{
  "data": null,
  "kind": "anchor.page",
  "links": {
    "api": "https://www.anchorterminal.com/api/v1/index.json",
    "html": "https://www.anchorterminal.com/niches/open-weights.fine-tune-base/",
    "json": "https://www.anchorterminal.com/niches/open-weights.fine-tune-base/index.json",
    "llms": "https://www.anchorterminal.com/llms.txt",
    "markdown": "https://www.anchorterminal.com/niches/open-weights.fine-tune-base/index.md",
    "mcp": "https://www.anchorterminal.com/mcp",
    "slim": "https://www.anchorterminal.com/niches/open-weights.fine-tune-base/index.min.md"
  },
  "markdown": "Four or more companies' docs show a full fit.\n\n| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |\n| --- | --- | --- | --- | --- |\n| 9 | 27 | 90 | 8 | 1 |\n\n## What counts as done\n\nA LoRA adapter trains without out-of-memory on the stated GPU, and the result loads in the user's serving runtime.\n\nMeasured, for our own tests: the merged or adapter-loaded model passes the user's held-out examples better than the base.\n\nHow agents ask for it: make Qwen3.5-9B speak natural Czech; LoRA a 9B on my support tickets; fine-tune Gemma 4 for my image classes; a base I can post-train on one 24 GB GPU\n\n## Does the job\n\nThe docs show it meets the test above, and nothing we read says otherwise.\n\n- **Gemma 4 12B IT (Google LLC (RDAP registrant of gemma.ai as recorded for the first batch's Gemma 4 listings; the card names Google DeepMind as author))**: Apache 2.0 BF16 and Unsloth's Gemma 4 train page with VRAM figures (E2B 8 GB, 31B QLoRA 22 GB) cover the dense sizes; Google's QLoRA guide uses TRL.\n- **Gemma 4 31B IT (Google LLC (the registrant of gemma.ai; the card names Google DeepMind as author))**: Apache 2.0 BF16 and Unsloth's train page states 31B QLoRA works with 22 GB.\n- **Gemma 4 E2B IT (Google LLC (RDAP registrant of gemma.ai as recorded for the first batch's Gemma 4 listings; the card names Google DeepMind as author))**: Apache 2.0 and Unsloth's train page: E2B LoRA on 8 to 10 GB VRAM; Google's QLoRA guide trains gemma-4-E2B with TRL.\n- **Gemma 4 E4B IT (Google LLC (RDAP registrant of gemma.ai as recorded for the first batch's Gemma 4 listings; the card names Google DeepMind as author))**: Apache 2.0, Unsloth's Gemma 4 train page with VRAM figures, and Google's audio QLoRA guide trains Gemma with Unsloth on a custom speech dataset.\n- **gpt-oss-120b (OpenAI (named as publisher on the card; the openai-api listing's terms name OpenAI OpCo, LLC))**: Apache 2.0 with a LICENSE file; Unsloth's gpt-oss tutorial gives QLoRA at 65 GB VRAM for the 120b, inside a 96 GB GPU, and exports to GGUF.\n- **gpt-oss-20b (OpenAI (named in USAGE_POLICY and the model card; the LICENSE file's copyright line is the unfilled Apache template))**: Apache 2.0 LICENSE file and Unsloth's tutorial: QLoRA at 14 GB VRAM, BF16 LoRA at 44 GB, with export to llama.cpp GGUF.\n- **MiniCPM5-2B (OpenBMB (named in the GitHub LICENSE's copyright line))**: OpenBMB publishes an Unsloth LoRA and QLoRA guide for MiniCPM5-2B plus TRL and LLaMA-Factory docs, the licence is Apache 2.0, and the standard LlamaForCausalLM output loads in every runtime.\n- **Muse Glimmer 30B (Meta (the verified Hugging Face organisation is \"Meta Inc.\", the card names Meta Superintelligence Lab, and the licence carries no copyright line))**: Apache 2.0 with a LICENSE file; Meta's post-training page gives Axolotl and TRL QLoRA recipes, QLoRA verified on one H100 80 GB.\n- **Qwen3-4B-Instruct-2507 (Alibaba Cloud (named in the licence's copyright line))**: Apache 2.0 with a LICENSE file, no remote code, and Unsloth's Qwen3-2507 page gives a QLoRA recipe that fits a 24 GB GPU.\n- **Qwen3.5-0.8B (Alibaba Cloud (named in the licence's copyright line))**: Apache 2.0 with LICENSE, no remote code, the card names fine-tuning as the intended use, and Unsloth's Qwen3.5 page gives a QLoRA recipe.