runtime with direct access to D1, R2 and Workers AI, with no JavaScript glue needed. For a solo builder this makes a Python-first edge deployment a supported production option rather than a beta experiment. If you run Python glue services on a VPS, check whether they fit Workers limits before your next deploy. - [R] [google-ai-changelog] Google AI Studio: "Try it" link for gemini-3.8-flash (Korean locale) — https://aistudio.google.com/prompts/new_chat?model=gemini-3.8-flash&hl=ko — A Google AI changelog entry that is only a Korean-locale "try it" link to gemini-3.8-flash in AI Studio, with no content. - [P] [vercel-changelog] Vercel Connect now supports Microsoft Teams — https://vercel.com/changelog/vercel-connect-microsoft-teams — Vercel Connect now has a managed Microsoft Teams connector. Vercel registers the Entra app and Azure Bot in your tenant, so you don't have to store a client secret, and forwards verified Teams messages to your project so an agent can reply as a bot. This lowers the effort of shipping an agent into enterprise Teams channels, a possible distribution route for B2B agent products. - [P] [vercel-changelog] Deployments now show billable duration and CPU minutes — https://vercel.com/changelog/deployments-now-show-billable-duration-and-cpu-minutes — Vercel deployments now show billable build duration (build plus post-build time, rounded up to the minute) and CPU minutes (duration × vCPUs) in the dashboard, in `vc inspect` (CLI 59.23.1+) and in the REST API. This helps you find which builds are eating your usage. - [P] [vercel-changelog] Xiaomi MiMo V2.6 (Pro, Flash, UltraSpeed) now callable via Vercel AI Gateway — https://vercel.com/changelog/mimo-v2-6-models-now-available-on-ai-gateway — Xiaomi's MiMo V2.6 family is available on Vercel AI Gateway: Pro (1.02T total / 42B active MoE), Flash (309B / 15B active) and Pro UltraSpeed (claimed up to 20x output speed). All have a 1M-token context, up to 128K output tokens, structured outputs and multimodal input. Latent Space reports Pro as the new top open-weights model, trained for about $3M. If you route agent or coding work through a gateway, it costs little to add as a candidate and benchmark against your current default. - [M] [vercel-changelog] AI Gateway now supports TypeSafe clients and an HTTP API for Jev — https://vercel.com/changelog/ai-gateway-now-supports-typesafe-clients-and-http-api-for-jev — Vercel AI Gateway now proxies Jev from TypeSafe AI, a "probabilistic decision model" that returns typed boolean, choice or score answers with probabilities instead of generated text, through the TypeSafe client, an HTTP API (POST /v1/evaluate) or the AI SDK. Vercel's eve agent framework uses it for model selection and tool approvals. It could be interesting for agent gating decisions, but it's early and vendor-driven. - [M] [vercel-changelog] Grok 4.7 now available and 40% off on AI Gateway, fx, and eve — https://vercel.com/changelog/grok-4-7-now-available-and-40-off-on-ai-gateway-fx-eve — Grok 4.7 (500K context, four reasoning levels) is on Vercel AI Gateway at 40% off through September 27. The discount is time-limited, so only relevant if you are already evaluating it. - [P] [lobsters] Self-Hosting Behind CGNAT — https://david.alvarezrosa.com/posts/self-hosting-behind-cgnat/ — A builder's write-up on self-hosting services behind carrier-grade NAT, where you can't port-forward. Title-based only. Useful if you plan to run anything from a home connection. - [R] [lobsters] Named and Optional Arguments are Awesome — https://botahamec.dev/named-optional-args — An opinion essay on language design that argues for named and optional arguments. - [R] [lobsters] rift - a tiling window manager for macos — https://github.com/acsandmann/rift/ — rift, an open-source tiling window manager for macOS. - [R] [lobsters] Textbook review: Is Parallel Programming Hard, And, If So, What Can You Do About It? — https://ahelwer.ca/post/2026-09-21-concurrency-textbook/ — A review of Paul McKenney's concurrency textbook. - [P] [lobsters] Looking forward to Git 2.56 - and 3.0 — https://lwn.net/SubscriberLink/1094575/2385e98583715c2b/ — LWN previews Git 2.56 and the upcoming Git 3.0, the release where long-planned breaking default changes land. Title-based only: the article was not read. Skim it before 3.0 ships so CI scripts and tooling that assume old defaults don't break. - [R] [lobsters] evocation (experimental computation project README) — https://code.irenes.space/evocation/tree/README.txt — The README for an experimental, artistically framed computation project. - [R] [lobsters] I said no and Apple said yes — https://dbushell.com/2026/09/22/apple-intelligence/ — Comments - [R] [lobsters] Relation algebra is not relational algebra — https://remy.wang/blog/ra-ra.html — Comments - [R] [lobsters] Do you use :visited on links? — https://kevquirk.com/do-you-use-visited-on-links — Comments - [R] [lobsters] Squalk: an old-school forum engine built on Nostr (NIP-29 groups, NIP-7D threads) — https://github.com/dtonon/squalk — I miss classic forums: slow, asynchronous threads that stay readable and searchable for years, instead of knowledge dissolving into chat scr… - [R] [lobsters] Creating a Blog in Gemini:// — https://brennan.day/creating-a-blog-in-gemini/ — Comments - [R] [lobsters] Why 0xCAFEBABE? — https://www.artima.com/insidejvm/whyCAFEBABE.html — Comments - [R] [lobsters] Cheap and easy throwaway VPN server — https://a.l3x.in/blog/throwaway-vpn/ — Comments - [R] [lobsters] What are you doing this week? — https://lobste.rs/s/fyngrv/what_are_you_doing_this_week — What are you doing this week? Feel free to share! Keep in mind it’s OK to do nothing at all, too. - [R] [lobsters] Ju! Ju! Tsu — https://arialdo.codeberg.page/ju-ju-tsu/ — Comments - [R] [lobsters] Markdown in /src — https://htmx.org/essays/markdown-in-src/ — Comments - [R] [lobsters] Attention is all you have — https://alicegg.tech/2026/09/21/attention — Comments - [R] [lobsters] Arguing about arguments — https://steveklabnik.com/writing/arguing-about-arguments/ — Comments - [R] [lobsters] Design your programming languages right (2024) — https://blog.veritates.love/design-it-right — Comments - [R] [lobsters] Making a game for the GBA and PC from the same codebase — https://mattgreer.dev/blog/making-a-game-for-gba-and-pc/ — Comments - [R] [lobsters] Canonical announces Zephyr 26.04 LTS — https://canonical.com/blog/zephyr-lts-announcement — Comments - [R] [lobsters] Running an Optimal Trace — https://anishathalye.com/optimal-trace/ — Comments - [R] [hn-top] MiMo v2.6 — https://mimo.xiaomi.com/mimo-v2-6 — Xiaomi's MiMo v2.6 announcement page, trending on Hacker News.