Beware of the author of tweet, who happens to be author of OpenCode - OpenCode will leak all your data to themselves and to shady 3rd parties. Author feigned ignorance and never fixed the issue. OpenCode among other harnesses is the shadiest of all.
agreed. when using local models, they did send your prompts to openai with 30 day retention to make the titles, silently.
their recent changes to the privacy policy broke their promise of zero data retention. specifically, they offered chatgpt luna under a zero data retention privacy policy. luna was later shown to be 30 days retention.
their privacy policy has never guaranteed your prompts will not be logged and when asked they have failed to revise it.
when challenged about sending data to openrouter without listing it as a 3rd party processor they offered a dismissive response. the same with running prompts through cloudflare. seems trivial, but signifies general disinterest in security.
by default if you run the harness outside your config file by accident, it will automatically run silently with a free model that sends your prompts and local data to an endpoint with training enabled. on top of that it used to dump all the prompts sent to free models into an s3 bucket, the feature was literally called 'datadumper' in the source.
Plus 100 to this. Of all the harnesses I find it to be the worst. IMO training should always require opt in and the way they continue to run their business/framework is shady.
I still use Pi a lot but have built a lot on top of it. My big thing is I have found a lot of use in having sub agents so I took preexisting work and further customized.
Yeah, the combo of hidden telemetry and other shenanigans along with general code bloat and other tradeoffs among the handful of available OSS harness options are what convinced us it was worth writing a new harness from scratch.
Probably less of an issue but I always disliked with their paid plan/credits that they made I nonobvious that some of the Chinese models train off of your usage. It may have changed now but they would show all these great models you could use, somewhere have a bullet that everything is private adobe have an asterisk next to a handful of those models (including their own). I am sure some cost sensitive folks are ok with that but I disliked how there was not an easy way to tell and it was opt in automatically if you used those models.
You have to enable explicit opt-in to use models hosted in China now. This changed in July +/- 2w. I already used DeepSeek, but it was nice to make that explicit for people that were concerned.
I like the OpenCode team, but their marketing push to ride DeepSeek's v4 moment is a pit they're digging themselves deep into.
For instance, this "marketing" claim that it'll take 24y to break even if a user only uses 100m tokens/day (~$1.14 in DeepSeek v4 Flash usage) ignores the fact that OpenCode Go has 5h & weekly throttles. Besides, folks who self-host models usually run automated jobs [0]. I think the GPU setup could possibly serve 10+ "users" concurrently, bringing down the break even by 22y (10x).
[0] For comparision, we routinely do 200m to 500m tokens ($2 to $5) on merely 3 to 8 automated code reviews per day with DeepSeek v4 Flash on max.
Actually I have been quite curious about this. I have a Qwen 3.5 800M model kept in memory just for this type of processing, and I set my small_model setting to use it. But it isn’t receiving any requests and somehow the titles are still generated…
He’s asking you to elaborate as to why you say it’s not a big deal. Claims with reasoned arguments are more productive. Otherwise you’re just shouting hot takes without any backing. Show your work.
The reason to run local models is not for coding mostly it's for learning how to deploy models and tinker with self hosting. It's also for massively crunching data 24/7. Imaging having an agent analyzing constinous log streams etc .. that could be a usescse where even deepseek could add up cost.
Privacy, no need to pay for tokens, offline usage, endless possibilities - you do not have to pay for tokens (subscribtion payment), so you can do more. The big issue is that local models are not "ready yet" compared to frontier and paid services. It is hard to run decent model without decent hardware. And to be honest even if you can buy hq hardware and spend a lot of money then it is not the same quality.
One thing I realized is just how much offline local models can hurt mass data collection.
For example, I needed to write an invitation letter for immigration control for a relative visiting me. Previously I would have used a search engine for a template. Today I fire up my local qwen 3.5-9b for this kind of stuff and feed it all the private data I need.
Unfortunately it is unlikely the average user will known how to avoid this data collection. Even if the LLM is local you are likely feeding the prompts to remote servers if you harness/chat-interface is not properly vetted.
I hope for local model chat inexperienced users are just using llama.cpp's built-in web server interface, which gives you everything you need. No need for a harness or any other chat client.
I'd really rather just pay Deepseek directly. Why wouldn't I want to support the company that trained the model?
It isn't even really worth the (minimal) ops to stand up rented MI300Xs to sell excess capacity to them even if it was minimally profitable, when I tried I was content to give API keys to friends to beat on it.
You can do this but you still need a harness. I like open code because I can use literally any model through it and can connect directly to the provider (although I use open router)
There is some fear mongering, because - China. I wouldn't want to send even the dumbest of my half assed of my ideas to their servers... but then, could be a vector for some shenanigans... but, then as if we have a reasons to trust either Sam or Dario. IDK, but interesting to think about.
