I feel like these types of benchmarks mostly test what the mainstream tells you is possible and or works. You rarely get to hear the different implementations of inference engines, different types of setups (e.g. hybrid ones or older hardware) and their pros and cons. Also it seems these were tested with very minimal token usage.
Currently on my 9 year old laptop with a GTX 1070 Max-Q 8GB vram, 16 gb ddr4 ram and i7 8750h. I am running a finetuned Qwen 3.5-35B-A3B (Tiel-Coder-35B-A3B-GGUF-MTP) with 4-bit quant on a llamacpp fork (Turboquant) with 131k context at around 16-19 tk/s and 350-500 tk/s prefil using Pi harness. It is amazing for writing scripts and doing medium coding tasks and summarizing text for me. After around 100k context it goes down to 9 tk/s but that is to be expected.
If you ever ask any frontier LLM or these types of benchmarks, they would probably say that the best model I can run on my system is a crappy 9B model or something, which is just completely false. So I really wonder how much more you can pull out from more powerful systems and much better open weight models.
Locally the MOE based models get a lot more out of the hardware than a dense model. Qwen 3.8-27B is the current golden child if you have >24GB VRAM but because its a dense model it is pretty slow to run and requires a lot of hardware. A lot of the guides are very much focussed on getting the dense models to run but in practice the MOE models with few active parameters are just faster and easier to run without too much loss of intelligence, they behave a lot more like their peak size than the active parameter size.
RAM is going to be the big bottleneck for MOE (at least for now) as it will determine the quantisations you could run.
On a 3080 10GB and 32GB of RAM I can run the qwen-next coder model and that is a 127B model and it gets 40-50 Token/s with 65k context. That is a lot more usable than a Quantised Qwen3.8-27B in my experience both because its faster and more capable.
I am looking forward to Qwen4 soon, I am hoping we get some more 27-35B MOE models, those are going to be really good for local hardware.
> in practice the MOE models with few active parameters are just faster and easier to run without too much loss of intelligence
Is this true for coding tasks? Genuinely asking because I don't yet have the hardware to be able to run/compare. I thought I remembered the sentiment that the 27B dense was better than the 35B MoE
It really depends on which models you are comparing. These days, parameter count and architecture really don't do a great job of predicting the model's capabilities.
The reigning champion for a fast, high-intelligence LLM right now is Qwen 3.8 Flash Next at 125B + 50B of embeddings (can be offloaded to SSD for ~free), 6B active parameters. It's somewhere in the Opus 4.6-4.8 range, noticeably stronger than 3.8 27B, despite having only 6B active. Compare Meta's Muse Spark Open, which has 816B total and 42B active, and does not even reach the performance of Flash Next.
The dense is better in its result, but is out of reach for most hardware compared to the MoE models, and the MoE models produce output tokens much much faster.
The A3B part of that model means that each "expert" in the mixture of experts (MOE) model use 3B parameters. At 4-bit quant, you can fit 2 parameters per byte. That gives you 1.5GB per expert (plus a bit for overhead), which makes it usable on an 8GB card.
It's all about trade-offs: how much of the model you want to load into VRAM; how much VRAM you want to reserve for context; how much quantization you/the model can tolerate; how many tokens/second you can tolerate; whether you want to use the VRAM for anything else (graphics, Text-to-Speech, etc.); etc.
A3B refers to 3 billion parameters active during a forward pass, not the size of an individual expert.
The 3 billion includes all active experts, and then other shared weights, attention components, embeddings.
In the qwen 3.5 35b a3b series, each expert is ~3 million parameters. There are 40 layers which each have 1 shared expert and 8 routed experts. Roughly 1.1B of the 3B active are experts.
Yes, this is all good knowledge in theory for testing local ai. There are some more technical stuffs to consider that are happening in the background that is well outside my expertise to explain. I have been testing every MoE model I can fit on my system every year to find out which is the best for me.
Last year, it would have been impossible to run a 35B A3B model on old Pascual gpu and make it somehow usable. But for some reason now it just works even with a better finetuned version. So clearly there is more to it that just having it theoretically fit the system. Look for example how Strata works on certain models and systems compared to its llamacpp counterpart.
Thanks for sharing! I would love for some go to source for this type of insider info from the people who actually have tried. (The best I've found so far are some guys on Twitter/X like Alex Fateev (https://x.com/superalesha) and AJ (https://x.com/ItsmeAjayKV))
I'm personally sitting on an extremely limited Geforce GTX 1660 6GB, and would like to see how far one could go even with such a crappy card.
At least it was quite amazing to see the difference between CPU-based and even this crappy GPU-based inference when trying out some tiny models in jan.ai. It's still like orders of magnitude faster at least.
Try to use your IGPU as well if you have enough RAM, almost everyone has dual GPUs in their computers.
Use Vulkan build of llama.cpp to run low parameter MoE models in your IGPU and CUDA llama.cpp to run dense models in your DGPU.