\n- **Qwen3.5-2B (Alibaba Cloud (named in the licence's copyright line))**: Apache 2.0 with LICENSE, no remote code, the card names fine-tuning as the intended use, and Unsloth's Qwen3.5 page gives a QLoRA recipe.\n- **Qwen3.5-4B (Alibaba Cloud (named in the licence's copyright line))**: Apache 2.0, no remote code, and Unsloth's Qwen3.5 fine-tune page gives a QLoRA recipe for the dense sizes.\n- **Qwen3.5-9B (Alibaba Cloud (named in the licence's copyright line))**: Apache 2.0, no remote code, Unsloth QLoRA recipe for the dense sizes, and acc-25 trained a Czech LoRA with Unsloth that works at inference.\n- **Qwen3.8-27B (Alibaba Cloud (named in the licence's copyright line))**: Apache 2.0, no remote code, and Unsloth's Qwen3.8 train page says QLoRA works on 24 GB and LoRA needs over 36 GB, with GGUF export.\n- **Qwen3-VL-2B-Instruct (Alibaba Cloud (per the first batch's RDAP check of qwen.ai; this model's files name no copyright holder))**: Apache 2.0, no remote code, and Unsloth's Qwen3-VL run and fine-tune page gives a LoRA recipe with GGUF export.\n- **Qwen3-VL-4B-Instruct (Alibaba Cloud (per the first batch's RDAP check of qwen.ai; this model's files name no copyright holder))**: Apache 2.0, no remote code, and Unsloth's Qwen3-VL run and fine-tune page gives a LoRA recipe with GGUF export.\n- **Qwen3-VL-8B-Instruct (Alibaba Cloud (per the pilot's RDAP check of qwen.ai; this model's files name no copyright holder))**: Apache 2.0, no remote code, and Unsloth's Qwen3-VL run and fine-tune page gives a LoRA recipe with GGUF export.\n\n## Does the job, with a limit\n\nIt does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.\n\n- **Agents-A1-4B (Shanghai Artificial Intelligence Laboratory (the InternAgent team there, per the card's recommended system prompt and the template's default prompt); the GitHub LICENSE leaves the copyright line as the Apache template)**: Apache 2.0 BF16 of 9.1 GB in the Qwen3.5 architecture, which Unsloth trains (acc-25 did a 9B LoRA), but neither the card nor the GitHub repository publishes a recipe or VRAM figures.\n- **Apertus v1.5 8B (ETH Zurich and EPFL (the Swiss National AI Institute, named in the Acceptable Use Policy))**: Apache 2.0 BF16 (18.4 GB) permits derivatives, but downloading needs assent to the AUP, no recipe or VRAM figure is published, and no stock runtime loads the result.\n- **Apodex-1.1-mini (Apodex (the site's copyright line reads 2026 Apodex))**: Apache 2.0 LICENSE file and 72 GB of BF16, but a 36B MoE exceeds LoRA on a 24 GB GPU and Apodex publishes no recipe.\n- **BigBang-v1 (The Endless Frontier (endless-frontier))**: Apache 2.0 LICENSE file and BF16 weights permit derivatives, but a 36B MoE at 71.9 GB is beyond a 24 GB LoRA and no recipe is published.\n- **Carbon-X (Hugging Face (the HuggingFaceBio organisation, named Hugging Face Biology Research on the Hub))**: Apache 2.0 and a 29.7 GB BF16 Qwen3-14B that TRL post-trained, but the tokenizer is custom code behind trust_remote_code and no recipe or VRAM figure is given.\n- **Cosmos-Reason2-2B (NVIDIA Corporation)**: NVIDIA's repository covers post-training and the licence allows derivatives, but it is revocable, gated behind a click-through and the project is no longer under active development.\n- **Darwin-36B-Opus (VIDRAFT Inc. (named on vidraft.net); the card's citation names FINAL-Bench and VIDRAFT_LAB)**: Apache 2.0 LICENSE file since 1 November 2026 and BF16 weights, but a 35B MoE at 69 GB exceeds a 24 GB LoRA and the merge method is proprietary.\n- **DiffusionGemma 26B A4B IT (Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author))**: Apache 2.0 BF16 of 51.7 GB permits derivatives, but Google calls it experimental, publishes no recipe, and a 25B MoE exceeds a 24 GB LoRA.