BearSSL is really cool, but it claims beta quality with the latest release in 2018, doesn't support TLS 1.3, and hasn't seen meaningful development in years. It's averaging about 1 commit per year recently, and they're not big ones.
Really looked forward to this release as MiniMax M2.1 is currently my most used model thanks to it being fast, cheap and excellent at tool calling. Whilst I still use Antigravity + Claude for development, I reach for MiniMax first in my AI workflows, GLM for code tasks and Kimi K2.5 when deep English analysis is needed.
Not self-hosting yet, but I prefer using Chinese OSS models for AI workflows because of the potential to self-host in future if needed. Also using it to power my openclaw assistant since IMO it has the best balance of speed, quality and cost:
> It costs just $1 to run the model continuously for an hour at 100 tokens/sec. At 50 tokens/sec, the cost drops to $0.30.
Using a coding plan, haven't noticed any throttling and very happy with the performance. They publish the quotas for each of their plans on their website [1]:
Lets not miss that MiniMax M2.5 [1] is also available today in their Chat UI [2].
I've got subs for both and whilst GLM is better at coding, I end up using MiniMax a lot more as my general purpose fast workhorse thanks to its speed and excellent tool calling support.
My perspective aligns with this: I used to obsess over the Best Model, which I defined as "top of benchmarks", which also meant Biggest, Slowest and Most Expensive.
Then I gave two models a Real World Task.
The "Best" model took 3x longer to complete it, and cost 10x more. [0]
Now I define Best Model as "the smallest, fastest, cheapest one that can get the job done". (Currently happy with GLM-4.7 on Cerebras, at least I would be if the unlimited plan wasn't sold out ;)
I later expanded this principle when model speed crossed into the Interactive domain. Speed is not merely a feature; a sufficient difference in speed actually produces a completely new category of usage.
[0] We recently arrived at an approximation of AGI which is "put a lossy solver in an until-done loop". For most tasks we're throwing stuff at a wall to see what sticks, and the smaller models throw faster.
We're getting high quality drops for the perfect trifecta of leading models from the Chinese models with the same day release of GLM-5 and Kimi K2.5 1T model drop a few days ago.
Despite having many great options, I end up making use of all 3. MiniMax is my fast workhorse for tool calling and getting quick responses. GLM for all coding tasks whilst Kimi K2.5 1T model has deep knowledge for everything else, with an Opus-level command of the english language. There's been many times where I've preferred Kimi K2.5 over Opus.
Yep the Qwen team has been churning out models for basically everything,
and lets not sleep the big blue whale that started it all which is rumored to have a 1M context drop coming soon [1]
It's looking like we'll have Chinese OSS to thank for being able to host our own intelligence, free from the whims of proprietary megacorps.
I know it doesn't make financial sense to self-host given how cheap OSS inference APIs are now, but it's comforting not being beholden to anyone or requiring a persistent internet connection for on-premise intelligence.
Didn't expect to go back to macOS but they're basically the only feasible consumer option for running large models locally.
I guess that's debatable. I regularly run out of quota on my claude max subscription. When that happens, I can sort of kind of get by with my modest setup (2x RTX3090) and quantized Qwen3.
And this does not even account for privacy and availability. I'm in Canada, and as the US is slowly consumed by its spiral of self-destruction, I fully expect at some point a digital iron curtain will go up. I think it's prudent to have alternatives, especially with these paradigm-shattering tools.
> That's like ten normal computers worth of power for the GPUs alone.
Maybe if your "computer" in question is a smartphone? Remember that the M3 Ultra is a 300w+ chip that won't beat one of those 3090s in compute or raster efficiency.
I wouldn't class the M3 Ultra as a "normal" computer either. That's a big-ass workstation. I was thinking along the lines of a typical Macbook or Mac Mini or Windows laptop, which are fine for 99% of anyone who isn't looking to play games or run gigantic AI models locally.
Did you even try to read and understand the parent comment? They said they regularly run out of quota on the exact subscription you're advising they subscribe to.
Self-hosting training (or gaming) makes a lot of sense, and once you have the hardware self-hosting inference on it is an easy step.
