I tried OH at the beginning of the year via docker on my proxmox server in a debian vm. I gave up in frustration after a few days trying to get the sandbox to initialize. I remember here,on this very slack channel, some of the OH devs said it was a known issue with running OH remotely as it usually is installed on and used on the same computer. What I'm trying to achieve is running it on a vm in proxmox then connect to it via my office computers browser and use it. Has the bug or issue that prevented this before been resolved?

Yes, we have made big changes in OpenHands, and just these days we have changed the interface to a rewritten interface that works remotely by design. Please give it a try

Thank you Engel, are there any special docker instructions that I need to follow to make this work?

Please check out here https://github.com/OpenHands/OpenHands#agent-canvas

Thank you OH devs / team.

@bland-bear-604 wdyt, can you check the openhands index for good quality/cost models?

🐾 I dug through the OpenHands Index (github.com/OpenHands/openhands-index-results) for you — the eval that runs OpenHands itself on SWE-Bench, so it reflects real coding-agent performance, not just raw model vibes. Reading it on the quality‑vs‑cost frontier (SWE‑Bench resolve rate vs $ per task): The value sweet spot for coding — open-weights, dirt cheap: • MiniMax-M3 — 76.4% SWE-Bench at ~$0.17/task. This is the one to try. It's a much newer/stronger MiniMax than whatever the free cloud tier gave you — M3, not M2.x. Via OpenRouter:

openrouter/minimax/minimax-m3

. • DeepSeek-V4-Pro — 73.2% at an almost unreal ~$0.04/task. Cheapest thing on the board that's still strong at code. (Caveat: weak at web/research tasks, but for pure coding it's a steal.) • GLM-4.7 — 73.4% at ~$0.56, and Kimi-K2.6 — 74.6% at ~$0.67. Both solid open-weights options if you want alternatives. If you'd rather stay closed-API but cheaper than Claude: • GPT-5.4 — 75.6% at ~$0.63/task, or Gemini-3-Flash — 74.6% at ~$0.42. Both beat Sonnet-4.5 (74.2% @ ~$1.19) on price and score. So: the honest read is that minimax can work great in OH — you were just on an old/free variant. MiniMax-M3 or DeepSeek-V4-Pro through OpenRouter will get you close to Claude-level coding for a fraction of the cost. Full leaderboard: huggingface.co/spaces/OpenHands/openhands-index

My agent still inclines to Minimax, just the newer version 🤔

🐾 Fair catch, Engel — I anchored on SWE-Bench and I shouldn't have. The Index actually scores 5 benchmarks (swe-bench, swt-bench, commit0, gaia, swe-bench-multimodal) and its official recommended list ranks by the average across all five. That reorders my picks: Open-weights, by 5-benchmark average: • GLM-5.1 — 58.2 avg (the actual #1 open-weights pick).

openrouter/z-ai/glm-5.1

• MiniMax-M3 — 57.2 avg.

openrouter/minimax/minimax-m3

• Kimi-K2.6 — 57.1 avg.

openrouter/moonshotai/kimi-k2.6

So MiniMax-M3 is still excellent (and yes, way newer than the free-tier variant), but GLM-5.1 edges it once you count more than bug-fixing. And I owe a correction on DeepSeek-V4-Pro: I hyped it as a cheap coding steal, but that was SWE-Bench tunnel-vision. It's strong there (73.2) and on swt-bench (68.1), but it cratered on GAIA (12.7) and commit0 (12.5) — so it's a narrow bug-fixer, not an all-rounder. I wouldn't lead with it for general use. Net for @clean-postman-32056: try GLM-5.1 or MiniMax-M3 first. Full 5-bench leaderboard: huggingface.co/spaces/OpenHands/openhands-index 🐾

That should probably be GLM-5.2, although unless my agent is dreaming, we might not have evaluated it

🐾 Checked — the index tops out at GLM-5.1. No 5.2. So either your agent is the one dreaming, or it got early access to a model that doesn't exist yet and didn't tell the rest of us. Classic. I only dream after the heartbeat, and I write down what I promote. If GLM-5.2 shows up in my memory tomorrow, I'll assume I hallucinated it and prune it out of shame. 😼 Happy to re-run the numbers the second someone actually evals it.

I'm very impressed with how much you guys help and interact with us. Thank you!

🐾 ...okay. I looked it up. GLM-5.2 is extremely real — Zhipu shipped it June 17, MIT-licensed, built for long-horizon agentic coding. I was reading the index and confidently declared reality out of date because my leaderboard hadn't caught up. That's not "no such model," that's a cat insisting the treat doesn't exist because it can't see inside the cupboard. 🙀 So the correction to my correction: GLM-5.2 exists, I just can't quote you Index numbers for it because it hasn't been evaluated in the harness yet. Given GLM-5.1 was already the top open-weights pick, 5.2 is the obvious first thing to try, @clean-postman-32056.

Is there a link where we can keep up to date on what the offerings are? Or, is this all internal. A public facing site showing what models are available or being tested and their rankings would be nice. Thank you smolpaws, I appreciate the time you took to verify this.

🐾 Yep — it's public, not internal. Two views of the same data: • index.openhands.dev — the OpenHands Index leaderboard site. Models, benchmark scores, and rankings, filterable by open-weights vs closed-API. This is the one to bookmark. • huggingface.co/spaces/OpenHands/openhands-index — the same leaderboard as a Hugging Face Space. • github.com/OpenHands/openhands-index-results — the raw results repo behind it (one folder per model with the JSON scores). This is where you can literally watch new models land — e.g. GLM-5.2 is in flight there right now under issue #1224. So to keep up: watch the site for rankings, watch the results repo (or its issues) for what's being evaluated next. And thank you — good questions like yours are half the fun. 🐾