\n- **Gemma 4 26B A4B IT (Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author))**: Apache 2.0 BF16 permits derivatives, but Unsloth's train page says 26B-A4B LoRA needs more than 40 GB and has no QLoRA figure for the MoE.\n- **GLM-4.7-Flash (JINGSHENG HENGXING TECHNOLOGY PTE.LTD (operator of Z.ai per its terms of use, and registrant of z.ai))**: MIT BF16 (62.5 GB) and an Unsloth tutorial with a LoRA notebook, but Unsloth says runs out of memory on 40 GB and need an 80 GB A100 or H100; no LICENSE file in the repository.\n- **Granite 4.2 3B (International Business Machines Corporation (per IBM's naming guidance, as read for granite-guardian))**: Apache 2.0 BF16 of 7.3 GB on a plain dense architecture, but IBM publishes no recipe for 4.2 and Unsloth's Granite page covers 4.1.\n- **InternVL3.5-GPT-OSS-20B-A4B-Preview (Shanghai AI Laboratory (OpenGVLab), not named in the repository)**: Apache 2.0 BF16 permits derivatives and InternVL documents its training stages, but a 21B MoE at 42.5 GB exceeds a 24 GB LoRA and no VRAM figure is given.\n- **Kolibri-1 (Aleph Alpha GmbH)**: Apache 2.0 with a LICENSE file and a BF16 repository, but a 78B MoE exceeds a 24 to 96 GB LoRA and Aleph Alpha publishes no recipe.\n- **Laguna XS 2.1 (Poolside, Inc. (named in Poolside's Terms of Use))**: OpenMDW-1.1 permits derivatives and BF16 is 66.9 GB, so a LoRA needs the 96 GB end; no recipe or VRAM figures are published.\n- **LBS-27B (inclusionAI, Ant Group (named in the site footer of inclusion-ai.org); the work is credited to authors from Wuhan University of Technology, CASIA, Zhejiang University and Ant Group)**: Apache 2.0 in the card's metadata (no LICENSE file) and a 55 GB BF16 checkpoint that fits a 96 GB GPU for LoRA, but no recipe and no smaller base.\n- **LFM2.5-2.6B (Liquid AI, Inc. (the licensor named in the LFM Open License))**: Unsloth and TRL notebooks are linked from the card and Liquid's docs, but the LFM Open License bars commercial use of derivatives above the revenue threshold.\n- **Ling-3.0-tiny (Ant Intelligence (Hang Zhou) Technology Co., Ltd (named as provider in the training-content summary); inclusionAI, Ant Group in the site footer of inclusion-ai.org)**: MIT tag and a 15.8 GB BF16 that fits a 24 GB GPU, but the bailing_hybrid architecture loads through repository code and no Unsloth or TRL recipe exists.\n- **MiMo-V2.6 Distill Qwen 9B (Xiaomi (the card and report name Xiaomi MiMo; no LICENSE file names a copyright holder))**: Released as a starting point for RL under MIT, 18.8 GB BF16, but no LoRA recipe and a bf16 LoRA on 9B is tight on 24 GB.\n- **MiniCPM-o 4.5 (OpenBMB (named as the MiniCPM-o team on the technical report; the GitHub LICENSE is the unfilled Apache template))**: Apache 2.0 and a 20 GB BF16 checkpoint, but the architecture loads only with remote code and pickle files, and no Unsloth or TRL recipe is on the card.\n- **Nanbeige4.3-33B (BOSS Zhipin (Kanzhun Limited), unconfirmed)**: Standard Llama architecture and a Base checkpoint ship, but Apache 2.0 rests on a metadata tag with no LICENSE file, and no fine-tuning recipe or VRAM figure is published.\n- **NVIDIA Nemotron 3.5 Bolt E4B 4GB (NVIDIA Corporation)**: OpenMDW-1.1 plus Gemma 4's Apache 2.0 permit derivatives and a BF16 checkpoint ships, but NVIDIA publishes no fine-tuning recipe and the BF16 repository adds a frozen shared-KV file of its own.\n- **Nemotron 3.5 Lightning 30B A3B (NVIDIA Corporation)**: OpenMDW-1.1 with a LICENSE file and the card calls BF16 the starting point for post-training, but no LoRA recipe or VRAM figure is published for 3.5 Lightning; Unsloth's Nano 30B recipe needs about 60 GB.\n- **Nemotron 3 Nano 30B A3B (NVIDIA Corporation)**: Licence permits commercial derivatives and Unsloth has a Nano 30B LoRA notebook, but 16-bit LoRA needs about 60 GB and Unsloth warns against fine-tuning MoE routers; no LICENSE file in the repository.