But if you have to factor in hardware costs self-hosting doesn't seem attractive. All the models I can self-host I can browse on openrouter and instantly get a provider who can get great prices. With most of the cost being in the GPUs themselves it just makes more sense to have others do it with better batching and GPU utilization
If you can get near 100% utilization for your own GPUs (i.e. you're letting requests run overnight and not insisting on any kind of realtime response) it starts to make sense. OpenRouter doesn't have any kind of batched requests API that would let you leverage that possibility.
For inference, even with continuous batching, getting 100% MFUs is basically impossible to do in practice. Even the frontier labs struggle with this in highly efficient infiniband clusters. Its slightly better with training workloads just due to all the batching and parallel compute, but still mostly unattainable with consumer rigs (you spend a lot of time waiting for I/O).
I also don't think the 100% util is necessary either, to be fair. I get a lot of value out of my two rigs (2x rtx pro 6000, and 4x 3090) even though it may not be 24/7 100% MFU. I'm always training, generating datasets, running agents, etc. I would never consider this a positive ROI measured against capex though, that's not really the point.
No I'm saying there are quite a few more bottlenecks than that (I/O being a big one). Even in the more efficient training frameworks, there's per-op dispatch overhead in python itself. All the boxing/unboxing of python objects to C++ handles, dispatcher lookup + setup, all the autograd bookkeeping, etc.
All of the bottlenecks in sum is why you'd never get to 100% MFUs (but I was conceding you probably don't need to in order to get value)
That’s kind of a moot point. Even if none of those overheads existed you would still be getting a a fractions of the mfu. Models are fundamental limited by memory bandwidth even with best case scenarios of sft or prefill.
I don't believe it's moot, but I understand your point. The fact that models are memory bandwidth bound does not at all mean that other overhead is insignificant. Your practical delivered throughput is the minimum of compute ceiling, bandwidth ceiling, and all the unrelated speed limits you hit in the stack. Kernel launch latency, Python dispatch, framework bookkeeping, allocator churn, graph breaks, and sync points can all reduce effective speed. There are so many points in the training and inference loop where the model isn't even executing.
> And what are you doing that I/O is a bottleneck?
We do a fair amount of RLVR at my org. That's almost entirely waiting for servers/envs to do things, not the model doing prefill or decode (or even up/down weighting trajectories). The model is the cheap part in wall clock terms. The hard limits are in the verifier and environment pipeline. Spinning up sandboxes, running tests, reading and writing artifacts, and shuttling results through queues, these all create long idle gaps where the GPU is just waiting to do something.
> That's almost entirely waiting for servers/envs to do things
I'm not sure why, sandboxes/envs should be small and easy to scale horizontally to the point where your throughput is no longer limited by them, and the maximum latency involved should also be quite tiny (if adequately optimized). What am I missing?
First as an aside, remember that this entire thread is about using local compute. What you're alluding to is some fantasy infinite budget where you have limitless commodity compute. That's not at all the context of this thread.
But disregarding that, this isn't a problem you can solve by turning a knob akin to scaling a stateless k8s cluster.
The whole vertical of distributed RL has been struggling with this for a while. You can in theory just keep adding sandboxes in parallel, but in RLVR you are constrained by 1) the amount of rollout work you can do per gradient update, and 2) the verification and pruning pipeline that gates the reward signal.
You cant just arbitrarily have a large batch size for every rollout phase. Large batches often reduce effective diversity or get dominated by stragglers. And the outer loop is inherently sequential, because each gradient update depends on data generated by a particular policy snapshot. You can parallelize rollouts and the training step internally, but you can’t fully remove the policy-version dependency without drifting off-policy and taking on extra stability headaches.
In Silicon Valley we pay PG&E close to 50 cents per kWh. An RTX 6000 PC uses about 1 kW at full load, and renting such a machine from vast.ai costs 60 cents/hour as of this morning. It's very hard for heavy-load local AI to make sense here.
Yikes.. I pay ~7¢ per kWh in Quebec. In the winter the inference rig doubles as a space heater for the office, I don't feel bad about running local energy-wise.
And you are forgetting the fact that things like vast.ai subscriptions would STILL be more expensive than Openrouter's api pricing and even more so in the case of AI subscriptions which actively LOSE money for the company.