Ofcourse you need to set your expectations right on what problem you want to solve, I've not been happy with using quantized models for coding even on a decent laptop; where as for NLP, ASR, STT etc. work great with this setup.
I am running on a Ryzen 5950X, 64GB DDR4 Ram and a Radeon RX 6700 XT with 12GB... I don't expect wonders, but I have between 11 and 120 (!) t/s.
I told it an open ended question: "What would you improve in this file at {location} in {bigger project}" and the context window seems to be too small. It's running in circles.
I must be doing something wrong.
Again, I am not expecting wonders and I am totally willing to accept that "I am holding it wrong".
Ideally you want to find something that allows a context window for at least 65k for agent work. Over 100k is better but 32k is about the bare minimum for anything in a harness/agent like Pi. You might have to drop to a lower quantisation to make that possible but context is the key for longer lots of thinking questions like this.
If you drop the thinking level to low or even off the context will go further at the cost of the models intelligence. That might also help you get some answer but since such a question will read so many files into that context its the wrong approach. This is one of the bigger limitations of local LLMs, you can't do work that spans across a lot of files because you are almost always context limited. There is a limit of 1M even with the frontier so even that can't review an entire project. Focussing in on individual files or small groups is more achievable.
So just to be specific the -c needs to be >65536 and depending on the size of what is in the folder might very well not be enough.
Your setup is definitely a monster compared to mine so it should give much better results. I also believe you run dense models with decent results with that setup.
Not sure what llamacpp build you are using but on mine using this fork: https://github.com/TheTom/llama-cpp-turboquant on commit 407f3237bfb3eeaff61546797de3d8c1a96be748 compiled for Pascal, I am able to push my current system to 131k context with minimal loss in quality. For reference this is my config.ini file for this model (Cyber is the better uncensored version btw):
would drop it from fp16 to 8 bits. You don't want to go below 8 bits if you can possibly avoid it and if you really do need to squeeze just drop the v value to q4_0 but leave the key alone. Unlike model quantisation this really makes things worse quickly.
Yeah, you need both a smaller system prompt than OpenCode uses, and a larger context configured in llama.cpp. It's probably running into compaction cycles before it even gets to start thinking about the task at hand.
You'd have better luck with pi instead of OpenCode, and see how far you can get that context window higher. For any halfway actual work you need bare minimum 64k, ideally 128k+. You can use 8bit kv cache, despite whatever folklore you may hear otherwise.
It sounds great and I hope local LLM/AI will be more powerful and available even on old hardware.
I tried llama.cpp with offloading "all what could be" to my GPU (GTX 750 Ti 2 GB). I could run local Qwen2.5-Coder quite fast (I do not remember but something like 25-30 t/s). However this model generates bad code (I tried Python and JavaScript). So, for my setup it is better to chat with free ChatGPT and do copying-pasting between web chat and code editor.
Is it even worth for trying something on my GPU? What do you think?
And yes, I use Claude Code (Opus 5.5) and Codex in my job but for personal stuff I would prefer trying free solutions (it is the reason I tried local LLM). I do not want to buy subs, at leats for now.
Models and inference engines are getting better and more optimized everyday so it is definitely worth experimenting with your GPU. You should probably check on newer finetuned and or REAP models Huggingface that could fit on your system.
I believe in your case the model quality matters a lot. To my (trust me bro) experience any MOE model, finetuned or not, with params < 19B even with Q4-quant always give lackluster results. So any Qwen2.5 wouldn't suffice to even simple coding tasks. For your case maybe you can bump it up to a Gemma4 E2B/E4B or Qwen3.5-9B. Also make sure that your current build of llamacpp actually uses your GPU with CUDA.
Yes, I was using CUDA toolkit on my Debian. Anyway, thanks for reply. I think I could try something but definitely it looks like it will not have better quality, even for writing simple code function by function :P
Bench is laughably out sync with reality too. I run things like flash next on an rtx 4080 that it says “won’t run” and its rates are just all over the place. Pure slop.
By running vllm at work I learned that you will usually outperform benchmarks which you will find on the internet. Mostly by running proper parameter tuning. So you get impressive numbers. Still in reality those numbers do not tell the whole truth. We found that real usage is more 60k-140k prompt with a few k of output for realistic agentic coding workloads. Most of the time that will give you half to a third of decode throughput vs the benchmark. It also becomes much more about cache aware routing and kv cache management (how to offload and restore). Because prefilling 60-140k will simply take too long for interactive work.
It says 46 tok/sec for Qwen3.6-27b 4 bit on RTX5090, but I get 60-70 tok/sec on a small context like that (5 bit weights with 8 bit KV cache), although it does slow down to 40 tok/sec on large contexts. If I add ngram speculation, then it can increase to 100+ tok/sec on code.
There are many knobs to tweak to make it faster, which the site ignores.
Prices for everything are all in flux at this moment. Consumer cards like RTX 5090 are bonkers. Even 3090s have like doubled in the last six months. Meanwhile, prices for renting this class of hardware on vast.ai are spiking too, with rates increasing like 50% in September alone.