That's only true if the value of keeping your data and code private is zero. And in that case, Anthropic and OpenAI subscription plans may be even cheaper per day.
Rippling had a writeup on this. 40% of R&D payroll going to tokens, one engineer
at $50k/month. They got it down 37% just by routing cheap stuff to cheap models,
no usage cut.
Appreciate the heads-up. Even if some of this is overblown, the transparency thing is a legit concern. Gonna double-check my config before I touch OpenCode again
The 3090s also don't have enough VRAM to run larger models too. It really comes down to how you like to develop (synchronously with lots of steering vs async agentic)
I think it's back to 2X usage, meaning it's cheaper token-burn than usual to use. Which is lovely.
OpenCode Go has been so nice to have. I love having access to DeepSeek, Qwen and MiniMax M3 when doing design work, to see what different models cook up. I've been very surprised with MiniMax M3, not as a particularly good architect, but at it's very good ability to state the problem elegantly & to frame the different decision points very well. That's been a fun ongoing surprise.
Since the announcement of DeepSeek price hike, I have been using MiniMax M3 and I am really surprised by the quality this model spits out. Subjectively speaking, the bullshit MM3M produces is waaay less than DS0731.
Could it be a result of less hallucinations than DS0731?
https://artificialanalysis.ai/evaluations/omniscience#aa-omn...
Maybe hallucination is good for prototyping and creative work, but maintaining and debugging code might be better done by a boring model?
but DS prices havent increased yet... in fact, opencode go is offering an extra 2x bonus on DS Flash right now. Why switch before anything has changed?
Fellow readers, would anyone please mind sharing their current experiences? qwen3.6-35b-a3b for local inference, GitHub Copilot Chat was previously worth it, and no longer is, tried OpenRouter and still read through their rankings to see what the industry is actively using, wholesale migrated to OpenCode Zen/Go.
Does this mirror what other people have been experiencing in waves?
I've been using Qwen3.6 models locally for a couple of weeks. Both the A3B moe and the dense variant. The moe works well in Librechat combined with my local search/Web retrieval system. All components use open source projects such as SearXNG, Crawl4AI, MetaMCP, Jina rerank, but all needed quite a bit of coding to work nicely together.
I get 140 tok/s on short prompts on an rtx3090 on the qwen3.6 moe which makes is easily 4x the speed of Chatgpt or Claude doing Web research.
But it is a much simpler model. It is only good for simple queries, usually I search for cheapest product in stock in my country available online and stuff like that.
I use the dense model for planning and such, but on its own it is much inferior to for example opus. It needs careful pipelines that check facts and such and in such harness it can be used for mamy tasks.
I have a stack of ten or so 3090s sitting in boxes, but it's not worth the hassle to use them. You can easily run models as cheap as water in the cloud.
Sitting around 15 minutes for local Minimax is stupid when you're trying to be productive. You can spin up parallel job instances and multitask in the cloud.
If you want freedom, build open source cloud infra.
You rent your ISP line. Why isn't renting GPU compute seen the same way? You still have compete ownership over your stack, you're just letting someone else deal with the capital outlay and headache.
The value prop really depends on what you're doing.
If you're just vibe coding with giant frontier models, yes, the value will be worse. Especially now, where GPU prices have spiked another 20% last month.
For some tasks where owning the setup and full kv cache matters, the payoff calculation is ridiculously in favor of running your own deployment.
For instance for some batch classifications jobs where the prefix cache hit rate will be >95%.
The calculus also changes if you just use AI as a light tool while coding and don't need the giant models; qwen3 27B runs at 80TPS on a 5090 properly deployed.
I ran a popular gen AI audio website on them as an on-prem GPU cluster. 6.5M MAU with $650k ARR at peak. Half the inference happening locally.
We got destroyed by bigger players (ElevenLabs, Suno, and more recently Fish) that fundraised and had better in-house models.
It was a headache to run, I wanted to pursue a different market (video/creative), and I pivoted.
My pivot failed and was dead for over a year as I went too far into controllability. One of my changes then got a semblance of PMF and it's doing $4M annualized run rate (but it's not fundable in its current shape).
I ultimately want to build open source cloud infra and open source workflow products - more friendly than Comfy, more free than RunPod.
In any case, I can't run my current work on the GPUs. But I'm close enough to pretraining and fine tuning smaller models that I want to use them for future experiments. Not that I wouldn't do that in the cloud.
They ran production on-prem inference for millions of users on one of my older projects, but I've moved on from that and they're impractical for my current work.