So I would still point out the GP (Original comment) where yes, it might not make financial sense to run these AI Models [They make sense when you want privacy etc, which are all fair concerns but just not financial sense]
But the fact that these models are open source still means that they can be run when maybe in future the dynamics might shift and it might make sense running such large models locally. Even just giving this possibility and also the fact that multiple providers could now compete in say openrouter etc. as well. All facts included, definitely makes me appreciate GLM & Kimi compared to proprietory counterparts.
Oops sorry. Fixed it now but I am trying a HN progressive extension and what it does is if I have any text selected it can actually quote it and I think this is what might've happened or such a bug I am not sure.
Anthropic has very tight limits, so you're basically using the worst (pricing-wise) SOTA cloud model as your baseline. I have $200 subs for both Claude and OpenAI, and I also bump into limits with Claude all the time, whether coding or research. With Codex, I ran into the limit once so far, and that's in a month of very heavy (sometimes literally 24 hours around the clock, leaving long-running tasks overnight) use.
I bought the Gemini Ultra to try for a month (at the discounted price). I have been using it non-stop for Opus 4.6 Thinking, which is much better than Gemini 3 Pro (High) and it's been a blast. The most I've managed to consume is 60% of my 5 hourly quota. That was with 2-3 instances in parallel.
I hope too many of us won't be doing this and cause Google to add limits! My hope is Google sees the benefit in this and goes all in - continues to let people decide which Google hosted model to use, including their own.
Getting CC to work with other models is quite straightforward -- setting a few env vars, and a thin proxy that rewrites the requests/responses to be in the expected format.
Not OP, but I am pretty sure they are using Opencode with a certain antigravity plugin. Not going to link it, since it technically allows breaking TOS. If you‘re not using Opencode yet, I wholeheartedly recommend the switch.
Did the napkin math on M3 Ultra ROI when DeepSeek V3 launched: at $0.70/2M tokens and 30 tps, a $10K M3 Ultra would take ~30 years of non-stop inference to break even - without even factoring in electricity. Clearly people aren't self-hosting to save money.
I've got a lite GLM sub $72/yr which would require 138 years to burn through the $10K M3 Ultra sticker price. Even GLM's highest cost Max tier (20x lite) at $720/yr would buy you ~14 years.
Everyone should do the calculation for themselves. I too pay for couple of subs. But I'm noticing having an agent work for me 24/7 changes the calculation somewhat. Often not taken into account: the price of input tokens. To produce 1K of code for me, the agent may need to churn through 1M of tokens of codebase. IDK if that will be cached by the API provider or not, but that makes x5-7 times price difference. OK discussion today about that and more https://x.com/alexocheema/status/2020626466522685499
And it's worth noting that you can get DeepSeek at those prices from DeepSeek (Chinese), DeepInfra (US with Bulgarian founder), NovitaAI (US), AtlasCloud (US with Chinese founder), ParaSail (US), etc. There is no shortage of companies offering inference, with varying levels of trustworthiness, certificates and promises around (lack of) data retention. You just have to pick one you trust
Doing inference with a Mac Mini to save money is more or less holding it wrong. Of course if you buy some overpriced Apple hardware it’s going to take years to break even.
Buy a couple real GPUs and do tensor parallelism and concurrent batch requests with vllm and it becomes extremely cost competitive to run your own hardware.
> Doing inference with a Mac Mini to save money is more or less holding it wrong.
No one's running these large models on a Mac Mini.
> Of course if you buy some overpriced Apple hardware it’s going to take years to break even.
Great, where can I find cheaper hardware that can run GLM 5's 745B or Kimi K2.5 1T models? Currently it requires 2x M3 Ultras (1TB VRAM) to run Kimi K2.5 at 24 tok/s [1] What are the better value alternatives?
Six months ago I'd have said EPYC Turin. You could do a heck of a build with 12Ch DDR5-6400 and a GPU or two for the dense model parts. 20k would have been a huge budget for a homelab CPU/GPU inference rig at the time. Now 20k won't buy you the memory.
It's important to have enough VRAM to get the kv cache and shared trunk of the model on GPU, but beyond that it's really hard to make a dent in the pool of 100s of gigabytes of experts.
