{
  "data": {
    "author": "Vlad Cealicu",
    "faq": null,
    "kicker": "Blog · Opinion · 22 November 2026",
    "lede": "An open model can now be made much better at one job by a file a hundredth of its size, trained by somebody who isn't the lab that made it. I think that changes who improves models, and what a benchmark of one means.",
    "readingMinutes": 6,
    "words": 1337
  },
  "kind": "anchor.page",
  "links": {
    "api": "https://www.anchorterminal.com/api/v1/index.json",
    "html": "https://www.anchorterminal.com/blog/what-a-lora-is-and-what-it-changes",
    "json": "https://www.anchorterminal.com/blog/what-a-lora-is-and-what-it-changes.json",
    "llms": "https://www.anchorterminal.com/llms.txt",
    "markdown": "https://www.anchorterminal.com/blog/what-a-lora-is-and-what-it-changes.md",
    "mcp": "https://www.anchorterminal.com/mcp",
    "slim": "https://www.anchorterminal.com/blog/what-a-lora-is-and-what-it-changes.min.md"
  },
  "markdown": "By Vlad Cealicu*, 22 November 2026. 1,337 words, about 6 min to read. * With the editing help of Claude Opus or Claude Fable.\n\n- 35.3% to 51.0%: Qwen3.8-27B on the 51 hardest tasks of SWE-bench Pro, without and with the adapter, thinking off\n- 6 to 8 times: fewer tokens than the plain model needs with thinking on, for about the same accuracy\n- 10,000 times: fewer parameters to train than full fine-tuning, for GPT-3 in the paper that introduced LoRA\n- 10s of megabytes: one adapter for Apple's on-device model of about 3 billion parameters\n\nharness-hub's card for its OpenCode adapter [1], Hu et al. [3] and Apple [6], read 22 November 2026.\n\nIn November 2026 an organisation on Hugging Face called harness-hub published a file that makes Qwen3.8-27B better at working inside the OpenCode coding agent [[1]](#ref-1). The model's own weights don't change. On the 51 hardest tasks of SWE-bench Pro it went from 35.3% to 51.0% with thinking off, and on Terminal-Bench 2.1 from 46.4% to 52.4% [[1]](#ref-1). The file is 0.6 GB as published and 0.3 GB in the build llama.cpp loads [[1]](#ref-1), where the Q8_0 build of the model our lab runs is 28.6 GB [[9]](#ref-9).\n\nThat file is a LoRA adapter. The idea is from 2021 and it's been how people change image models for years, so none of this is new. What I think is new is that it has reached agents, and that the people publishing it measured the thing I care about, which is how much work a model does to finish a job.\n\n## What a LoRA is\n\nA model is a very large set of numbers, and fine-tuning used to mean changing all of them, which gives you a second copy of the whole model for every job you train it for. Hu and colleagues at Microsoft proposed in June 2021 to freeze the model and train a small pair of matrices beside each layer you want to change [[3]](#ref-3). The pair is thin (its rank is how thin), so there's very little to train and very little to store. For GPT-3 they report 10,000 times fewer parameters to train than full fine-tuning and a third of the GPU memory, with results on par or better [[3]](#ref-3).\n\nWhat comes out is a file that holds only those matrices. You load it on top of the model, or merge it in, and once it's merged the model runs no slower than it did [[3]](#ref-3). The harness-hub adapter is rank 32 on the attention and feed-forward layers [[1]](#ref-1). It only works with the model it was trained on, and its card says so [[1]](#ref-1).\n\n## Why video and images got there first\n\nAn image or video model is expensive to retrain and slow to run as it comes, and a LoRA helps with both. People train one on a few pictures or clips to get a style or a character, which is the use everybody knows. The use I find more interesting is speed. In November 2023 Luo and colleagues showed that the difference between a slow Stable Diffusion model and a fast distilled one could itself be stored as a LoRA and plugged into other fine-tuned models with no further training [[4]](#ref-4). Video works the same way now. The lightx2v adapters for Wan2.2 are rank 64 files that bring generation down to four steps, and they have to be used with the Wan2.2 base model [[5]](#ref-5).\n\nSo in video an adapter is how you get what you want and how you get it quickly, and the base model is the thing everybody shares.\n\n## What changed for code\n\nThe harness-hub adapters are trained for a harness, which is the coding agent a model works inside. Hugging Face's post on them, published on 20 November 2026, says they trained one adapter for Qwen3.8-27B for each of three agents (pi, Hermes Agent and OpenCode), with reinforcement learning inside the agent itself [[2]](#ref-2).\n\nThe accuracy went up, and I'd hold that loosely. The sets are small and the numbers are the publisher's own. The gain on SWE-bench Pro is 15.7 points give or take 13.9, and the gain on Terminal-Bench 2.1 is 6.0 give or take 10.7 [[1]](#ref-1).