Reminds of the TikTok ban for security and safety only for it to be sold to a fellow crony. Can't help but see this play going down again. Threaten / Ban / Control / Pressure a technology+company, then get your cronies a seat at the board.
The cynic in me suspects they were salivating so much over the Spacex IPO they wanted a finger in anthropics 2026 IPO. Banning fable ~1 day after.
I find it ironic, we now have to use lesser models to write potentially MORE buggy code, than greater models which would allow you to write LESS buggy code. It's paradoxical.
Looks like we are seeing small but mighty model breakthroughs, outpacing the pure capital firepower of SOTA providers. I love rooting for the little guy, but is it too soon to call it? To play devils advocate, could it just be the benchmarks are not efficient enough to capture success of real developer workflows?
I think people are going to continue to be surprised by the capability of small models.
Now, if you ask this model to have a conversation with you, it's gonna fail and be incoherent. But boy, does it sure reason through math problems well.
I've just started using qwen3.6:35b a couple days ago running on my framework desktop and rather impressed. It runs really well and reminds me of probably the first Claude model I used. It's the first local model that's actually working for me in a coding agent I've tried. Very exciting!
Interesting. I wonder if there is opportunity to train a set of small model variants to excel at a certain stacks. Eg Qwen3.6-27B for Node + React or Qwen3.6-27B for Rust + TUI
This is always how I've imagined small/consumer-hardware models going in time. If I only ever code in Python, give me a model that does just that (plus some general CS, algorithms, structure, etc.) and does it super-fast and well. Make it small enough that if I need a Python back end and an HTML front end, another specific model can load alongside and collaborate on the front end.
Or give me a pure shopping model that has a general understanding of products and product categories, and then will playwright/scrape/API into shopping sites to compare options and find me what I want. Etc.
Qwen 3.6 27B is an anomalously strong all-around model for its size, but when we run our evaluations, we generate 10 coding submissions/language/model (110 total). So full discosure, the per-language per-model performances can be noisy (I do not think Qwen3.6 27B is better than Fable 5 in agentic workflows when writing Kotlin, given enough samples, although we do find some interesting anomalies that hold up under large sample sizes).
It is, but it's way too slow on a Strix Halo due to its limited bandwidth.
(I'm still sad that they didn't make a 122B-A10B version of it, as it's the kind of model that fits best on a Strix Halo, and for 3.5 it was comparable in performance to the dense 27B version).
Yeah the speed is vastly different but it's getting ~10 tps. And the MoE model is like 50 or something.
Might use it if it proves to be much smarter I dont really monitor my agents while they're running.
Off the top of my head since it seems to be the quick info you're looking for: IIRC, with these two, the 27B is a dense model, meaning it's all active at inference. Meanwhile, the 35B is a Mixture of Experts (MoE), so only part of its network (3B?) is active at any time.
Thanks! Dense models have been slow on my compute, but I'll give it a try. If its not toooooo slow then it's fine I mostly fire and forget agents anyway.
Edit: seems fast! I'll try it out some more, thanks again.
I'm running qwen36.:35b:iq4 IQ4_XS quant. Takes 18 GB of RAM with 131k context window. Seems to be really good. Have it running local stuff via Hermes, using a cloud model via Ollama (Deepseek V4-Pro) for heavy lifting.
I settled on opencode after trying goose and aider as well. I'll probably try some more but opencode worked similar to Claude code which is my main agent.
I serve the model with ollama and am thinking about replacing ollama but haven't looked into it.
I have openwebui for chat if I want that too, but don't really use it.
I’m beginning to suspect the closed SOTA labs were doing all these optimisations, keeping quiet about it, and just charging us out the yinyang for inference.
> The catch is that regular EC2 is already a VM. AWS runs our host inside its own isolation layer, and then we run browser VMs inside that host. In other words, every browser is a VM inside a VM.
yes but i think there is specifically some ec2s which give you hypervisor access and thereby firecracker too - someone correct me if im wrong?
Unfortunately supply is quite limited. If you want to horizontally scale on these instances you need to have a good relationship with AWS so they'll give you a big allocation before c9i is a thing.
I haven't personally tried, so I can't say for certain, but Lambda has publicly stated they run on bare metal EC2 instances, presumably the supply of whatever instance types they use should be fairly healthy
The interesting part to me is less the exact hardware generation and more the control plane around placement, isolation, and startup latency. That is hard to copy outside AWS.
When we had need of quite big machines (AWS metal instances), we've found the performance differential between metal, and the equivalent size VM was 10-20% for CPU heavy workloads.
They just don't have access to giant pools of residential IPs, so too many sites end up blocking all the cloud providers by IP range/ASN anyway, even if they could get through a captcha.
google has a large amount of "caching servers (GGC)" located in data centers for residential providers all over the world.. They use these servers for a variety of services.. Most of the traffic I have seen from them have been for their "URL preview" service ..
they kind of do.. gcp has their lambda equivalent which i believe comes with chromium preinstalled, its how major search tools like jina work, sure thre problaby somethign about session management that they probably neuter to prevent abuse though
I wonder if with enough input data and transcription you could “fingerprint” where a speaker personality has habits of interjecting “ums” leading to more hardy analysis. Novel approach, but gets me thinking
Not yet. I have a PR with a WASM experiment based on an earlier build, but it's not integrated. It's on my list of things to try. I _did_ get it working for some things, so it's clearly possible, but I need to put some more effort into it.
Maybe it’s misplaced nostalgia but google search before google plus ruined the +, and the image search was top tier (not shopping ads) - truly incredible the signal to noise you could find.
Now all the junk comes to the top and the sites you get all have ads and modal popups or sales funnel flows
Of course - it's SEO's fault that Google search frequently ignores words in your queries.
It's SEO's fault that Google frequently prioritizes giving you results that are semantically or thematically related but ultimately irrelevant to what you searched for.
It's SEO's fault that Google refuses to place ads on websites not full of meandering irrelevant "content".
It's SEO's fault Google search all-but-ignores most single-purpose tool websites.
It's not the result of decisions made by Google's programmers and project managers - it's allllll the fault of SEO!
The SEO industry definitely isn't an all-purpose scapegoat for horribly unpopular decisions made by human beings at Google.