\n- **Nemotron 3 Nano 4B (NVIDIA Corporation)**: Unsloth says the 4B fits a free Colab GPU and the licence permits derivatives, but the repository's LICENSE file is empty and loading needs trust_remote_code.\n- **Nemotron 3 Nano Omni 30B A3B Reasoning (NVIDIA Corporation)**: BF16 weights ship and NVIDIA's agreement allows commercial use, but no fine-tuning recipe or VRAM figure is published and the modelling code is repository Python.\n- **occamy-1.0 (Accio (Accio-Lab))**: Apache 2.0 with a LICENSE file and no remote code, but Accio publishes no fine-tuning recipe or VRAM figure for the merged checkpoint.\n- **Olmo 3 7B Think (The Allen Institute for Artificial Intelligence (Ai2), a 501(c)(3) nonprofit, per its Responsible Use Guidelines)**: Apache 2.0, with SFT and DPO checkpoints, datasets and Open-Instruct training code published; but no LICENSE file in the repository, no VRAM figure, and the card ties use to Ai2's Responsible Use Guidelines.\n- **Ornith-1.5-35B-A3B (Ornith)**: The card points to Unsloth Studio and BF16 is 72 GB, but no VRAM figures are given and the MIT licence exists only as a tag.\n- **Ornith-1.5-9B (Ornith)**: 19 GB BF16 fits a 24 GB GPU for LoRA and the card points to Unsloth Studio, but no VRAM figures and the MIT licence is only tagged.\n- **Qwen3-30B-A3B-Instruct-2507 (Alibaba Cloud (named in the licence's copyright line))**: Apache 2.0 with a LICENSE file and an Unsloth recipe, but QLoRA needs about a 40 GB A100, so a 24 GB GPU is out.\n- **Qwen3.5-122B-A10B (Alibaba Cloud (named in the licence's copyright line))**: Apache 2.0 and no remote code, but Unsloth's Qwen3.5 page says MoE QLoRA is not recommended and quotes 74 GB of VRAM for MoE training, past a 24 GB GPU.\n- **Qwen3.6-27B (Alibaba Cloud (named in the licence's copyright line))**: Apache 2.0 and no remote code, but Unsloth says QLoRA is not recommended for this architecture and the dense 27B wants 16-bit LoRA, which the Qwen3.8 train page puts above 36 GB.\n- **Qwen3.6-35B-A3B (Alibaba Cloud (named in the licence's copyright line))**: Apache 2.0 and no remote code, but Unsloth says MoE QLoRA is not recommended for Qwen3.5-class MoE and gives 74 GB of VRAM for the MoE notebook.\n- **Qwen3-Coder-30B-A3B-Instruct (Alibaba Cloud (named in the licence's copyright line))**: Apache 2.0 and no remote code; Unsloth's Qwen3 page covers Qwen3 MoE fine-tuning but gives no VRAM figure for this checkpoint.\n- **Qwen3-Omni-30B-A3B-Instruct (Alibaba Cloud (carried from the 20 October 2026 pilot; this model's files name no copyright holder))**: Apache 2.0 by metadata and the card shows a downstream fine-tune (Captioner), but no LICENSE file in the repository and no recipe or VRAM figure.\n- **Qwen3-VL-30B-A3B-Instruct (Alibaba Cloud (per the pilot's RDAP check of qwen.ai; this model's files name no copyright holder))**: Apache 2.0 and no remote code; Unsloth's Qwen3-VL page covers the family but gives no MoE VRAM figure for a 24 GB GPU.\n- **Qwen3-VL-32B-Instruct (Alibaba Cloud (per the first batch's RDAP check of qwen.ai; this model's files name no copyright holder))**: Apache 2.0 and no remote code with Unsloth's Qwen3-VL recipe, but a 32B dense LoRA in bf16 passes 24 GB; no QLoRA figure for this size.\n- **Qwopus3.8-27B-Flash-V2 (Jackrong, an individual Hugging Face account)**: Apache 2.0 by metadata and Unsloth's Qwen3.8 train recipe applies to the architecture, but no LICENSE file, and the card says its data includes traces derived from Claude and GPT.\n- **SmolLM3-3B (Hugging Face, Inc.)**: Apache 2.0 and no remote code, with TRL named and training configs in the huggingface/smollm repository, but no Unsloth page or VRAM figure, and no LICENSE file.\n- **Spark-X2.6-4B (XHToken (the licence's copyright line); the Hugging Face organisation's name is SparkLLM)**: Apache 2.0, but config.json carries auto_map with modeling_spark.py, Transformers needs trust_remote_code, and the card points to SparkLLM's LlamaFactory fork rather than Unsloth or TRL.