I just released a new feature for Hyperfolio and the Hyperfolio API: Yield. You can now browse thousands of yield opportunities on HyperEVM across nearly 30 protocols

1st of April: Qwen release Qwen 3.6 Plus 14th of April: Qwen release Qwen 3.6 35B A3B 1st of June: Qwen release Qwen 3.7 Plus 14th of June: Qwen release Qwen 3.7 35B A3B (?) I believe

A bit worried about the upcoming Qwen 3.7 open source models A Qwen team member recently deleted a comment where he said they would likely release another 27B model The Summary section of the 3.7 Plus blog post doesn’t mention any upcoming open-source models, whereas the 3.6 Plus blog explicitly said they would be open-sourcing smaller-scale models We also didn’t get the other two Qwen 3.6 models from the poll, 9B and 122B I still think we’ll probably get some open-source models from the 3.7 series, but it’s unclear which sizes they’ll be or when they’ll arrive

oh boy, I'm running Qwen 3.6 27B Q6_K using tensor split on both my RTX 3090 and 3080 with almost full context (250k). yesterday, my friend sent me this reddit post: reddit.com/r/LocalLLaMA/comm… turns out llama.cpp build b9455 allows for good tensor splitting across two GPUs good news is: the 3090 and the 3080 have the same bandwidth (936 GB/s), so even though the 3080 has less VRAM, it won't drag down the 3090 this means I can split the model 70/30 between the two cards based on my first tests, the MTP version gives me 69 to 44 tok/s in decode depending on context (it'll probably be less during a real coding session, I'll test it out later) ngl, I'm pretty happy to be able to run bigger quants than Q4 I also tried the Q6_K_XL, which works a bit slower and has a tighter context window Config: Q6_K, KV q8_0/q4_0, tensor split 70/30, ctx 252000, MTP, 3090 3080 Ti at 300W Model: huggingface.co/unsloth/Qwen3…

I've ran some tests on Gemma 4 12b on my RTX 3060 12gb it seems to handle context pretty good and I have around 30tok/s almost all the way to 130k context I asked GPT to test it out a bit for coding and comparing to Qwen 3.6 35B, and it seems quite close I have tried the Q4_K_M from Unsloth, I want to see if the Q6 does any better quality wise

We're building the Docker for Local AI. A simple way to package and share complete AI environments including models, runtimes, dependencies, and configuration. One command. Same setup. Run anywhere. Free and Open-source we’re live on Product Hunt today.

I really don't understand the point of not having like the full benchmark tested and rely on "Overall Performance" or "E-Commerce" aggregated data model looks super cool, but idk I just feel rugged when labs do such things also why everyone keeps having 3.5 on their benchmark even tho 3.6 exists

next up: 35B A3B then 27B right?

27B running on my 3090 that delegate 2 code reviews to two 35B subagents running on my 3060 and 3080 pi is such a perfect harness for local AI is this AGI?

I've tried tvall43's 14B REAP of Qwen's 35B A3B over the last few days on my 3060 and I have to say I'm quite surprised Q4_K_M is 8.47 GB so this fits nice on the 12gb VRAM we have room for 64k context and --parallel 2 for sub agentic work until now I felt that all the REAP I've tried where brain dead, or too broken to be useful this 14B actually does stuff tho I've benched it against 35B on 270-cases from different benchmarks (was mainly focused on coding and tool-use) 14B vs 35B: - BFCL subset (120 items): 38.3% vs 72.5% - LiveCodeBench-lite subset (80 items): 41.3% vs 52.5% - Qodo review subset (40 items): 90.0% vs 92.5% - small coding set (30 items): 33.3% vs 46.7% overall on that 270-case mixed run, it retained ~69% of 35B's pass rate, while being faster and fitting on 12gb without any offload I need to test it more in real coding with 27B as main agent and 14B as sub agent

fitted 3 GPUs in my homelab, had to cut the case a bit though now let’s test how the case temperatures hold up under pressure

We’re aware of the security reports linked to rewards payout. User funds and market resolution are safe. Findings point to a private key compromise of a wallet used for internal top-up operations, not contracts or core infrastructure. More updates to follow.

guess who just happens to have a 3080 ti 12gb in the mail instead of the 3090? 🤡

while I really love the approach and the compute put to the work, I feel that those distill don't really bring much... I've tested out this coder version vs the 9B Base: 9B Base: - HumanEval full: 150/164 (91.5%) - MBPP sanitized 100: 87/100 (87.0%) 9B Coder: - HumanEval full: 124/164 (75.6%) - MBPP sanitized 100: 80/100 (80.0%) maybe those are not the best benchark to run against tho