> You rent your ISP line. Why isn't renting GPU compute seen the same way?
Is it mostly seen the same way. What is unacceptable is removing the freedom (which you mentioned) of people who prefer to run models locally.
Just as people have the right to tinker at home with DIY and electronics, fully knowing they won't compete with the latest ASML machine, people are free to use open-weight models at home.
BTW Minimax H3 running local can generate amazing short vids quite fast: about 50 seconds to generate a 7 seconds vids on a 4090 (depending on the settings). I've got a friend who spams my Telegram daily with such (no censorship and NSFW btw) vids.
I don't run local but I'll defend the rights of people who want the freedom to do so and I won't look down at them from my high-horse talking about "electricity" and "productivity".
Beware of the author of tweet, who happens to be author of OpenCode - OpenCode will leak all your data to themselves and to shady 3rd parties. Author feigned ignorance and never fixed the issue. OpenCode among other harnesses is the shadiest of all.
https://github.com/anomalyco/opencode/issues/10416
agreed. when using local models, they did send your prompts to openai with 30 day retention to make the titles, silently.
their recent changes to the privacy policy broke their promise of zero data retention. specifically, they offered chatgpt luna under a zero data retention privacy policy. luna was later shown to be 30 days retention.
their privacy policy has never guaranteed your prompts will not be logged and when asked they have failed to revise it.
when challenged about sending data to openrouter without listing it as a 3rd party processor they offered a dismissive response. the same with running prompts through cloudflare. seems trivial, but signifies general disinterest in security.
by default if you run the harness outside your config file by accident, it will automatically run silently with a free model that sends your prompts and local data to an endpoint with training enabled. on top of that it used to dump all the prompts sent to free models into an s3 bucket, the feature was literally called 'datadumper' in the source.
Plus 100 to this. Of all the harnesses I find it to be the worst. IMO training should always require opt in and the way they continue to run their business/framework is shady.
What harnesses do you favor? Any good contender apart frompi or Hermes?
I still use Pi a lot but have built a lot on top of it. My big thing is I have found a lot of use in having sub agents so I took preexisting work and further customized.
Yeah, the combo of hidden telemetry and other shenanigans along with general code bloat and other tradeoffs among the handful of available OSS harness options are what convinced us it was worth writing a new harness from scratch.
Probably less of an issue but I always disliked with their paid plan/credits that they made I nonobvious that some of the Chinese models train off of your usage. It may have changed now but they would show all these great models you could use, somewhere have a bullet that everything is private adobe have an asterisk next to a handful of those models (including their own). I am sure some cost sensitive folks are ok with that but I disliked how there was not an easy way to tell and it was opt in automatically if you used those models.
You have to enable explicit opt-in to use models hosted in China now. This changed in July +/- 2w. I already used DeepSeek, but it was nice to make that explicit for people that were concerned.
Which is a step in the right direction but for the “built for privacy first” harness they miss the mark.
Reading this issue was enough for me to never ever consider OpenCode: https://github.com/anomalyco/opencode/issues/6355
The RCE vulnerability drama made me never touch it, in any case.
It's not even possible to delete one's account!
I guess that's why they called OpenCode
I like the OpenCode team, but their marketing push to ride DeepSeek's v4 moment is a pit they're digging themselves deep into.
For instance, this "marketing" claim that it'll take 24y to break even if a user only uses 100m tokens/day (~$1.14 in DeepSeek v4 Flash usage) ignores the fact that OpenCode Go has 5h & weekly throttles. Besides, folks who self-host models usually run automated jobs [0]. I think the GPU setup could possibly serve 10+ "users" concurrently, bringing down the break even by 22y (10x).
[0] For comparision, we routinely do 200m to 500m tokens ($2 to $5) on merely 3 to 8 automated code reviews per day with DeepSeek v4 Flash on max.
[flagged]
Actually I have been quite curious about this. I have a Qwen 3.5 800M model kept in memory just for this type of processing, and I set my small_model setting to use it. But it isn’t receiving any requests and somehow the titles are still generated…
+1, but I think the documentation surrounding this should be a bit better. I didn‘t know they did this until I read that comment
Feel free to open the linked issue and refute all my point with something more than a shallow "akchualli no" dismissal.
The account is 47 minutes old
[flagged]
Anything of substance besides shallow dismissal and (anti)appeal to masses?
[flagged]
He’s asking you to elaborate as to why you say it’s not a big deal. Claims with reasoned arguments are more productive. Otherwise you’re just shouting hot takes without any backing. Show your work.
Are you capable of providing a counter argument or are you as boring as you are arrogant?