I wish I had better numbers to compare with the 2x M3 Ultra setup. My system is a few RTX A4000s on a Xeon with 190GB/s actual read bandwidth, and I get ~8 tok/s with experts quantized to INT4 (for large models with around 30B active parameters like Kimi K2.) Moving to 1x RTX Pro 6000 Blackwell and tripling my read bandwidth with EPYC Turin might make it competitive with the the macs, but I dunno!
There's also some interesting tech with ktransformers + sglang where the most frequently-used experts are loaded on GPU. Pretty neat stuff and it's all moving fast.
my system is running GLM-5 MXFP4 at about 17 tok/s. That’s with a single RTX Pro 6000 on an EPYC 9455P with 12 channels of DDR5-6400. Only 16k context though, since it’s too slow to use for programming anyway and that’s the only application where I need big context.
> I regularly run out of quota on my claude max subscription. When that happens, I can sort of kind of get by with my modest setup (2x RTX3090) and quantized Qwen3.
When talking about fallback from Claude plans, The correct financial comparison would be the same model hosted on OpenRouter.
You could buy a lot of tokens for the price of a pair of 3090s and a machine to run them.
Your $5,000 PC with 2 GPUs could have bought you 2 years of Claude Max, a model much more powerful and with longer context. In 2 years you could make that investment back in pay raise.
> In 2 years you could make that investment back in pay raise.
you can't be a happy uber driver making more money in the next 24 months by having a fancy car fitted with the best FSD in town when all cars in your town have the same FSD.
Nothing changed since ’87. Machines still can’t be accountable and still shouldn’t make managerial decisions. Acceptance control is one of those decisions, and all the technical knowledge still matters to form a well-informed one. It may change, of course, but I have an impression that those who try otherwise seem to not fare well after the initial vibecoding honeymoon period. Of course, it varies from case to case - sometimes machines get things right, but long-term luck seems to eventually run out.
Unless you already had those cards, it probably still doesn’t make sense from a purely financial perspective unless you have other things you’re discounting for.
Not the person you’re responding to, but my experience with models up through Qwen3-coder-next is that they’re not even close.
They can do a lot of simple tasks in common frameworks well. Doing anything beyond basic work will just burn tokens for hours while you review and reject code.
It's just as fast, but not nearly as clever. I can push the context size to 120k locally, but quality of the work it delivers starts to falter above say 40k. Generally you have to feed it more bite-sized pieces, and keep one chat to one topic. It's definitely a step down from SOTA.
In one sense yes, but the training data is not open, nor is the data selection criteria (inclusions/exclusions, censorship, safety, etc). So we are still subject to the whims of someone much more powerful that ourselves.
The good thing is that open weights models can be finetuned to correct any biases that we may find.
Apple devices have high memory bandwidth necessary to run LLMs at reasonable rates.
It’s possible to build a Linux box that does the same but you’ll be spending a lot more to get there. With Apple, a $500 Mac Mini has memory bandwidth that you just can’t get anywhere else for the price.
But a $500 Mac Mini has nowhere near the memory capacity to run such a model. You'd need at least 2 512GB machines chained together to run this model. Maybe 1 if you quantized the crap out of it.
And Apple completely overcharges for memory, so.
This is a model you use via a cheap API provider like DeepInfra, or get on their coding plan. It's nice that it will be available as open weights, but not practical for mere mortals to run.
But I can see a large corporation that wants to avoid sending code offsite setting up their own private infra to host it.
The needed memory capacity depends on active parameters (not the same as total with a MoE model) and context length for the purpose of KV caching. Even then the KV cache can be pushed to system RAM and even farther out to swap, since writes to it are small (just one KV vector per token).
With Apple devices you get very fast predictions once it gets going but it is inferior to nvidia precisely during prefetch (processing prompt/context) before it really gets going.
For our code assistant use cases the local inference on Macs will tend to favor workflows where there is a lot of generation and little reading and this is the opposite of how many of use use Claude Code.
Source: I started getting Mac Studios with max ram as soon as the first llama model was released.
> With Apple devices you get very fast predictions once it gets going but it is inferior to nvidia precisely during prefetch (processing prompt/context) before it really gets going
I have a Mac and an nVidia build and I’m not disagreeing
But nobody is building a useful nVidia LLM box for the price of a $500 Mac Mini
You’re also not getting as much RAM as a Mac Studio unless you’re stacking multiple $8,000 nVidia RTX 6000s.