It's hard to know if we're at the beginning or the end of this trend. If manufacturing continues to focus on high end datacenter gpus at $40k a pop, and we continue to have supply constraints, this could last a while.
On the other hand, if Chinese models continue to nip at the tails of the frontier, cheaper by 20x, running on new Chinese chips, the whole bottom could fall out of the US market, and things could get very weird.
around this time, i was thinking about running a local LLM for the company i work for. i'm trying to figure out what kind of hardware i'd need and what level of performance i could expect. i tried a few things on your website but how much do the results reflect real world usage?
in my use case, the model might need to analyze a given task first then generate and execute a database query and finally send the results back to the client. to what extent could your setup handle this kind of workflow and how accurate are the results?
This is a great idea but I'm far less concerned about output token speed and way more about the ability to parallelize prefill - for agentic coding that's pretty much way more important than output speed.
Projects like Strata show that this sort of thing will become outdated very fast (unless they can start including estimations for stuff like that as well).
Increasingly, people are finding clever technical solutions to get the most out of whatever hardware you have.
just go rent one for a few minutes and try it out ... it'll cost you well under $0.50 if you're just trying to get a feel for if it's right for you. There's a lot of things like context window that this doesn't capture.
Depending on who you sign up for, that's within the introductory credits (usually like 5-10$) so it's actually $0.
yeah opencode just runs with it with lmstudio as llm engine.
It won't win benchmarks of course, but 20 tps generation is okay and letting it chug away on my local kubernetes cluster to troubleshoot is pretty nice. First start is pretty slow though, parsing is around 100-200 tps. Which is crap for long questions or harnesses like opencode which start around with thousands of tokens. Once cache is warm, pretty enjoyable.
I think it could have way more use cases where it's just 'good enough'. Email categorization, voice commands for the home. 1 year ago this was still a very miserable experience, but qwen 9B + better cache quantization and more tricks in the latest llm engines makes it viable.
Will it do difficult architectural refactors by just giving a one-shot prompt like opus allows, nope, not gonna happen.
On 8GB VRAM (RTX 2060 Super) + 64GB DRAM I can run Qwen 3.8 Flash Next in usable speeds. The last 2 weeks have resulted in quite some speed ups in Strata land.
Yeah simple, but menial refactors where IDEs don't have clear cut refactors yet. Or languages like CUE which has awesome tooling, but not so great refactors in its lsp yet.
My take as well. The dual-3090 and dual-spark configs have an outsized voice in this conversation. I'd add pipeline parallelism vs tensor parallelism and in the case of TP setting your lanes and speeds per card.
Would love to have multi-gpu support eventually but it's a balance between whats popular and what actually has good quality benchmarks available. Multi-gpu numbers can also vary between setups
1. It is difficult to figure out if dropping $10k+ on hardware will make a meaningful difference for running local models.
2. Downloading each model and each quantization of that model to see the speed seems rude to free services like hugging face, so now I can see things like "with my hardware, going to a Q8 only results in 3 more tokens per second"
My one complaint is picking my hardware (9070XT) and Qwen3.6 27B FP16 results in the message:
> This combination won't run. Qwen 3.6/3.8 27B at FP16 needs 54.6 GB — the AMD RX 9070 XT 16GB has 14.9 GB usable. No quantization of this model fits; pick a smaller model or more memory.
While it is true that it won't fit in your GPU memory, it demonstrably does run on my hardware. Llama.cpp supports putting some layers on the GPU and running others on the CPU.
Also, that message says "No quantization of this model fits" but switching to a different quantization of that model (for example, Q8_0) makes the message go away?
It is nice to have simulated confirmation of the M4 Max 128gb limitations on Gemma4 31B vs Gemma4 26B A4B at Q4:
Time to first token (TTFT) on 5,687 tokens, like pasting a short code block into the prompt, takes 36 seconds vs 4.7 seconds. It also breaks down the seconds for each step like KV and prefilling.
Gemma4 31B TTFT on the M5 Max 128gb with same setup is showing as 19 seconds.
What surprised me is how fast the Mac Studio is and how slow the Spark is. I thought they were at least comparable, but the former is getting 20 tok/s and the latter 4.6 tok/s with e.g. Qwen 3.8 27B and Q8.
It would be great if numbers weren't off, would include optimised versions (ie. speculative decoding), context size, include allowed parallel inference (single agents use parallel inference a lot) etc.
I don't think in general people care about long list of models, they just care about the best one.
Also dual gpu is common setup, there should be ie. 2x 4090.
A random number I checked feels kinda optimistic, or there's something wrong with my setup; Gemma 26B-A4B on a 16GB 5060Ti is reported to get 21-88tok/s with Q4_K_M, meanwhile I can just barely eke out 70tok/s with Q3 (which fully fits in vram), and 40tok/s with Q4 (with recently-merged MoE cache on llama-cpp). I guess the average fits, but the peak feels implausible.