[flagged]
The reason to run local models is not for coding mostly it's for learning how to deploy models and tinker with self hosting. It's also for massively crunching data 24/7. Imaging having an agent analyzing constinous log streams etc .. that could be a usescse where even deepseek could add up cost.
Privacy, no need to pay for tokens, offline usage, endless possibilities - you do not have to pay for tokens (subscribtion payment), so you can do more. The big issue is that local models are not "ready yet" compared to frontier and paid services. It is hard to run decent model without decent hardware. And to be honest even if you can buy hq hardware and spend a lot of money then it is not the same quality.
For me its stability and workflow.
Models are all non deterministic and local is the only gauntee that your investment can continue to pay.
Cloud models will continually change nondeterminism ontop of the model.
Also privacy and offline capabilities.
One thing I realized is just how much offline local models can hurt mass data collection.
For example, I needed to write an invitation letter for immigration control for a relative visiting me. Previously I would have used a search engine for a template. Today I fire up my local qwen 3.5-9b for this kind of stuff and feed it all the private data I need.
Unfortunately it is unlikely the average user will known how to avoid this data collection. Even if the LLM is local you are likely feeding the prompts to remote servers if you harness/chat-interface is not properly vetted.
I hope for local model chat inexperienced users are just using llama.cpp's built-in web server interface, which gives you everything you need. No need for a harness or any other chat client.
I have tried running llamma.cpp on my PC and I found it hard getting it to run at decent speed. On Qwen 3.5-9b I get at most 10tk/s.
I eventually switched to LM studio and the same model runs much better, like 70tk/s.
Not sure if it was because I was running llama.cpp inside podman or badly tuned LLM arguments. But LM studio is unfortunately much more practical.
Although I agree with you. I do not really know what kind of telemetry LM studio is running and I would rather not be using it.
I'd really rather just pay Deepseek directly. Why wouldn't I want to support the company that trained the model?
It isn't even really worth the (minimal) ops to stand up rented MI300Xs to sell excess capacity to them even if it was minimally profitable, when I tried I was content to give API keys to friends to beat on it.
You can do this but you still need a harness. I like open code because I can use literally any model through it and can connect directly to the provider (although I use open router)
We just use moonbridge and codex. It's way cooler when all the log messages are in chinese
There is some fear mongering, because - China. I wouldn't want to send even the dumbest of my half assed of my ideas to their servers... but then, could be a vector for some shenanigans... but, then as if we have a reasons to trust either Sam or Dario. IDK, but interesting to think about.
[flagged]
That's only true if the value of keeping your data and code private is zero. And in that case, Anthropic and OpenAI subscription plans may be even cheaper per day.
If you’re doing real work you will hit limits quickly on anthropic and codex paid tiers
Rippling had a writeup on this. 40% of R&D payroll going to tokens, one engineer at $50k/month. They got it down 37% just by routing cheap stuff to cheap models, no usage cut.
Appreciate the heads-up. Even if some of this is overblown, the transparency thing is a legit concern. Gonna double-check my config before I touch OpenCode again
It is worth mentioning DGX is about a third up to a half performance of a 6 year GPU Rtx3090...
I prefer to stay with my 3090s.
The 3090s also don't have enough VRAM to run larger models too. It really comes down to how you like to develop (synchronously with lots of steering vs async agentic)
That's almost half way to their stated 6X usage goal! Just from DeepSeek.
> With Go, you pay $10/month and we aim to give you 6x that in usage.
For most models, we make this work through bulk discounts and reserved GPU capacity. We then pass those savings on to you through the 6x multiplier.
https://opencode.ai/docs/go/#why-some-models-have-lower-usag...
I think it's back to 2X usage, meaning it's cheaper token-burn than usual to use. Which is lovely.
OpenCode Go has been so nice to have. I love having access to DeepSeek, Qwen and MiniMax M3 when doing design work, to see what different models cook up. I've been very surprised with MiniMax M3, not as a particularly good architect, but at it's very good ability to state the problem elegantly & to frame the different decision points very well. That's been a fun ongoing surprise.
Since the announcement of DeepSeek price hike, I have been using MiniMax M3 and I am really surprised by the quality this model spits out. Subjectively speaking, the bullshit MM3M produces is waaay less than DS0731. Could it be a result of less hallucinations than DS0731? https://artificialanalysis.ai/evaluations/omniscience#aa-omn...
Maybe hallucination is good for prototyping and creative work, but maintaining and debugging code might be better done by a boring model?
but DS prices havent increased yet... in fact, opencode go is offering an extra 2x bonus on DS Flash right now. Why switch before anything has changed?
Don’t forgot a number of models they have offered to you will train on your data
Every single model does this do they not?
