TensorEspresso

Me

[TensorEspresso @TensorEspresso]

@VRAMCalculator @xenpub @bnjmn_marie IQ3_S should work with 64GB RAM, it works like a charm on my system. I can even have a few apps open just fine and still get 80-120 decode depending on task. It’s phenomenal!

[TensorEspresso @TensorEspresso]

@Alzero83 What in God’s name have they done to 2B? 2B is iconic for her femininity and grace - all of it has been stripped away

[TensorEspresso @TensorEspresso]

@ItsmeAjayKV That's a pretty darn good result!

[TensorEspresso @TensorEspresso]

@bnjmn_marie Thank you! I'm finding IQ3_S GSQ-RCO very capable and enjoyable to use. And it flies with the Strata engine. Appreciate you sharing your work!

[TensorEspresso @TensorEspresso]

@MSripadarao @bnjmn_marie Forgot to mention, to your point on comparing to Q6, you might be right! This one might be my new default. I hope Qwen 4 keeps the same size!

[TensorEspresso @TensorEspresso]

@ItsmeAjayKV 27B has been so good, but with this new Strata engine that allows more of us to run Flash Next, I hope we get these 2 models that we can run on gaming PCs! Flash Next might be the new default for me!

[TensorEspresso @TensorEspresso]

@MSripadarao @bnjmn_marie Sorry for the confusion, I just meant that my analysis wasn’t rigorous; I didn’t do anything formal for testing. I just used it on some side projects to see how it was doing. Really love it though. A game changer if this allows us to use bigger MoEs!

[TensorEspresso @TensorEspresso]

@DarkMasamune87 That’s one of the best ways to tell that it’s a truly special game!

[TensorEspresso @TensorEspresso]

@Zona_Buckeye @hoosierstweets I do watch quite a bit, although I don’t do QB evaluation/coaching professionally. Seems to be a hard thing to do even at the top of the profession. I gave a mini assessment, what’s yours other than he’s mid?

[TensorEspresso @TensorEspresso]

@MarcelHolter @ItsmeAjayKV @coldniko Forgot to mention about selecting 8bit for the KV cache

[TensorEspresso @TensorEspresso]

@MarcelHolter @ItsmeAjayKV @coldniko I’d be happy to! What would you like to see? All I did was run the startup script you get from the Strata download, then chose the IQ3_S quant, 262k ctx, chose vision encoder, then let it do its thing. Rest is default! But happy to share whatever else you need!

[TensorEspresso @TensorEspresso]

@hoosierstweets He’s got a little gunslinger in him. You can tell he’s comfortable in air raid style offense, wish we would do it more. Needs to work a bit on placement. Love what I see though!

[TensorEspresso @TensorEspresso]

@ItsmeAjayKV @coldniko It does! Some stats for a day of Hermes agentic coding:
Session totals (684 requests, ~12h uptime)
- Aggregate decode: 134.4 tok/s over 1.6M+ output tokens
- Aggregate prefill: ~2,520 tok/s on new tokens
- KV reuse: 95.9% of 16M tokens from cache
- Draft acceptance: 82.6%

[TensorEspresso @TensorEspresso]

@RohitP_tweets @ItsmeAjayKV @coldniko I totally agree - I was skeptical that it would meet 27B, much less exceed!

[TensorEspresso @TensorEspresso]

@needmorevram How are you liking it? It’s been very good for me so far - might be my new default!

[TensorEspresso @TensorEspresso]

@solkry12 @Biggest I know my question sounds completely stupid 😆

But go with me, new model updates aren’t strictly always an upgrade - I remember some Opus updates that were received poorly. Sometimes hype can lead to disappointment, hence, my question. Cheers!