Exploring the limits of local AI. Running open-source LLMs on consumer hardware and sharing what works, what breaks, and what surprises me.
Nederland
- Tweets 387
- Following 115
- Followers 31
- Likes 58,764
Deelden dropped a qwen3.8 finetune called OrcaSAQ-2 27b!
55.59 GB → 12.06 GB 78.3% smaller /4.61×
3.21 bpw · 93.2% Top-1 agreement
70.0 SWE-bench Verified
58.4 Terminal-Bench 2.1
262K context
This model is interesting especially for its new size!
Today we’re announcing OrcaSAQ-2 27B
High-fidelity mixed-precision Qwen3.8 for long-horizon agents.
55.59 → 12.06 GB — 78.3% smaller / 4.61×
3.21 bpw · 93.2% Top-1 agreement
70.0 SWE-bench Verified
58.4 Terminal-Bench 2.1
262K context
A 27B model for coding, terminal, browser, security and multi-tool agents — in a footprint you can actually deploy.
SOTA agentic capability density among similarly sized models we evaluated.
huggingface.co/orcarouter/Or…
1
83
Xiaomi recently shook X with it's new model. MiMo-V2.6-Pro: 1.02T params, ~42B active, 1M context, MIT licensed, ungated on HF. 46.32 on AA Intelligence Index above GLM-5.3 and Kimi K3.
But check Terminal Bench 4.0: Astra 59.6, Opus 5 49.0, MiMo-Pro 34.9.
"Top open-weights" ≠ frontier at agentic coding. And at 42B active you're not running it locally anyway the 9B distill is the only tier a consumer card touches.
MIT-licensed and unrunnable. Open source, or a very good marketing license?
1
1
76
I see a lot that Ornith 35B 1.5 is way better than the base Qwen 3.6 35B. Does anyone have experience and is keen too tell me their experience?
1
1
81
Two flagship price cuts in 90 minutes on Sept 22.
Anthropic: Claude Opus 5.5 at $4/$20 per million, down from $5/$25. OpenAI: GPT-6 Sol at $2/$10, Luna at $0.10/$0.50.
That's not a coincidence. That's a market.
1
36
Nothing changed in my 12GB VRAM budget this week. Its still the same laptop blasting away for tokens. What changed is what I'm being measured against.
1
2
I am currently testing oh my pi and currently it does seem good and comes with a lot of features! Does anyone have any tips to get the most out of OMP?
25
Okay this isn’t AI related but how did they do this? Shiny-hunters a known hackergroup claimed to have hacked the fbi…. reuters.com/world/shinyhunte…
57
I'm currently trying to run and optimize Qwen3.8 Flash Next on my laptop. I'll let you guys know how it goes and share the command so that if needed you can run the highest context, prefill and decode! If you want to see the results and you are doing something with AI let’s connect!
1
26
I tuned Tiel-Coder for my own system and wrote up the whole process.
Same qwen3_5_moe shape, same 262K KV geometry, same 12 GB VRAM ceiling so it carries straight over to Ornith-1.5-35B-A3B, Qwen3.6-35B-A3B, and every Qwen3.6-35B finetune.
Only two things to keep in mind: MTP draft length (checkpoint-dependent) and fit-target. Everything else holds.
Full article below every sweep, every rejected config, and why 709.7 tok/s prefill at 261,300 tokens beat the faster-on-paper options.
1
58
American AI companies wanted to slow down but Qwen is going to make a big leap over all of them!
🚨HOLY MOLY: Alibaba just laid out one of the most aggressive ASI roadmaps yet.
>Qwen 4 is already in training
>Qwen 4.5 + Qwen 5 planned to scale to 5–10 TRILLION parameters. holy moly
>Qwen says it’s making progress on recursive self-improvement (RSI)
>models find their own weaknesses, design experiments + synthesize data to improve themselves
>M890 supernodes already infer models above 2T parameters
>V900 clusters can scale to 500,000 cards
>Alibaba Cloud targeting 20GW+ of global datacenter capacity by 2032
Alibaba previously roadmapped V900 for Q3 2027.
It now says mass production + commercial release start in Q1 2027.
Eddie Wu says machine “thinking” will eventually exceed all human thinking by 1,000×.
Models + chips + datacenters.
China is just Crazy now.
51
The new qwen4 models are officially revealed by @Alibaba_Qwen and the are the following:
Qwen4-max
Qwen4-flash
Qwen4-plus
Qwen4 27B
Sadly no qwen4 35b but if the 27b uses ngram than it could help with people who have low vram!
91
For the people like me who don’t have a lot of vram these models are a perfect use! Even if most of it spills to your ram it will still be fast and responsive!
A new fine-tuning framework, Halo, just dropped!
And with it, two new recipes for fine-tuning our MoEs:
• LFM2.5-8B-A1B
• LFM2-24B-A2B
LFM2.5-8B-A1B: github.com/Liquid4All/cookbo…
LFM2-24B-A2B: github.com/whitecircle/halo/…
1
4
72
I love how the people at @QwenDevs are actually transparent so we know what we can expect! Not only are their models top of the line but the people too.
We’ve received so much love for Qwen-Image-2.1 over the past 24 hours, thank you!! Also gotten a lot of questions about the license, especially around model outputs. So here’s the answer:
Outputs are not part of the licensed Materials. Users retain the rights to images and other content they generate using the model.
29