\n\nThe number I'd hold on to is the work. On the SWE-bench Pro tasks both solve, the adapted model takes 13 turns where the plain one takes 48 [[1]](#ref-1). Against the plain model with thinking on, it reaches about the same accuracy on 6 to 8 times fewer tokens, a median of 3,900 a task against 31,800 on SWE-bench Pro [[1]](#ref-1). Thinking is a model working out, every time, what it could have been taught once. On an API that's a bill. On your own machine it's time, and the best model on our lab box already takes 71 seconds a question [[10]](#ref-10).\n\n## What it changes\n\nA lab isn't the only one who can make an open model better any more. A new checkpoint of a 27B model takes the lab that made it. These adapters took a team with tasks to train on and two GPUs each [[2]](#ref-2), and they ship as a file a hundredth of the size. I think most of the improvement in open models in 2027 will arrive this way, between releases.\n\nThe harness is now part of the model. These adapters were trained inside one agent and measured inside it, and the card says its numbers can't be compared with the ones on Qwen's own card, which used other harnesses [[1]](#ref-1). So \"how good is Qwen3.8-27B at coding\" has no single answer. It has one for each agent it runs in and each adapter it's wearing, and a table that doesn't say which is a table I'd skip.\n\nOne base model can do many jobs on a small machine. Apple described this for its phones in June 2024. It ships one on-device model of about 3 billion parameters and an adapter for each feature, each adapter in the tens of megabytes, loaded and swapped as needed [[6]](#ref-6). llama.cpp's server can load several adapters and switch between them while it runs [[7]](#ref-7), so the same thing works on one machine at home.\n\nIt reaches robots too. OpenVLA, a 7 billion parameter model that turns what a robot sees and is told into what it does, can be fine-tuned on consumer GPUs with low-rank adaptation [[8]](#ref-8). I'd expect the same in any field where one general model exists and each job is narrow, and where the people who know the job aren't the people who can train a model from nothing.\n\nThe same door opens for people you'd rather it didn't. In October 2023 Lermen and colleagues undid the safety training of Llama 2-Chat with a LoRA for under $200 on one GPU, and got the 70B model's refusals down to about 1% [[11]](#ref-11). Safety training in open weights is a default, and a cheap one to change.\n\nAnd the base model becomes hard to leave. An adapter is useless without the exact model it was trained on [[1]](#ref-1) [[5]](#ref-5), so every adapter published for a model is a reason to keep using it after a better one comes out. I think the open models that last will be the ones with the most adapters, more than the ones with the best scores on the day.\n\n## What we're doing with it\n\nOur lab's harness isn't OpenCode. Its tools, its prompts and the Go programs a model writes in it are ours [[9]](#ref-9). So the OpenCode adapter is going on the lab's list beside the plain model at the same settings, to see whether an adapter trained for one harness helps, does nothing or hurts in another [[12]](#ref-12). Neither has run yet.\n\n## What isn't known\n\nWhether any of this holds outside the publisher's own tests. Whether it survives the smaller builds people run at home, since the card says the llama.cpp build of the adapter hasn't been evaluated again [[1]](#ref-1). And whether adapters stack. Hugging Face found its three were almost unrelated to each other and merged them into one, and the plain sum made the agents too brief to do the work, with 13 tool calls a task against the plain model's 49 [[2]](#ref-2). They had to turn it down by half to get a file that works in all three [[2]](#ref-2).\n\n---\n\nMore in [Opinion](https://www.anchorterminal.com/blog/opinion/index.md) · Older opinion [A machine that proved the Riemann hypothesis, or at least we think it did](https://www.anchorterminal.com/blog/a-machine-that-proves-riemann.md)\n\n- Older: [An open model on a £3,879 box scored 98 on our test. Claude scored 100](https://www.anchorterminal.com/blog/local-models-on-one-box-against-claude.md)\n- All posts: https://www.anchorterminal.com/blog/index.md\n\nAnchor Terminal grades the tools AI agents run on, with a source for every score. [The top tools](https://www.anchorterminal.com/top/index.md) · [Categories](https://www.anchorterminal.com/categories/index.md) · [Compare two tools](https://www.anchorterminal.com/compare/index.md) · [Prices](https://www.anchorterminal.com/prices/index.md) · [Sunsets](https://www.anchorterminal.com/sunsets/index.md) · [API, MCP server and CLI](https://www.anchorterminal.com/agents/index.md)\n\n## References\n\nRead on 22 November 2026.