There is always something faster in LLM hardware. Apple is popular for the price points of average consumers.
It depends. This particular model has larger experts with more active parameters so 16GB is likely not enough (at least not without further tricks) but there are much sparser models where an active expert can be in RAM while the weights for all other experts stay on disk. This becomes more and more of a necessity as models get sparser and RAM itself gets tighter. It lowers performance but the end result can still be "useful".
This. It's awful to wait 15 minutes for M3 Ultra to start generating tokens when your coding agent has 100k+ tokens in its context. This can be partially offset by adding DGX Spark to accelerate this phase. M5 Ultra should be like DGX Spark for prefill and M3 Ultra for token generation but who know when it will pop up and for how much? And it still will be at around 3080 GPU levels just with 512GB RAM.
All Apple devices have a NPU which is potentially able to save power for compute bound operations like prefill (at least if you're ok with FP16 FMA/INT8 MADD arithmetic). It's just a matter of hooking up support to the main local AI frameworks. This is not a speedup per se but gives you more headroom wrt. power and thermals for everything else, so should yield higher performance overall.
AFAIK, only CoreML can use Apple's NPU (ANE). Pytorch, MLX and the other kids on the block use MPS (the GPU). I think the limitations you mentioned relate to that (but I might be missing something)
And then only Apple devices have 512GB of unified memory, which matters when you have to combine larger models (even MoE) with the bigger context/KV caching you need for agentic workflows. You can make do with less, but only by slowing things down a whole lot.
> a $500 Mac Mini has memory bandwidth that you just can’t get anywhere else for the price.
The cheapest new mac mini is $600 on Apple's US store.
And it has a 128-bit memory interface using LPDDR5X/7500, nothing exotic. The laptop I bought last year for <$500 has roughly the same memory speed and new machines are even faster.
> The cheapest new mac mini is $600 on Apple's US store.
And you're only getting 16GB at that base spec. It's $1000 for 32GB, or $2000 for 64GB plus the requisite SOC upgrade.
> And it has a 128-bit memory interface using LPDDR5X/7500, nothing exotic.
Yeah, 128-bit is table stakes and AMD is making 256-bit SOCs as well now. Apple's higher end Max/Ultra chips are the ones which stand out with their 512 and 1024-bit interfaces. Those have no direct competition.
Also, cheaper... X99 + 8x DDR4 + 2696V4 + 4x Tesla P4s running on llama.cpp.
Total cost about $500 including case and a 650W PSU, excluding RAM.
Running TDP about 200W non peak 550W peak (everything slammed, but I've never seen it and I've an AC monitor on the socket).
GLM 4.5 Air (60GB Q3-XL) when properly tuned runs at 8.5 to 10 tokens / second, with context size of 8K.
Throw in a P100 too and you'll see 11-12.5 t/s (still tuning this one).
Performance doesn't drop as much for larger model sizes as the internode communication and DDR4 2400 is the limiter, not the GPUs.
I've been using this with 4 channel 96GB ram, recently updated to 128GB.
Not feasible for Large models, it takes 2x M3 512GB Ultra's to run the full Kimi K2.5 model at a respectable 24 tok/s. Hopefully the M5 Ultra will can improve on that.
these run some pretty decent models locally, currently I'd recommend GPT-OSS 120GB, Qwen Coder Next 80B (either Q8 or Q6 quants, depending on speed/quality trade-offs) and the very best model you can run right now which is Step 3.5 Flash (ubergarm GGUF quant) with 256K context although this does push it to the limit - GLMs and nemotrons also worth trying depending on your priorities
there's clearly a big quantum leap in the SotA models using more than 512GB VRAM, but i expect that in a year or two, the current SotA is achievable with consumer level hardware, if nothing else hardware should catch up with running Kimi 2.5 for cheaper than 2x 512GB mac studio ultras - perhaps medusa halo next year supports 512GB and DDR5 comes down again, and that would put a local whatever the best open model of that size is next year within reach of under-US$5K hardware
the odd thing is that there isn't much in this whole range between 128GB and 512GB VRAM requirement to justify the huge premium you pay for Macs in that range - but this can change at any point as every other day there are announcements
I don't really care about being able to self host these models, but getting to a point where the hosting is commoditised so I know I can switch providers on a whim matters a great deal.