No clue about details, but Sebastian Raschka's site reports its kv cache usage at 210kB per token, which is like 6x more than this site claims ("0.273 GB per 8k"). Also, the site claims qwen 27b to use 2x more memory per token than gemma, which is the opposite from what I know.
Software is improving quickly so take these numbers as a snapshot performance at best. Strata has really shown how much performance is left on the table. I have a dual A4000 16gb (Ampere) on an old X99 dual Xeon E5-2680 v4 with 128gb of memory and Qwen 3.8 Flash Next at the IQ3-S size (~96Gb) runs at 65-75 t/s which is extremely useable.
Model list is a bit limited, and whole harness affects the speed, so the libs used in harness and their versions might be useful to show for each individual model setup.
I’ll look into adding harnesses. The benchmark sources usually don’t contain harness so the problem is limited data. What models would you like to see?
He's exagerating, I hope. The only bottleneck a harness attachs is pre-context fill and whatever tooling is installed. A proper harness reduces these to negligable startup costs.
Unless of course he's driving some weird lambo that dumps 200k of context into their chats. Who knows, I feel like we're on a yellow submarine.
Exagerating? I setup llm endpoints on corporate level in a company on H200 gpus, that serves hundreds of software engineers, I have observed differences, even on different versions of nvidia drivers. I shared my day to day experience. It is more and more important if you have very limited VRAM like 24gb on a 4090 RTX desktop or RTX 5090 Laptop gpu. I set up and use llms on my personal devices also, when I can't use cloud models on classified company materials. Here is an example problem what I daily report and deal with.
Note that this assumes a single chat (single cached context). As soon as you want to use it for several tasks in parallel or in frequent alternation, hardware requirements increase or performance drops.
I feel like these types of benchmarks mostly test what the mainstream tells you is possible and or works. You rarely get to hear the different implementations of inference engines, different types of setups (e.g. hybrid ones or older hardware) and their pros and cons. Also it seems these were tested with very minimal token usage.
Currently on my 9 year old laptop with a GTX 1070 Max-Q 8GB vram, 16 gb ddr4 ram and i7 8750h. I am running a finetuned Qwen 3.5-35B-A3B (Tiel-Coder-35B-A3B-GGUF-MTP) with 4-bit quant on a llamacpp fork (Turboquant) with 131k context at around 16-19 tk/s and 350-500 tk/s prefil using Pi harness. It is amazing for writing scripts and doing medium coding tasks and summarizing text for me. After around 100k context it goes down to 9 tk/s but that is to be expected.
If you ever ask any frontier LLM or these types of benchmarks, they would probably say that the best model I can run on my system is a crappy 9B model or something, which is just completely false. So I really wonder how much more you can pull out from more powerful systems and much better open weight models.
Locally the MOE based models get a lot more out of the hardware than a dense model. Qwen 3.8-27B is the current golden child if you have >24GB VRAM but because its a dense model it is pretty slow to run and requires a lot of hardware. A lot of the guides are very much focussed on getting the dense models to run but in practice the MOE models with few active parameters are just faster and easier to run without too much loss of intelligence, they behave a lot more like their peak size than the active parameter size.
RAM is going to be the big bottleneck for MOE (at least for now) as it will determine the quantisations you could run.
On a 3080 10GB and 32GB of RAM I can run the qwen-next coder model and that is a 127B model and it gets 40-50 Token/s with 65k context. That is a lot more usable than a Quantised Qwen3.8-27B in my experience both because its faster and more capable.
I am looking forward to Qwen4 soon, I am hoping we get some more 27-35B MOE models, those are going to be really good for local hardware.
> in practice the MOE models with few active parameters are just faster and easier to run without too much loss of intelligence
Is this true for coding tasks? Genuinely asking because I don't yet have the hardware to be able to run/compare. I thought I remembered the sentiment that the 27B dense was better than the 35B MoE
It really depends on which models you are comparing. These days, parameter count and architecture really don't do a great job of predicting the model's capabilities.
The reigning champion for a fast, high-intelligence LLM right now is Qwen 3.8 Flash Next at 125B + 50B of embeddings (can be offloaded to SSD for ~free), 6B active parameters. It's somewhere in the Opus 4.6-4.8 range, noticeably stronger than 3.8 27B, despite having only 6B active. Compare Meta's Muse Spark Open, which has 816B total and 42B active, and does not even reach the performance of Flash Next.
The dense is better in its result, but is out of reach for most hardware compared to the MoE models, and the MoE models produce output tokens much much faster.
The A3B part of that model means that each "expert" in the mixture of experts (MOE) model use 3B parameters. At 4-bit quant, you can fit 2 parameters per byte. That gives you 1.5GB per expert (plus a bit for overhead), which makes it usable on an 8GB card.
