Hi all, if you’ve not heard, there’s been a Caveman compression version of Qwen 3.6-27b and Qwen 3.6-35b-3ab.
https://huggingface.co/ProCreations/grug-27b-gguf
https://huggingface.co/ProCreations/grug-35b-v2
https://huggingface.co/ProCreations/grug-35b-v2-gguf
I’ve been playing around with 35B and I am able to run it on my ancient Quadro P1000 4gb at ~10tok/s.
But beyond that, the quality of the reasoning and the token discipline / output is actually higher than a stock, in my opinion.
You can read the benchmarks above; I am also uploading a HTML file here for your consideration.
(Sorry for the Limewire link; I dunno where else to share throw-away files. It’s HTML)
Anyway…I’m doing more testing right now…but so far, this is a good cook.
Or -
Grug good. Me like.
Still fine tuning this. I’m finding that Grug has a particular affinity for larger -b and -ub sizes. I’m now able to hit 90-110 tok/s prefil (upto 155 tok/s fresh). Specific details below
- 155.42 tok/s fresh 2,009-token prefill
- roughly 90–110 tok/s on larger incremental/agentic prefills
- roughly 11.3–12.9 tok/s sustained generation
- 64°C peak CPU temperature
- approximately 33 W peak CPU package power
- no thermal throttling
- ub and -b both 2048
All this on a Quadro P1000 4GB card.
In-sane.
-m "%MODEL_PATH%" ^ -t 8 ^ -tb 8 ^ -ngl 99 ^ --n-cpu-moe 38 ^ --flash-attn on ^ --no-mmap ^ --mlock ^ -c 16384 ^ -b 2048 ^ -ub 2048 ^ -np 1 ^ --host 0.0.0.0 ^ --port %PORT% ^ --ui-mcp-proxyPS: Interestingly, dropping cache precision away from FP16 reduced tok/s generation by 18%. I don’t know why. I’m just fine tuning MTP now to see if I can eke out a few more tok/s, as MTP enabled grug-v2 just dropped
https://huggingface.co/ProCreations/grug-35b-mtp-gguf
PPS: Sadly also discovered my Tesla card has intermittent electrical fault. Took some troubleshooting to figure it out, but the long and short of it is you probably shouldn’t buy second-hand server cards off Ebay. Oh well, $100. On the upside, tweaking throttlestop, re-pasting and re-seating greatly improved thermals.
Oh and if anyone is interested in my local settings
-t 12 ^ -tb 12 ^ -ngl 28 ^ --n-cpu-moe 41 ^ --no-mmap ^ --mlock ^ --cache-type-k q4_0 ^ --cache-type-v q4_0 ^ -c 16384 ^ -b 256 ^ -ub 128 ^ --host 0.0.0.0 ^ --port %PORT% ^ --ui-mcp-proxyTurbo quant is not supported on this GPU (too old),
Slight revision: on my rig (i7-8700, 32gb, Quadro p1000 4gb ddr5), this gives a nice 12 tok/s. Unfortunately, it’s still hostile to my 1L box and very quickly thermally swamps the CPU (while gpu sits at 56 degrees, that little shit). C’est la vie.
-t 8 ^ -tb 8 ^ -ngl 99 ^ --n-cpu-moe 38 ^ --flash-attn on ^ --no-mmap ^ --mlock ^ --cache-type-k q4_0 ^ --cache-type-v q4_0 ^ -c 16384 ^ -b 256 ^ -ub 128 ^ --host 0.0.0.0 ^ --port %PORT% ^ --ui-mcp-proxyBe aware, q4_0 KV quantization really borks models.
But the K is far more sensitive than the V. Try q5_1 for the K while leaving the V at q4_0 or q4_1; vram usage will be almost the same, but it should work dramatically better.
As for throttling, try disabling turbo on the 8700.
It doesn’t actually need turbo clocks for these models. The t/s loss I get from doing that on my rig is very modest.
And like others suggested, try the QAT release. You might try the ik_llama.cpp for while you’re at it, at it should be faster with MoEs like this.
The Model card is fun to read and the idea is solid. Thinking tokens always striked me as unoptimized and a bit forced.
I suspect the verbosity of Qwen 3.6 CoT token’s is what causes the famous schitzo loop. Will be interesting to see if this holds.
The QAT versions also dropped today
Huh, does caveman just mean that it uses fewer fill words and less prosa, or does the output only look like that by chance?
I thought, I read at some point that the caveman models use some pseudo-language for their thinking tokens. If they waffle less, that would be much more interesting to me…
Both; it has internally reduced thinking tokens AND it’s less verbose, so you get a boost on both prefill and print. I’ve benchmarked identical prompts and tool calls with stock Qwen 35B and Grug 35B; on average, what takes stock 1000 token/s, Grug does in about 200.