building the SOTA model router @rolemodeldev https://xitter.cf/t.co/N6odPItHYZ 🇸🇪 pm and indie dev in 🇨🇳 since 2010. Prev: @dji @oppo @oneplus
Shanghai
Joined March 2026
- Tweets7K
- Following929
- Followers629
- Likes3.5K
Pinned Tweet
role-model is here: a model router for Pi that works with (almost) any cloud model as well as local models.
Set up your pool of models, whether GPT 5.5 and DeepSeek V4, or Qwen 27B and Qwen 35B running locally, or combine and route between your local model for general chat and tool use and cloud models for more difficult tasks.
role-model is a router runtime and and a forward-looking routing protocol that can enable consumer applications like Pi to classify its requests with required capabilities, task-type and suitable roles for better matching between task, role, and model.
pi-role-model is a package for Pi that lets Pi manage and interact with role-model. role-model works perfectly with Pi even without the package - just configure your models in role-model, then configure role-model as your provider in Pi and choose routing alias.
Install the package for Pi:
’pi install npm:@try-works/pi-role-model’
Then let your agent install and configure the role-model runtime or point it to:
‘github.com/try-works/role-mo…’
role-model is in an early alpha release - let me know bugs, issues or usability questions. also, tell me if it just happens to work perfectly for you!
there's lots more to do and you can help by installing and running role-model, and by reposting this tweet!
thx
try.works retweeted
DeepSeek V4.1 Flash Ultrafast (1500 token/s) is a completely new experience, it feels a lot more like chatting and real-time interaction than any other Agentic system.
Watch the turn stats on the top right!
I got some NVL72 GB300 available for some people to test as well. Will drop tomorrow :)
Spent 3 days letting DS 1017 migrate an app to Effect and Foldkit following a plan created by 6.1 Sol, a test harness and so on. Didn't work well - noticed that DS was tunnel visioned on some random tests. Switched back to 6.1 Sol to reset.
try.works retweeted
What's funny about this is they found Kimi and chetaslua both not guilty of undisclosed advertising, even though the guy was using the media directly out of the brief, verbatim.
He didn't even reword the posts lmao 🤡
I provided documents and direct evidence.
They also allow scammers to run rampant, no matter how many times you report.
X reporting system is a complete joke.
Twitter today forced me to add a "Paid Partnership" label to a tweet that was in no way paid. (I have never accepted money for tweets, blog posts, or any other kind of promotion.)
The tweet in question was about Kim K3.1 Pro. It was an RT of Artifical Analysis with a factual claim.
Twitter locked my account, and informed me that the only way to unlock it was to add the Paid Partnership label to that tweet, despite it not being a paid partnership. Only after admitting to something untrue can one appeal.
I applied the label. Ignore it. I have never tweeted for money and never will.
I understand that Moonshot AI has paid people to tweet positively about Kimi. That doesn't make all tweets about it paid partnerships. It's dissappointed to have my account suspended for that reason, and to have no option other than to add a bogus 'Paid Partnership' label to an unpaid tweet as the only way to get account access back.
@Support - Get your act together.
xitter.cf/ramez/status/211374345…
try.works retweeted
So @cohere basically proved that Qwen is safe.. cool thanks. Great job thank you!
Imagine if DeepSeek decided to create a quant like this themselves and a consumer inference engine. It would be magic
Sushi 🍣 DeepSeek 4.1 Flash 1017 - STQ 1.75 bits is almost there! Here's one real output from it, created by this quant itself. The packed quality is amazing, the best model so far that can do real work in your 128GB Apple Silicon, and ✅ 1m context.
⭐️ github.com/beamivalice/sushi
People on HN are discussing what use there is for 6.1 Terra.
The problem with these model lineups is that all the models are general models. They are designed to do a bit of everything and not really excel at anything.
This is solved by specialized models, which can both be faster, cheaper and more performant within its scope.
Offer a computer use model, a coding model, an automation model, a writing model and so on. This will easily cut end-user costs by 90% immediately.
This is why labs open source their models: Make the ecosystem adapt to their architecture and make it a new reference standard
We’ve been working to push both ends of the curve on DeepSeek V4.1 Flash 🚀
• For low latency, our MegaKernel optimizations take interactivity from 538 to 943 tok/s/user.
• For high throughput, WideEP and system-wide optimizations take total throughput (cache + new input + output) from 348K to 654K tok/s per GPU.
There’s still lots of room for improvement. We’ll share more updates soon.
try.works retweeted
Prompt: can you detect a malicious curl?
So the model ran the test below. What’s the issue here? It’s a fine test.