It's all about trade-offs: how much of the model you want to load into VRAM; how much VRAM you want to reserve for context; how much quantization you/the model can tolerate; how many tokens/second you can tolerate; whether you want to use the VRAM for anything else (graphics, Text-to-Speech, etc.); etc.
A3B refers to 3 billion parameters active during a forward pass, not the size of an individual expert.
The 3 billion includes all active experts, and then other shared weights, attention components, embeddings.
In the qwen 3.5 35b a3b series, each expert is ~3 million parameters. There are 40 layers which each have 1 shared expert and 8 routed experts. Roughly 1.1B of the 3B active are experts.
Yes, this is all good knowledge in theory for testing local ai. There are some more technical stuffs to consider that are happening in the background that is well outside my expertise to explain. I have been testing every MoE model I can fit on my system every year to find out which is the best for me.
Last year, it would have been impossible to run a 35B A3B model on old Pascual gpu and make it somehow usable. But for some reason now it just works even with a better finetuned version. So clearly there is more to it that just having it theoretically fit the system. Look for example how Strata works on certain models and systems compared to its llamacpp counterpart.
Thanks for sharing! I would love for some go to source for this type of insider info from the people who actually have tried. (The best I've found so far are some guys on Twitter/X like Alex Fateev (https://x.com/superalesha) and AJ (https://x.com/ItsmeAjayKV))
I'm personally sitting on an extremely limited Geforce GTX 1660 6GB, and would like to see how far one could go even with such a crappy card.
At least it was quite amazing to see the difference between CPU-based and even this crappy GPU-based inference when trying out some tiny models in jan.ai. It's still like orders of magnitude faster at least.
Try to use your IGPU as well if you have enough RAM, almost everyone has dual GPUs in their computers.
Use Vulkan build of llama.cpp to run low parameter MoE models in your IGPU and CUDA llama.cpp to run dense models in your DGPU.
Ofcourse you need to set your expectations right on what problem you want to solve, I've not been happy with using quantized models for coding even on a decent laptop; where as for NLP, ASR, STT etc. work great with this setup.
I might write more about this setup soon.
Interesting advice, gotta look into that, thank you!
This tool has very tiny quants available for each model, and defaults to the tiny quants, so I'm not so sure you... tried it?
Ya but the library matters as well and what will get 20tps on one might get 30-40tps on another, especially with finicky cards like Intel arc.
I need to play around with this.
I am running on a Ryzen 5950X, 64GB DDR4 Ram and a Radeon RX 6700 XT with 12GB... I don't expect wonders, but I have between 11 and 120 (!) t/s.
I told it an open ended question: "What would you improve in this file at {location} in {bigger project}" and the context window seems to be too small. It's running in circles.
I must be doing something wrong.
Again, I am not expecting wonders and I am totally willing to accept that "I am holding it wrong".
llama-server \ -m ~/models/tiel-coder/Tiel-Coder-35B-A3B-MTP-UD-IQ4_XS.gguf \ -ngl 20 \ -c 16384 \ --jinja \ --alias tiel-coder
using it in opencode
{ "$schema": "https://opencode.ai/config.json", "provider": { "llamacpp": { "npm": "@ai-sdk/openai-compatible", "name": "Local llama.cpp", "options": { "baseURL": "http://127.0.0.1:8080/v1" }, "models": { "tiel-coder": { "name": "Tiel-Coder 35B A3B", "limit": { "context": 16384, "output": 4096 } } } } }, "model": "llamacpp/tiel-coder" }
Ideally you want to find something that allows a context window for at least 65k for agent work. Over 100k is better but 32k is about the bare minimum for anything in a harness/agent like Pi. You might have to drop to a lower quantisation to make that possible but context is the key for longer lots of thinking questions like this.
If you drop the thinking level to low or even off the context will go further at the cost of the models intelligence. That might also help you get some answer but since such a question will read so many files into that context its the wrong approach. This is one of the bigger limitations of local LLMs, you can't do work that spans across a lot of files because you are almost always context limited. There is a limit of 1M even with the frontier so even that can't review an entire project. Focussing in on individual files or small groups is more achievable.
So just to be specific the -c needs to be >65536 and depending on the size of what is in the folder might very well not be enough.
Your setup is definitely a monster compared to mine so it should give much better results. I also believe you run dense models with decent results with that setup.