\n\n[1] harness-hub, Qwen3.8-27B-opencode-adapted, model card on Hugging Face, November 2026. Source for the adapter's results, their intervals, the turns and tokens a task, its rank, its sizes, the size of its file (and of the GGUF build in harness-hub/Qwen3.8-27B-opencode-adapted-GGUF), that it needs its base model, that its numbers aren't comparable with Qwen's card and that its GGUF build hasn't been evaluated again. https://huggingface.co/harness-hub/Qwen3.8-27B-opencode-adapted\n\n[2] Hugging Face, Introducing harness adapters, 20 November 2026. Source for one adapter for each harness, trained inside it on two GPUs, the three harnesses, and what happened when the three adapters were merged. https://huggingface.co/blog/harness-adapters-at-scale\n\n[3] Hu and others, LoRA: Low-Rank Adaptation of Large Language Models, arXiv 2106.09685, 17 June 2021. Source for what a LoRA is, 10,000 times fewer trainable parameters, a third of the GPU memory and no added latency. https://arxiv.org/abs/2106.09685\n\n[4] Luo and others, LCM-LoRA: A Universal Stable-Diffusion Acceleration Module, arXiv 2311.05556, 9 November 2023. Source for a speed-up stored as a LoRA and plugged into other fine-tuned models without training. https://arxiv.org/abs/2311.05556\n\n[5] lightx2v, Wan2.2 Distilled LoRA Models, model card on Hugging Face. Source for four-step video generation from a rank 64 LoRA that has to be used with the Wan2.2 base model. https://huggingface.co/lightx2v/Wan2.2-Distill-Loras\n\n[6] Apple Machine Learning Research, Introducing Apple's On-Device and Server Foundation Models, 10 June 2024. Source for one on-device model of about 3 billion parameters with an adapter for each feature, tens of megabytes each, loaded and swapped. https://machinelearning.apple.com/research/introducing-apple-foundation-models\n\n[7] ggml-org, llama.cpp, the server's README. Source for loading several LoRA adapters and changing them while the server runs. https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md\n\n[8] Kim and others, OpenVLA: An Open-Source Vision-Language-Action Model, arXiv 2406.09246, 13 June 2024. Source for a 7 billion parameter robot model fine-tuned on consumer GPUs with low-rank adaptation. https://arxiv.org/abs/2406.09246\n\n[9] Anchor Terminal, anchorlab, MODELS.md and README.md. Source for the 28.6 GB build of Qwen3.8-27B the lab runs and for the lab's own harness. https://github.com/Anchor-Terminal/anchorlab/blob/main/MODELS.md\n\n[10] Anchor Terminal, Lab score, open-weight models on our own evaluation. Source for the best model on the lab box taking 71 seconds a question. https://www.anchorterminal.com/benchmark/lab/evaluation/models/\n\n[11] Lermen and others, LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B, arXiv 2310.20624, 31 October 2023. Source for under $200, one GPU and refusals of about 1%. https://arxiv.org/abs/2310.20624\n\n[12] Anchor Terminal, anchorlab, pull request 36. Source for the adapter and the plain model at its settings going on the lab's list. https://github.com/Anchor-Terminal/anchorlab/pull/36\n",
  "meta": {
    "attribution": "Anchor Terminal (https://www.anchorterminal.com)",
    "attributionRequired": true,
    "credit": "CC BY 4.0, credit Anchor Terminal and link the page or API address you took it from (https://www.anchorterminal.com/terms/#data)",
    "docs": "https://www.anchorterminal.com/docs/",
    "generatedAt": "2026-11-22",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "1.0",
    "openapi": "https://www.anchorterminal.com/openapi.json",
    "preview": false,
    "run": "2026-10-01",
    "runLabel": "October 2026 research run",
    "terms": "https://www.anchorterminal.com/terms/#data"
  },
  "page": {
    "breadcrumbs": [
      {
        "name": "Home",
        "url": "https://www.anchorterminal.com/"
      },
      {
        "name": "Blog",
        "url": "https://www.anchorterminal.com/blog/"
      },
      {
        "name": "Opinion",
        "url": "https://www.anchorterminal.com/blog/opinion/"
      },
      {
        "name": "What a LoRA is",
        "url": ""
      }
    ],
    "description": "Vlad Cealicu on LoRA adapters. In November 2026 a 0.6 GB adapter took Qwen3.8-27B from 35.3% to 51.0% on the hardest tasks of SWE-bench Pro without changing the model. The same idea is how video models get faster, how a phone runs many features on one model and how a robot learns a new job, and in 2023 it was how the safety training of an open model was undone for under $200.",
    "facts": [
      "Vlad Cealicu",
      "22 November 2026"
    ],
    "h1": "A small file took an open model from 35% to 51% on hard coding tasks. What a LoRA is, and what it changes outside code",
    "image": "https://www.anchorterminal.com/assets/og/blog-what-a-lora-is-and-what-it-changes.png",
    "path": "/blog/what-a-lora-is-and-what-it-changes",
    "published": "2026-11-22",
    "section": "blog",
    "title": "What a LoRA adapter is, and why it matters for coding, video and beyond | Anchor Terminal",
    "toc": null,
    "updated": "2026-11-22",
    "url": "https://www.anchorterminal.com/blog/what-a-lora-is-and-what-it-changes"
  },
  "tokens": {
    "markdown": 3100,
    "slim": 1280
  },
  "version": 1
}
