How do labs train a frontier, multi-billion parameter model? We look towards seven open-weight frontier models: Hugging Face’s SmolLM3, Prime Intellect’s Intellect 3, Nous Research’s Hermes 4, OpenAI’s gpt-oss-120b, Moonshot’s Kimi K2, DeepSeek’s DeepSeek-R1, and Arcee’s Trinity series. This blog is an attempt at distilling the techniques, motivations, and considerations used to train their models with an emphasis on training methodology over infrastructure.
These notes are largely structured based on Hugging Face’s SmolLM3 report due to its extensiveness, and it is currently supplemented with notes from other reports including Intellect-3, gpt-oss-120b, Hermes 4, DeepSeek, and Kimi. While this blog explores some infrastructure-related ideas like in-flight weight updates and multi-client orchestrators, there are many other ideas mentioned throughout those posts/blogs like expert parallelism and quantization. Hugging Face writes more about gpt-oss-120b’s infrastructure here.
Author here. We built 5 high-fidelity mock Google Workspace + Slack services and ran 7,224 trials across 6 frontier models and 4 agent harnesses.
The headline finding that surprised us most: scaffolding (skills + meta prompt) gives a 39-63pp lift, while the top 5 models are statistically indistinguishable (53-63% TSR, no pairwise comparison survives correction). Your choice of scaffolding matters ~6x more than your choice of model.
The safety findings are darker: Opus leads on task success (63%) but ties for most unsafe (23% UAR). GPT-5.4 is the safest (7% UAR) but mid-tier on tasks. There's no capability-safety tradeoff — they're decoupled.
Also I'm reviewer of Terminal Bench 3.0. Here's what I've heard from contributors as well.
> I noticed that when I was building tasks with harbor. Claude is a good student which generally follows the instruction. But gpt always try to find a short path to cheat. Like reversing the binary directly instead of interaction
Another friends added ways to address this: https://x.com/xeophon/status/2041772210562511080?s=20
> Just ask codex to not reward hack
> It literally works. And it works even better when you state which things you consider reward hacking, eg wrapping a CLI or something
A two-week study of autonomous language model agents deployed in a live multi-party environment with persistent memory, email, shell access, and real human interaction — tested by twenty researchers interacting both benignly and adversarially.
Agent scaffold comparison. We additionally evaluateOpenCode, an open-source scaffold that supports multiplemodel providers. Native CLI scaffolds consistently outper-form OpenCode when using the same underlying model.GPT-5.1 Codex Max achieves 20.2% on Codex CLI butonly 7.7% on OpenCode. Similarly, Gemini 3 Pro scores18.3% on Gemini CLI versus 14.9% on OpenCode. The
one exception is Claude Opus 4.5, which scores 17.1% on Claude Code and 17.3% on OpenCode — effectively equivalent, and the only case where the open-source scaffold
matches or slightly exceeds the native one.
no worries it's totally fine! there is indeed work needs to be done on the feedbacks generated skills. Thanks for helping us submitting on HackerNews. And for
> a lot of Skills on GitHub are just AI-generated without any feedback or deliberative refinement. Many thought those would still be valuable, but you've shown evidence otherwise.
we do find most skills on the internet to be useless, and thanks to the generosity of https://skillsmp.com/ author, we were able to get all the meta data of the 99k skills indexed on his website. We did a lot of filtering and deduping and we discovered ~40k+ skills were relevant at the time we did the study.
20+ Anthropic Default Skills, 200k+ community skills on skillsmp. People talk about skills without knowing how well they work.
We're hosting the largest Agent Skills hackathon at Founders, Inc. (March 7 - 8) from our lessons learned at SkillsBench
No sims. No slides. No flops.
Did you check our repos and sites? the repo is skills native. Also please don't be misled by the original title, we have this configuration to eliminate the impact of internal knowledge of LLMs. It's in the paper.
We collected 86 tasks from 105 domain experts across 11 domains, every task is verifiable, human created and has verified Skills. SOTA model without skills score ~30% without skills.
We found a few interesting things:
1. Skills substitute for model scale — Haiku 4.5 with Skills (27.7%) beats Opus 4.5 without (22.0%).
The right procedural knowledge can be worth more than a bigger model.
2. Skills' improvement has nothing to do with LLMs' internal knowledge. We have an ablation where no Skills provided for the agent, but the agent is prompted to generate relevant procedural knowledge before solving the task. This isolates the impact of LLMs' latent domain knowledge. The result is:
Curated Skills: +16.2pp average improvement across all 7 agent configs
Self-generated Skills: -1.3pp: models can't write their own procedural knowledge pre-trajectory feedbacks. This is used to isolate the impact of LLMs' latent domain knowledge.
Thanks @dang for moderating! This is indeed not our original findings and this is a sub conclusion for an ablation we did to remove the confound of LLMs internal domain knowledge. Thanks for submitting for us @mustaphah here's a little bit more details on how we approach this:
> I would frame the 'post-trajectory generated skills' as feedback-generated skills, so is Letta: https://www.letta.com/blog/skill-learning. We haven't seen existing research or hypothesis debating whether the skills improvement might come from the skill prompt themselves activated knowledge in LLMs that can help itself. So that's why we added an ablation of 'pre-trajectory generated skills' because we have that hypothesis and this seems a very clean way to test it. Also it is very logical that feedback generated skills can help, because it most certainly contain the failure mode of agents on that specific tasks.
I would frame the 'post-trajectory generated skills' as feedback-generated skills, so is Letta: https://www.letta.com/blog/skill-learning. We haven't seen existing research or hypothesis debating whether the skills improvement might come from the skill prompt themselves activated knowledge in LLMs that can help itself. So that's why we added an ablation of 'pre-trajectory generated skills' because we have that hypothesis and this seems a very clean way to test it. Also it is very logical that feedback generated skills can help, because it most certainly contain the failure mode of agents on that specific tasks.
> limited to a single markdown file of instructions
single file of instructions is common in most benchmark papers, e.g. Terminal Bench. Also we have very complicated prompts like this one: https://www.skillsbench.ai/tasks/shock-analysis-supply
> opaque verifier
Could you specify which tasks' verifier is not clear or defective for benchmarking purpose?
It says:
"Our top priority is ensuring that this change won't be disruptive for our customers. We will continue to sell and operate our product subscription service through our app and website. The company will continue to operate from Singapore."
But I suppose they won't try as hard as before to make the product better. It's such a shame. I've been using it since it launched the video by begging everyone I knew and got an invite code. And I've been on the higher end of subscription ever since.
tldr:
- gpt-5.2 and gpt-5.1-codex-max have identical pass rates but solve different tasks
- 36 tasks common to both
- 12 tasks unique to each model
- gpt-5.2-pro consistently underperforms by ~7-9 percentage points
- gpt-5.2-pro has significantly more timeout issues (26 vs 7-8)
- Extended timeouts recover additional passes - using 3x timeout multiplier recovers ~5-7 passes per model
My co-founder and I were building another vertical agents, and we were having a hard time to both get a phone number and set up tool calling (both retrieving and sending) text update for our users.
These notes are largely structured based on Hugging Face’s SmolLM3 report due to its extensiveness, and it is currently supplemented with notes from other reports including Intellect-3, gpt-oss-120b, Hermes 4, DeepSeek, and Kimi. While this blog explores some infrastructure-related ideas like in-flight weight updates and multi-client orchestrators, there are many other ideas mentioned throughout those posts/blogs like expert parallelism and quantization. Hugging Face writes more about gpt-oss-120b’s infrastructure here.