Not sure what llamacpp build you are using but on mine using this fork: https://github.com/TheTom/llama-cpp-turboquant on commit 407f3237bfb3eeaff61546797de3d8c1a96be748 compiled for Pascal, I am able to push my current system to 131k context with minimal loss in quality. For reference this is my config.ini file for this model (Cyber is the better uncensored version btw):
[*] # Global defaults jinja = 1 metrics = 1 threads = 12 threads-batch = 12 flash-attn = on sleep-idle-seconds = 99999 timeout = 99999 no-mmap = 1 mlock = 1 ctk = turbo4 ctv = turbo2
[qwen36-a3b] model = /root/llm/models/Cyber-Tiel-Coder-35B-A3B-MTP-UD-IQ4_XS.gguf alias = qwen36-a3b n-cpu-moe = 32 temp = 0.6 top-p = 0.95 top-k = 20 min-p = 0.0 presence-penalty = 0.0 repeat-penalty = 1.0 reasoning = on n-predict = 32768 reasoning-format = deepseek reasoning-budget = 16384 reasoning-budget-message = "Proceed to final answer" reasoning-preserve = true swa-full = true np = 2 batch-size = 4096 ubatch-size = 2048 cache-reuse = 256 ngl = auto ctx-checkpoints = 33 ctx-size = 262144 no-mmproj = true no-mmproj-offload = true spec-type = draft-mtp spec-draft-n-max = 2
I am also certain that these flags could be improved somehow depending on your system.
I would give it significantly more context, if you don't have enough room you can quantize the kv cache with some flags that I don't remember ottomh.
I would go ahead and give it at least 64k but the more the better really!
I run the kv cache at q8 for keys and q4 for values, but going below q8 is generally not recommended.
--cache-type-k q8_0 --cache-type-v q8_0
would drop it from fp16 to 8 bits. You don't want to go below 8 bits if you can possibly avoid it and if you really do need to squeeze just drop the v value to q4_0 but leave the key alone. Unlike model quantisation this really makes things worse quickly.
Yeah, you need both a smaller system prompt than OpenCode uses, and a larger context configured in llama.cpp. It's probably running into compaction cycles before it even gets to start thinking about the task at hand.
You'd have better luck with pi instead of OpenCode, and see how far you can get that context window higher. For any halfway actual work you need bare minimum 64k, ideally 128k+. You can use 8bit kv cache, despite whatever folklore you may hear otherwise.
It sounds great and I hope local LLM/AI will be more powerful and available even on old hardware.
I tried llama.cpp with offloading "all what could be" to my GPU (GTX 750 Ti 2 GB). I could run local Qwen2.5-Coder quite fast (I do not remember but something like 25-30 t/s). However this model generates bad code (I tried Python and JavaScript). So, for my setup it is better to chat with free ChatGPT and do copying-pasting between web chat and code editor.
Is it even worth for trying something on my GPU? What do you think?
And yes, I use Claude Code (Opus 5.5) and Codex in my job but for personal stuff I would prefer trying free solutions (it is the reason I tried local LLM). I do not want to buy subs, at leats for now.
Models and inference engines are getting better and more optimized everyday so it is definitely worth experimenting with your GPU. You should probably check on newer finetuned and or REAP models Huggingface that could fit on your system.
I believe in your case the model quality matters a lot. To my (trust me bro) experience any MOE model, finetuned or not, with params < 19B even with Q4-quant always give lackluster results. So any Qwen2.5 wouldn't suffice to even simple coding tasks. For your case maybe you can bump it up to a Gemma4 E2B/E4B or Qwen3.5-9B. Also make sure that your current build of llamacpp actually uses your GPU with CUDA.
Yes, I was using CUDA toolkit on my Debian. Anyway, thanks for reply. I think I could try something but definitely it looks like it will not have better quality, even for writing simple code function by function :P
I haven't managed to run anything on my 8GB card that isn't blown away by free tier Gemini.
Have you tried the free models on OpenCode Zen?
Bench is laughably out sync with reality too. I run things like flash next on an rtx 4080 that it says “won’t run” and its rates are just all over the place. Pure slop.
By running vllm at work I learned that you will usually outperform benchmarks which you will find on the internet. Mostly by running proper parameter tuning. So you get impressive numbers. Still in reality those numbers do not tell the whole truth. We found that real usage is more 60k-140k prompt with a few k of output for realistic agentic coding workloads. Most of the time that will give you half to a third of decode throughput vs the benchmark. It also becomes much more about cache aware routing and kv cache management (how to offload and restore). Because prefilling 60-140k will simply take too long for interactive work.
It says 46 tok/sec for Qwen3.6-27b 4 bit on RTX5090, but I get 60-70 tok/sec on a small context like that (5 bit weights with 8 bit KV cache), although it does slow down to 40 tok/sec on large contexts. If I add ngram speculation, then it can increase to 100+ tok/sec on code.
There are many knobs to tweak to make it faster, which the site ignores.
Prices for everything are all in flux at this moment. Consumer cards like RTX 5090 are bonkers. Even 3090s have like doubled in the last six months. Meanwhile, prices for renting this class of hardware on vast.ai are spiking too, with rates increasing like 50% in September alone.
It's hard to know if we're at the beginning or the end of this trend. If manufacturing continues to focus on high end datacenter gpus at $40k a pop, and we continue to have supply constraints, this could last a while.
On the other hand, if Chinese models continue to nip at the tails of the frontier, cheaper by 20x, running on new Chinese chips, the whole bottom could fall out of the US market, and things could get very weird.
around this time, i was thinking about running a local LLM for the company i work for. i'm trying to figure out what kind of hardware i'd need and what level of performance i could expect. i tried a few things on your website but how much do the results reflect real world usage?
in my use case, the model might need to analyze a given task first then generate and execute a database query and finally send the results back to the client. to what extent could your setup handle this kind of workflow and how accurate are the results?
good job.