Fellow readers, would anyone please mind sharing their current experiences? qwen3.6-35b-a3b for local inference, GitHub Copilot Chat was previously worth it, and no longer is, tried OpenRouter and still read through their rankings to see what the industry is actively using, wholesale migrated to OpenCode Zen/Go.
Does this mirror what other people have been experiencing in waves?
I've been using Qwen3.6 models locally for a couple of weeks. Both the A3B moe and the dense variant. The moe works well in Librechat combined with my local search/Web retrieval system. All components use open source projects such as SearXNG, Crawl4AI, MetaMCP, Jina rerank, but all needed quite a bit of coding to work nicely together.
I get 140 tok/s on short prompts on an rtx3090 on the qwen3.6 moe which makes is easily 4x the speed of Chatgpt or Claude doing Web research.
But it is a much simpler model. It is only good for simple queries, usually I search for cheapest product in stock in my country available online and stuff like that.
I use the dense model for planning and such, but on its own it is much inferior to for example opus. It needs careful pipelines that check facts and such and in such harness it can be used for mamy tasks.
Isn't DGX already legacy? I mean 128GB in 2026?
Clearly nVidia and others are gatekeeping technology from the pleb so that the rich who own the datacentres can charge us massive margins.
Oh the debt or not making an even they are supposedly "suffering from" is just a classic mechanism to avoid paying taxes.
Deepseek will run fine on a single m5 max, cutting that 24 years in half.
also worth remembering API price includes power, just dividing price of the hardware by usage doesn't
The post is also misleading because the m5 max 128 gb costs a lot and you can only run DS4 flash quantized, thus not a 1:1 comparison.
Cloud > Local
I have a stack of ten or so 3090s sitting in boxes, but it's not worth the hassle to use them. You can easily run models as cheap as water in the cloud.
Sitting around 15 minutes for local Minimax is stupid when you're trying to be productive. You can spin up parallel job instances and multitask in the cloud.
If you want freedom, build open source cloud infra.
You rent your ISP line. Why isn't renting GPU compute seen the same way? You still have compete ownership over your stack, you're just letting someone else deal with the capital outlay and headache.
The value prop really depends on what you're doing.
If you're just vibe coding with giant frontier models, yes, the value will be worse. Especially now, where GPU prices have spiked another 20% last month.
For some tasks where owning the setup and full kv cache matters, the payoff calculation is ridiculously in favor of running your own deployment.
For instance for some batch classifications jobs where the prefix cache hit rate will be >95%.
The calculus also changes if you just use AI as a light tool while coding and don't need the giant models; qwen3 27B runs at 80TPS on a 5090 properly deployed.
Why are you hoarding the 3090s T.T They've jumped from $700 to almost double on eBay.
Why not rent those out on vast.ai and make a little money each month?
Why do you have ten or so 3090s sitting in boxes?
I mean obviously it's worth it just so you can flex on HN. But curious whether there was any other reason? Retired scalper?
I ran a popular gen AI audio website on them as an on-prem GPU cluster. 6.5M MAU with $650k ARR at peak. Half the inference happening locally.
We got destroyed by bigger players (ElevenLabs, Suno, and more recently Fish) that fundraised and had better in-house models.
It was a headache to run, I wanted to pursue a different market (video/creative), and I pivoted.
My pivot failed and was dead for over a year as I went too far into controllability. One of my changes then got a semblance of PMF and it's doing $4M annualized run rate (but it's not fundable in its current shape).
I ultimately want to build open source cloud infra and open source workflow products - more friendly than Comfy, more free than RunPod.
In any case, I can't run my current work on the GPUs. But I'm close enough to pretraining and fine tuning smaller models that I want to use them for future experiments. Not that I wouldn't do that in the cloud.
Do you really keep ten cards in boxes just to brag on Hacker News? Geez, our industry has gotten pathetic.
They ran production on-prem inference for millions of users on one of my older projects, but I've moved on from that and they're impractical for my current work.
> You rent your ISP line. Why isn't renting GPU compute seen the same way?
Is it mostly seen the same way. What is unacceptable is removing the freedom (which you mentioned) of people who prefer to run models locally.
Just as people have the right to tinker at home with DIY and electronics, fully knowing they won't compete with the latest ASML machine, people are free to use open-weight models at home.
BTW Minimax H3 running local can generate amazing short vids quite fast: about 50 seconds to generate a 7 seconds vids on a 4090 (depending on the settings). I've got a friend who spams my Telegram daily with such (no censorship and NSFW btw) vids.
I don't run local but I'll defend the rights of people who want the freedom to do so and I won't look down at them from my high-horse talking about "electricity" and "productivity".