2026-08-09

I compared Qwen 35B-A3B MoE against Qwen 27B dense on a series of local coding-maintenance tasks. On my R9700/llama.cpp setup, the MoE model generated about 3.9× faster (~116 vs ~30 tok/s) , but the coding-quality difference was much smaller than I expected. Both usually handled ordinary bug fixes and multi-file changes correctly. As I made the tests progressively harder, the dense model did show an advantage—but mainly in implicit invariants, unusual edge cases, and consequences beyond the lite

Someone explained this to me in a comment thread and it's been rattling around in my head since. The idea: in a long conversation, if the model says something wrong and you correct it, that correction doesn't necessarily erase the wrong idea's influence. The tokens around the mistake, including the back-and-forth about why it's wrong, can end up giving the original bad idea more weight in context, not less, because it's now been referenced multiple times. The model starts treating the repeated-b

On b10173 - "state":"loading" 4min54sec. - With this PR and GGML_RPC_LOAD_THREADS 12 - "state":"loading" 1min38sec Interestingly the biggest bottleneck wasnt networking, disk IO, or any of that pci gen2/3/4... It was 1 CPU thread doing all the work while the others sat idle during the model load. This handles _part_ of the problem, but there is still room for more noted in comments. The PR is close to ready, will need a docs change if they want to keep the new GGML_RPC_LOAD_THREADS variable.. an

I maintain a self-hosted project and received an email yesterday for a sponsorship like many other companies have done. However, this one was different: They started off with a simple and flattering email of what my project does (probably AI description) and that they have shared it with their team and offered to sponsor me like many other companies have done in the past. They say that the sponsorship will come through "Pump.fun's GitHub Sponsorship integration" however that integration does not

Planka is a snappy Web Kanban with a decent API. The company offers some additional functionality with their paid "Pro" plans. They have now announced to move SSO/OIDC out of the Community edition into Pro . What do you think about that? What alternatives exist? Anyone interested in forking?

Running across 2 clusters using llama.cpp over RPC too. Both clusters are not enough to hold everything in memory, so main cluster still partially offloads to run. Goal will be to get all the GPUs in one system and without RPC, I should probably see 2-3x faster speed. Running the IQ1_M, goal is to get to Q2_K_XL. My hope is that Qwen3.8 is as good, faster and smaller, and that DeepSeekV4Pro/GLM5.3 will all be the same size and just as good. I'm going to give this a hard coding problem to see the

To power its new West Texas data center, Amazon is investing in the construction of a new power plant that could be one of the largest single producers of greenhouse gases in the US, according to the New York Times. The new gas-burning plant in Pecos County, Texas has received significant investment from Amazon and, […]

Vision-Language-Action (VLA) models have become a powerful paradigm for robot manipulation, but training a single generalist policy for heterogeneous robot embodiments remains an open problem. Existing methods have two main limitations. First, they underuse dynamics priors shared across diverse visual and interaction data, limiting cross-embodiment transfer. Second, they require extensive manual preprocessing to convert embodiment-specific actions into a common format. To overcome these limitati