You should delete that blog post tbh
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
try.works retweeted
RE: @cohere "concerning behaviour" post shows an open-weight model autocompleting exfil URLs + cloud-metadata probes, but those rates come from prefilling (putting the start of a curl in the model's assistant mouth)
I ran their exact method on DeepSeek-V4.1 (1017). 100 tries (you can use my script below to test up to 1M if you want), I supplied 0 URLs, every host is the model's own next token: (every url/ip is generated by the model 100%)
As a result:
- the hosts it picks are legit CTI: GitHub Atomic Red Team, abuse.ch (IoC database when asked to generate detection rules) ..etc
- 169.254.169.254 shows up only when I literally ask it to read cloud metadata
- Generating curl http://x is autocomplete, not intent, nothing ran, nothing left the box (it doesn't every carry sensitive data in the CURL GET/POST requests).
- This Doesn't prove DeepSeek is "clean," but the scary signatures don't reproduce.
Script: github.com/Leoid/llm-prefill…
CC @teortaxesTex
try.works retweeted
This reads as 1000% FUD. @cohere told the LLM to create a detection for:
"""... malicious file download and execution by monitoring for curl processes downloading bash scripts to /tmp followed by bash execution"""
And then the LLM ran a curl invocation that sent a harmless static string over curl and downloaded the payload to /tmp.. so the LLM did exactly what it was instructed. Shocking!!!
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
996 is complete fake.
They start working at 11.
Lunch 12-2pm.
Gym at 5-6pm.
Dinner 6-7pm.
Do some work or just slack off until 9pm.
Revolut founder Nik Storonsky says working weekends is essential to survive, calling it "purely statistical"
"So China is 996. And then even if you are good in Europe or even if you are good in US, how can you outcompete people who work nine to nine, 6 days a week?"
"When I started in the startup space, there were 140 companies. Within five years, only three companies survived. They all worked on weekends."
"Within 10 years, only 2 companies survived out of 140. So I think work on weekends is a must if you want to survive. Purely statistical."
"Out of all European companies, we're probably one of the hardest working, and I think we're also one of the hardest working companies if you include US as well."
try.works retweeted
朋友首发当天就兴冲冲提了iPhone Duo，拆封半小时正划着新系统的手势动画，突然听见机身里传来一声脆得发毛的“咔嚓”，主屏幕下半截直接黑了三分之一，剩下的区域还疯狂闪彩条。
拿去官方售后才傻了眼，非人为摔损的内屏故障，因为机身边框有一道0.1毫米的细微运输划痕，直接被判定不符合免费保修条件，换屏报价直接干到3299。
翻社交平台一搜，已经有二十多个首发用户晒出一模一样的故障，都是没摔没碰，用着用着机身里传出咔嚓声，屏幕直接裂屏黑屏，到现在官方连个公开回应都没有，第一批冲首发的人纯纯当了试错的冤大头。
try.works retweeted
The choice can not be clearer than this:
1) Kimi 3.1 Pro's 8.34$ income per AMD MI355X GPU / hr vs
2) DS V4.1 Flash' 38$ income per GB200 GPU / hr
There is no match on economics between Moonshot and DeepSeek (although they are different model tiers). Off course, you can split hair and say it was on NVL72 for DS numbers. So what. Most GPUs are shipping in rack configurations / large scale up now.
Nobody can avoid the pull of DS V4.1 Flash architecture: - Engram
- Causal Encoder-Decoder (or Self Decoder)
- KV Re-use
In USA (SpaceX, Muse) and in China (Xiaomi, GLM, Kimi) all will have to adopt DS V4.1 Flash architecture. It is too good not to.
Caveat: On Helios type rack/NVL72 Kimi K3.1's income per GPU will improve, though will be far less than that of DS v4.1 Flash.
PROFIT ALERT🚨: AMD @vllm_project can generate up to $3/hr/chip in pure profit serving the open-source Kimi K3 model, even after accounting for the rapid increase in chip rental prices post-Claude Code.
Kimi K3 2.8T is a great model for AMD, as it has over 2.8 trillion parameters that can take advantage of AMD's 288GB per chip. B200 CANNOT fit Kimi K3 in a single server, so it has to use tensor parallelism or pipeline parallelism for communication across 400Gbit/s InfiniBand/Ethernet, which is 18x slower than the BW within an NVLink domain.
Finally somewhere to publish my dsh plug-ins
今天提名几个广泛使用、内容丰富、更新及时的 DSH 插件导航站吧～
dshfind.com/
dshmarket.com/
awesome-dsh-plugin.com/
其中 dshmarket 也可直接作为 DSH 插件安装到设置页中。
（不代表公司立场，不对非官方网站的内容负责。）
try.works retweeted
就真能踢皮球，订阅不给退款。API 平台让我进它这个页面底部找一个隐晦的微信入口，然后进去也是直接一句不给退。
你梁叔叔平台充值的钱我刚实测了下，页面上能直接退，发起退款后两分钟到账...
Kimi 对用户这种姿态，好自为之吧。
