
Head of Models @Cloudflare. Founder at @jamdotdev.
Austin, TX
Joined December 2011
- Tweets3.5K
- Following903
- Followers1.6K
- Likes14.4K
Irtefa retweeted
shipped last week at @cloudflare (just a random week in november):
→ durable objects
- improved startup placement: reactivated dos now start on the machine that last ran them (less cold starts for sqlite-backed ones)
- better request routing: fewer round trips on location/forwarding lookups → lower p99 connection latency when relocating
→ spend limits
set spending limits + email alerts for workers and durable objects (block usage or just get notified)
→ kitesurf
now open source (apache 2.0)! the browser engine for ai agents that runs entirely on workers — inspect, run locally, or deploy your own
→ hyperdrive
raised limits on workers paid: 100 configs (was 25) and 1500 origin connections per config (was 100)
→ realtime
- custom validity windows for auth tokens (set not_before / expires_on + single-token revoke)
- noise cancellation package for web apps
- collaborative whiteboard plugin for meetings
→ new workers ai models
- embeddinggemma 2 (multimodal embeddings from google deepmind)
- inkling 2 small (thinking machines’ 594b moe)
→ ai search
- d1 as a data source (beta)
- terraform support for instances in any namespace
plus more (d1 virtual table exports, browser run local quick actions, etc.)
full list / keep an eye on: developers.cloudflare.com/ch…
Irtefa retweeted
7 Quotes Michelle Chen @michellechen at @aiDotEngineer Code Summit
1. "Life can only be understood backwards, but it must be lived forwards."
• @Cloudflare made bets
• Look back to learn, then move forward
2. "Models got larger and larger."
• 2023 bet: models get smaller
• Instead, 7B grew to trillion-parameter models
• So they pivoted to serving giant LLMs
3. "With these small and fast models, the network latency became important again."
• System 1 models: one forward pass
• Small, fast models bring latency back into play
• Edge compute wins again
4. "Cloudflare was not wrong here. We actually were just early."
• Being early can look like being wrong
• Keep the bet ready for when it pays off
• Move fast when it does
5. "Everything is ready here. All that's missing is the east wind."
• A Chinese proverb she shared
• You can't predict the shift
• Be ready to react when it comes
6. "They both ended up getting to the answer at the end, so the quality was not affected."
• All-LLM agent vs System 1 + LLM
• Almost 2x faster, 72% cheaper
• 94% fewer tokens
7. "We're not limited to just LLMs."
• Route decisions to small models
• Extract typed JSON from messy input
• Save the LLM for taking action
Irtefa retweeted
inkling 2 small is available on workers ai! excited to host a frontier US model, and our first from @thinkymachines
developers.cloudflare.com/wo…
Inkling 2 Small’s weights are now available for download on HuggingFace under an Apache 2.0 license.
Make Inkling your own!
huggingface.co/thinkingmachi…
Irtefa retweeted
Inkling 2 Small, now available on Cloudflare Workers AI.
Thinking Machines used this 594B-parameter model to autonomously train Qwen 3.5-4B on Tinker to forecast weather and Wikipedia page views. In 4.5 hours, it more than doubled the model’s forecasting score.
Try it yourself: developers.cloudflare.com/wo…
nitter.cf/thinkymachines/status/…
Inkling 2 Small’s weights are now available for download on HuggingFace under an Apache 2.0 license.
Make Inkling your own!
huggingface.co/thinkingmachi…
Irtefa retweeted
i see a lot of you are becoming durable objects pilled. the growth for DO has been incredible
we’re committed to making it the best primitive for running agents. seen many people asking for these improvements. enjoy better placement and fewer cold start
I'm new to the Durable Objects team, and I've been seriously amazed by how locked in everyone is on making DOs and Workers the best platform to build on.
Two big wins delivered from the team - together, these help your app & agent get to work sooner.
Both of these improvements are live, with no code changes needed.
fewer cold starts:
developers.cloudflare.com/ch…
improved relocation request routing:
developers.cloudflare.com/ch…
Irtefa retweeted
thank you for laughing at my chinese girl in ai joke and being receptive to my philosophy larp
.@michellechen on how @Cloudflare became a NeoLab.
All it took is for the right piece to come along. That piece, JEV from @typesafeai 👏
Cloudflare got their "east wind" and then trained Clef and Clairic and started serving them on their own inference.
Super fast!
Muse Ember from @AIatMeta is now available on @Cloudflare AI Gateway. With top tier speed and performance in its class, the model is efficient and a great fit for subagent workloads.
developers.cloudflare.com/ai…
1
15
761
Cloudflare’s Matthew Prince brilliant thoughts on neoclouds in their most recent 3Q26 earnings call:
”We are an engineering company. We are not a financial engineering company...
We are focused on how we can actually deliver real operational excellence…
You can’t sell $1 for $0.50 and be in business over the long term.”
Irtefa retweeted
Google's new video model from @GoogleDeepMind
- 15-second clips, up from 10
- More natural dialogue and lip-sync
- Characters stay consistent across shots
- Edit the same clip with a new prompt































