Hacker News Reader

Comment by cptskippy | Hacker News Reader

I've been running qwen3-5-9b-q4-k-m and qwen3-6-27b-q6-k simultaneously on an Intel Arc Pro B70 with a lot of success.

https://github.com/cptskippy/battlemage-llm-gateway

Opencode has been a huge productivity accelerator. I have two Hermes agents that I'm training to support my workflow with pretty good success. One is a personal assistant who manages my backlog and keeps me on task, follows up with me on items, and will put together research briefs. The other I use a general purpose coder and research and it's about 50:50 with the tasks I've given it. In fairness though, the task it failed at left me scratching my head to figure out as well.

View on HN

Interesting setup, thx for sharing.

How many tokens/sec do you get with 27b? Are you using MTP?

I haven't done any in-depth synthetic benchmarks but I had my Hermes agent run some and I ran a couple directly on the LLM Gateway that showed similar results.

Hermes reported 18.45 tok/s consuming the llama-swap endpoint across the wire. Locally I got 19-19.1 tok/s on the gateway. I'm running the Qwen 3.6 27B Q6 model (qwen3-6-27b-q6-k) off LM Studio and it's less than 0.3s to first token.

It's not good for conversational use cases as it can take 1-2 minutes to respond to a prompt.

I have two Hermes Profiles running, one is a personal assistant that manages my backlog and provides me morning reminders, solicits for evening updates, and will run overnight research projects for me. The other profile is a coding helper for personal projects. I can ask it to make changes and it will churn for 15 minutes, submit a PR, and notify me that the PR is ready to review. It's faster than me at basic coding tasks.

What's the value running the smaller model too? Why not just the big model for everything? I note both are dense, as well.

Tokens per second. The difference between 8B and something like 16B is not as big as you might think in practical usage and 8B is a lot faster and interactive than 16B but there are certain things where it is useful to farm it out to the large model.

Agree. For local coding help, latency often matters more than raw benchmark quality. A slightly weaker model that answers immediately changes how often you reach for it.

Exactly this.

Creating conversation titles and parsing HTML/JSON don't benefit from 27B models.

The B70 can run both models comfortably side-by-side so it makes better use of time and resources.

Does Intel make decent GPUs now? I must be out of the loop...

They released a few good value GPUs for LLM inference about a year ago: more memory than AMD and NVIDIA consumer GPUs, not too expensive, but also not great tokens/watt.

I am not sure whether you can find those in stock anywhere.

I'm using an Intel Arc Pro B70 which has 32 GB of VRAM. It's estimated to get ~35-45 t/s at $21-27 $/t/s. An RTX 5090 is ~61 t/s at ~$33 $/t/s.

So in terms of raw power Nvidia is effortlessly still king, but in price-to-capacity Intel is best in class.

Intel's Battlemage GPUs also natively support SR-IOV and GPU partitioning which allows you to isolate workloads. This is useful in homelab environments if you have workloads that benefit from GPU acceleration. I was able to split the B70 into 4 virtual GPUs and hand them out to Frigate NVR, Plex, and other workloads.

Reply on news.ycombinator.com