Test models on OpenRouter first.
This is a great idea but I'm far less concerned about output token speed and way more about the ability to parallelize prefill - for agentic coding that's pretty much way more important than output speed.
Projects like Strata show that this sort of thing will become outdated very fast (unless they can start including estimations for stuff like that as well).
Increasingly, people are finding clever technical solutions to get the most out of whatever hardware you have.
just go rent one for a few minutes and try it out ... it'll cost you well under $0.50 if you're just trying to get a feel for if it's right for you. There's a lot of things like context window that this doesn't capture.
Depending on who you sign up for, that's within the introductory credits (usually like 5-10$) so it's actually $0.
Shame they are not including old crap like my RX580 which is doing some impressive stuff with Qwen 9B Models.
Tell us more about what your RX580 can do with the small Qwen models. What workflows are you doing, and what kind of performance are you seeing?
yeah opencode just runs with it with lmstudio as llm engine.
It won't win benchmarks of course, but 20 tps generation is okay and letting it chug away on my local kubernetes cluster to troubleshoot is pretty nice. First start is pretty slow though, parsing is around 100-200 tps. Which is crap for long questions or harnesses like opencode which start around with thousands of tokens. Once cache is warm, pretty enjoyable.
I think it could have way more use cases where it's just 'good enough'. Email categorization, voice commands for the home. 1 year ago this was still a very miserable experience, but qwen 9B + better cache quantization and more tricks in the latest llm engines makes it viable.
Will it do difficult architectural refactors by just giving a one-shot prompt like opus allows, nope, not gonna happen.
Thank you
that's 8G vram? you're using it for coding?
On 8GB VRAM (RTX 2060 Super) + 64GB DRAM I can run Qwen 3.8 Flash Next in usable speeds. The last 2 weeks have resulted in quite some speed ups in Strata land.
Yep, it's 8GB vram from 2017 :D
Yeah simple, but menial refactors where IDEs don't have clear cut refactors yet. Or languages like CUE which has awesome tooling, but not so great refactors in its lsp yet.
I would like to see in such simulators:
- multi GPU configuration (hetero and homo)
- PCIe gen chooser (3.0/4.0/5.0)
- is nvlink enabled (2 cards, 2 x 2 cards, 3 x 2 cards)
- is PCIe bifurcated or CPU native (threadripper pro for example)
Nvlink just is not available above 3090 and below B200 (datacenter stuff). E.g. https://www.gpuyard.com/tutorials/howto/setup-nvlink-blackwe...
My take as well. The dual-3090 and dual-spark configs have an outsized voice in this conversation. I'd add pipeline parallelism vs tensor parallelism and in the case of TP setting your lanes and speeds per card.
Would love to have multi-gpu support eventually but it's a balance between whats popular and what actually has good quality benchmarks available. Multi-gpu numbers can also vary between setups
This is great. It solves two of my issues:
My one complaint is picking my hardware (9070XT) and Qwen3.6 27B FP16 results in the message:> This combination won't run. Qwen 3.6/3.8 27B at FP16 needs 54.6 GB — the AMD RX 9070 XT 16GB has 14.9 GB usable. No quantization of this model fits; pick a smaller model or more memory.
While it is true that it won't fit in your GPU memory, it demonstrably does run on my hardware. Llama.cpp supports putting some layers on the GPU and running others on the CPU.
Also, that message says "No quantization of this model fits" but switching to a different quantization of that model (for example, Q8_0) makes the message go away?
It is nice to have simulated confirmation of the M4 Max 128gb limitations on Gemma4 31B vs Gemma4 26B A4B at Q4:
Time to first token (TTFT) on 5,687 tokens, like pasting a short code block into the prompt, takes 36 seconds vs 4.7 seconds. It also breaks down the seconds for each step like KV and prefilling.
Gemma4 31B TTFT on the M5 Max 128gb with same setup is showing as 19 seconds.
What surprised me is how fast the Mac Studio is and how slow the Spark is. I thought they were at least comparable, but the former is getting 20 tok/s and the latter 4.6 tok/s with e.g. Qwen 3.8 27B and Q8.
There are also platform specific sites which have leaderboards for different models and setups. Having an OEM Spark, this one has been very helpful
https://sparkbench.dev/
It would be great if numbers weren't off, would include optimised versions (ie. speculative decoding), context size, include allowed parallel inference (single agents use parallel inference a lot) etc.
I don't think in general people care about long list of models, they just care about the best one.
Also dual gpu is common setup, there should be ie. 2x 4090.