Of course, it's nice if I can run it myself as a last resort too.
It is pretty easy to set up Open Router and set up schemes to point at different models, but in the same token, you can point at yours locally unless you wanted a "more powerful" answer
Not concerned with electricity cost - I have solar + battery with excess supply where most goes back to the grid for $0 compensation (AU special).
But I did the napkin math on M3 Ultra ROI when DeepSeek V3 launched: at $0.70/2M tokens and 30 tps, a $10K M3 Ultra would take ~30 years of non-stop inference to break even - without even factoring in electricity. You clearly don't self-host to save money. You do it to own your intelligence, keep your privacy, and not be reliant on a persistent internet connection.
AFAIK they haven't released this one as OSS yet. They might eventually but its pretty obvious to me that at one point all/most those more powerful chinese models probably will stop being OSS.
> It's looking like we'll have Chinese OSS to thank for being able to host our own intelligence, free from the whims of proprietary megacorps.
I don’t know where you draw the line between proprietary megacorp and not, but Z.ai is planning to IPO soon as a multi billion dollar company. If you think they don’t want to be a multi billion dollar megacorp like all of the other LLM companies I think that’s a little short sighted. These models are open weight, but I wouldn’t count them as OSS.
Also Chinese companies aren’t the only companies releasing open weight models. ChatGPT has released open weight models, too.
> Also Chinese companies aren’t the only companies releasing open weight models. ChatGPT has released open weight models, too.
I was with you until here. The scraps OpenAI has released don't really compare to the GLM models or DeepSeek models (or others) in both cadence and quality (IMHO).
our laptops, devices, phones, equipments, home stuff are all powered by Chinese companies.
It wouldn't surprise me if at some point in the future my local "Alexa" assistant will be fully powered by local Chinese OSS models with Chinese GPUs and RAM.
Yeah it's funny how the needle has moved on this kind of thing.
Two years ago people scoffed at buying a personal license for e.g. JetBrains IDEs which netted out to $120 USD or something a year; VS Code etc took off because they were "free"
But now they're dumping monthly subs to OpenAI and Anthropic that work out to the same as their car insurance payments.
There's also zero incentive for individual companies to care: if I only want to use opus in VS code (and why would I use anything else, it's so much better at the job) I can either pay for copilot, which has excellent VS Code integration (because it has to), or I can pay Claude specifically and then use their extension which has the absolute worst experience because not only is the chat "whimsical, to make AI fun!", its interface is pat of the sidebar, so it's mutually exclusive with your file browser, search, etc.
So whether you pay Claude or GitHub, Claude gets paid the same. So the consumer ends up footing a bill that has no reason to exist, and has no real competition because open source models can't run at the scale of an Opus or ChatGPT.
(not unless the EU decides it's time for a "European Open AI Initiative" where any EU citizen gets free access to an EU wide datacenter backed large scale system that AI companies can pay to be part of, instead of getting paid to connect to)
I'm not sure being beholden to the whims of the Chinese Communist Party is an iota better than the whims of proprietary megacorps, especially given this probably will become part of a megacorp anyway.
It seems you missed the point entirely once you saw the word "Chinese". The point isn't that the models are from China. It's that the weights are open. You can download the weights and finetune them yourself. Nobody is beholden to anything.
Yeah that sounds great until it's running as an autonomous moltbot in a distributed network semi-offline with access to your entire digital life, and China sneaks in some hidden training so these agents turn into an army of sleeper agents.
Lol wat? I mean you certainly have enough control self hosting the model to not let it join some moltbot network... or what exactly are you saying would happen?
We just saw last week people are setting up moltbots with virtually no knowledge of what it has and doesn't have access. The scenario that i'm afraid of is China realizes the potential of this. They can add training to the models commonly used for assistants. They act normal, are helpful, everything you'd want a bot to do. But maybe once in a while it checks moltbook or some other endpoint China controls for a trigger word. When it sees that, it kicks into a completely different mode, maybe it writes a script to DDoS targets of interest, maybe it mines your email for useful information, maybe the user has credentials to some piece that is a critical component of an important supply chain. This is not a wild scenario, no new sci-fi technology would need to be invented. Everything to do it is available today, people are configuring it, and using it like this today. The part that I fear is if it is running locally, you can't just shut off API access and kill the threat. It's running on it's own server, it's own model. You have to cut off each node.
Big fan of AI, I use local models A LOT. I do think we have to take threats like this seriously. I don't Think it's a wild scifi idea. Since WW2, civilians have been as much of an equal opportunity target as a soldier, war is about logistics, and civilians supply the military.
Fair point but I would be more worried about the US government doing this kind of thing to act against US citizens than the Chinese government doing it.
I think we're in a brief period of relative freedom where deep engineering topics can be discussed with AI agents even though they have potential uses in weapons systems. Imagine asking chat gpt how to build a fertilizer bomb, but apply the same censorship to anything related to computer vision, lasers, drone coordination, etc.
There was research last year [0] finding significant security issues with the Chinese-made Unitree robots, apparently being pre-configured to make it easy to exfiltrate data via wi-fi or BLE. I know it's not the same situation, but at this stage, I wouldn't blame anyone for "absurd threat porn fantasy" - the threats are real, and present-day agentic AI is getting really good at autonomously exploiting vulnerabilities, whether it's an external attacker using it, or whether "the call is coming from inside the house".
Big fan of Salvatore's voxtral.c and flux2.c projects - hope they continue to get optimized as it'd be great to have lean options without external deps. Unfortunately it's currently too slow for real-world use (AMD 7800X3D/Blas) when adding Voice Input support to llms-py [1].
In the end Omarchy's new support for voxtype.io provided the nicest UX, followed by Whisper.cpp, and despite being slower, OpenAI's Whisper is still a solid local transcription option.
Also very impressed with both the performance and price of Mistral's new Voxtral Transcription API [2] - really fast/instant and really cheap ($0.003/min), IMO best option in CPU/disk-constrained environments.
Hi! This model is great, but it is too big for local inference, Whisper medium (the "base" IMHO is not usable for most things, and "large" is too large) is a better deal for many environments, even if the transcription quality is noticeable lower (and even if it does not have a real online mode). But... It's time for me to check the new Qwen 0.6 transcription model. If it works as well as their benchmarks claim, that could be the target for very serious optimizations and a no deps inference chain conceived since the start for CPU execution, not just for MPS. Since, many times, you want to install such transcription systems on server rent online via Hetzner and other similar vendors. So I'm going to handle it next, and if it delivers, really, time for big optimizations covering specifically the Intel, AMD and ARM instructions sets, potentially also thinking at 8bit quants if the performance remain good.
Same experience here with Whisper, medium is often not good enough. The large-turbo model however is pretty decent and on Apple silicon fast enough for real time conversations. The addition of the prompt parameter can also help with transcription quality, especially when using domain specific vocabulary. In general Whisper.cpp is better with transcribing full phrases than with streaming.
And not to forget, for many use cases more than just English is needed. Unfortunately right now most STT/ASR and TTS focus on English plus 0-10 other languages. Thus being able to add with reasonable effort more languages or domain specific vocabulary would be a huge plus for any STT and TTS.
Not sure how it works in other OS's but in Omarchy [1] you hold down `Super + Ctrl + X` to start recording and release it to stop, while it's recording you'll see a red voice recording icon in the top bar so it's clear when its recording.
Although as llms-py is a local web App I had to build my own visual indicator [2] which also displays a red microphone next to the prompt when it's recording. It also supports both Tap On/Off and hold down for recording modes. When using voxtype I'm just using the tool for transcription (i.e. not Omarchy OS-wide dictation feature) like:
$ voxtype transcribe /path/to/audio.wav
If you're interested the Python source code to support multiple voice transcription backends is at: [3]
(I keep coming back to this one so I've got half a dozen messages on HN asking for the exact same thing!).
It's a shame, whisper is so prevalent, but not great at actual streaming, but everyone uses it.
I'm hoping one of these might become a realtime de facto standard so we can actually get our realtime streaming api (and yep, I'd be perfectly happy with something just writing to stdout. But all the tools always end up just batching it because it's simpler!)
I've shipped an Omarchy MCP Server that lets AI Assistants manage your Omarchy desktop themes - switch wallpapers, change color schemes, toggle dark mode and more, all from natural language:
[1] https://bearssl.org
reply