A random number I checked feels kinda optimistic, or there's something wrong with my setup; Gemma 26B-A4B on a 16GB 5060Ti is reported to get 21-88tok/s with Q4_K_M, meanwhile I can just barely eke out 70tok/s with Q3 (which fully fits in vram), and 40tok/s with Q4 (with recently-merged MoE cache on llama-cpp). I guess the average fits, but the peak feels implausible.
No clue about details, but Sebastian Raschka's site reports its kv cache usage at 210kB per token, which is like 6x more than this site claims ("0.273 GB per 8k"). Also, the site claims qwen 27b to use 2x more memory per token than gemma, which is the opposite from what I know.
I suspect we're going to see a resurgence of local computing because of LLMs, although not in exactly the same way as before.
Software is improving quickly so take these numbers as a snapshot performance at best. Strata has really shown how much performance is left on the table. I have a dual A4000 16gb (Ampere) on an old X99 dual Xeon E5-2680 v4 with 128gb of memory and Qwen 3.8 Flash Next at the IQ3-S size (~96Gb) runs at 65-75 t/s which is extremely useable.
I have two cards from 2017 running flash next at 200+ t/s on the same model file. Engines are wildly under optimized still. :)
Model list is a bit limited, and whole harness affects the speed, so the libs used in harness and their versions might be useful to show for each individual model setup.
I’ll look into adding harnesses. The benchmark sources usually don’t contain harness so the problem is limited data. What models would you like to see?
He's exagerating, I hope. The only bottleneck a harness attachs is pre-context fill and whatever tooling is installed. A proper harness reduces these to negligable startup costs.
Unless of course he's driving some weird lambo that dumps 200k of context into their chats. Who knows, I feel like we're on a yellow submarine.
Exagerating? I setup llm endpoints on corporate level in a company on H200 gpus, that serves hundreds of software engineers, I have observed differences, even on different versions of nvidia drivers. I shared my day to day experience. It is more and more important if you have very limited VRAM like 24gb on a 4090 RTX desktop or RTX 5090 Laptop gpu. I set up and use llms on my personal devices also, when I can't use cloud models on classified company materials. Here is an example problem what I daily report and deal with.
https://github.com/ggml-org/llama.cpp/issues/27330
I can believe different drivers make a difference, but that wasn't the point under debate
For example you might add Strata & Qwen 3.8 Flash Next with some quantizations. Nowadays it is a go to selection and very sought after MoE model.
Qwen 3.8 flash next is already in the works and should be live soon.
New model/hardware benchmarks is now live
This is not accurate, I get like 80t/s on Q3.8 27b UD Q4_K_XL on my 4090 with 3 drafts and about 110K context window, sim shows mid 20s
I get about 50 t/s with the same model on a 3090, with about 100K context. I usually set harnesses to 90k.
It would be neat to also include NVFP4 variants for the Blackwell hardware.
Cool examples, this is interesting to test. I've noticed that OpenAI and Meta play a lot with animations and fades to mask the delays.
Cool idea but I'd like to also be able to set a custom value of tok/s, ignoring all the other options
A ton of RTX Spark (supposedly similar to DGX Spark) laptops just came out. We really need this for those.
Asus ProArt P14, Surface Ultra, Yoga Pro 9n, etc.
I am hoping the cheaper N1 models will also come out soon, but might be several months.
Also would be good to see a few more quant and model options.
I don't know what basis these numbers are estimated on, but my M2 Ultra 128GB runs almost twice as fast as what's reported for the M3 Ultra 512GB.
I wished there was such a tool before picking my hardware, it's useful. But the estimates needs to be somewhat correct.
Thank you for the feedback. What model/quant did you test so I can look in it?
Note that this assumes a single chat (single cached context). As soon as you want to use it for several tasks in parallel or in frequent alternation, hardware requirements increase or performance drops.
I mean I haven't checked but I'd guess it'd be possible to save a kv cache to an ssd or ram and then pull it back into vram when needed.
Doesn't seem to work? It just spits out a fixed answer about RAM. qwen3.8 27b
This is just a tok/s simulator. It displays the dummy text at appropriate speeds for configurations selected through buttons below the chatbox.
It would be cool if I could just search for a GPU instead of scrolling through the whole list. Overall, great website!
Depending on what hardware you're considering, you might rent a vast.ai box with similar specs for a month or two to get a feel for it.
Please modify to include multiples of each as well. Most serious people aren't going to use a single PCIe card if they can help it.
Would be nice if we could choose, say, multiple 4090s or otherwise customize the configuration (or maybe I just missed the button for that?)
How accurate are the simulations?
Missing: Dual Strix Halo, Dual Nvidia Spark
How about context length? Isn't it still just a few thousand tokens on local models?
No Qualcomm Snapdragon X Elite?
The hardware that you cant afford and will never pay for itself? Exciting!
Just chiming in that the idea is cool, but the speeds are way wrong
Can't go below 12GB?
Why can't tools show a progress bar for the prefilling step?
It would be nice if it would give a feel for the quality too ...
Doesn't provide for hybrid setups RTX5070 + RTX5080 for example.