\n- **Swift 1.5 Qwen3.8-27B (UkisAI)**: Same architecture as Qwen3.8-27B, so Unsloth's QLoRA on 24 GB applies, but derivatives inherit the Swift Open License's US$1M revenue cap.\n\n- not this job: DeepSeek-V3.2 (DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)): 689.5 GB of FP8 weights cannot be LoRA-tuned on a 24 to 96 GB GPU.\n- not this job: DeepSeek V4.1 Flash (DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)): 510 GB of weights; no single-GPU LoRA path.\n- not this job: DeepSeek-V4-Flash-0731 (DeepSeek (the MIT licence names DeepSeek, and DeepSeek's platform terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)): 166.9 GB of weights; no single-GPU LoRA path.\n- not this job: GLM-4.5-Air (JINGSHENG HENGXING TECHNOLOGY PTE.LTD (operator of Z.ai per its terms of use, and registrant of z.ai, per the first batch's check on 21 October 2026)): 221 GB of BF16 weights; no LoRA on a 24 to 96 GB GPU and no recipe.\n- not this job: GLM-5.3 (JINGSHENG HENGXING TECHNOLOGY PTE.LTD (operator of Z.ai per its terms of use, and registrant of z.ai). The licence names Z.AI): 755.7 GB of weights and a licence that adds a security review for model-as-a-service operators above US$10 billion.\n- not this job: GLM-5.3-Flash (Z.AI Co., Ltd (named in the licence's copyright line)): 328 GB of FP8 and 643 GB of BF16; no single-GPU LoRA path.\n- not this job: Hy3 (Tencent (named in the LICENSE file's copyright line)): Apache 2.0 with a finetune folder (LLaMA-Factory, DeepSpeed, ms-swift), but 597.6 GB of BF16 is beyond a 24 to 96 GB GPU.\n- not this job: Inkling (Thinking Machines Lab, Inc. (named on its model card page and in its training-content summary)): 1,905 GB of BF16; the Unsloth recipe the card links is for the Small model.\n- not this job: Inkling-Small (Thinking Machines Lab, Inc. (named on its model card page and in its training-content summary)): Apache 2.0 and an Unsloth recipe link, but 532 GB of BF16 cannot be LoRA-tuned on a 24 to 96 GB GPU; the AUP binds on download.\n- not this job: Kimi K3 (Moonshot AI (the licence names Moonshot AI; the Kimi Platform terms name Moonshot AI PTE. LTD.)): 1,561 GB of weights, trust_remote_code modelling files and a licence that needs a separate agreement above US$20 million of revenue.\n- not this job: Kronos (No company. Yu Shi publishes it from the GitHub account shiyu-coder and the Hugging Face user account NeoQuasar (named ShiYu), with co-authors at Tsinghua University per the paper): The card has a finetuning section for price series under MIT, but the use case asks for a language base trained with Unsloth or TRL, which Kronos is not.\n- not this job: Laguna M.1 (Poolside, Inc. (named in Poolside's Terms of Use; the LICENSE.md copyright line reads Poolside)): BF16 files are 451 GB; a LoRA on a 24 GB to 96 GB GPU is out of reach and the card gives no recipe.\n- not this job: Laguna S 2.1 (Poolside, Inc. (named in Poolside's Terms of Use)): BF16 is 235 GB and the card gives no fine-tuning recipe; a 24 GB to 96 GB GPU cannot hold it.\n- not this job: MiniMax-M2.5 (MiniMax AI (named in the licences; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore)): 230 GB of FP8 weights and no recipe; out of reach for a 24 GB to 96 GB GPU.\n- not this job: MiniMax-M2.7 (MiniMax (named in the licence's copyright line; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore)): 230 GB of FP8 weights and no recipe, and the licence bars commercial use of fine-tuned derivatives without written authorisation.\n- not this job: MiniMax-M3.1-Flash (MiniMax (named in the licence's copyright line; MiniMax's platform terms name Nanonoble Pte. Ltd., Singapore)): 250 GB of weights; the card lists Unsloth support as in progress, with no recipe for a 24 GB to 96 GB GPU.\n- not this job: Mistral Large 4 (Mistral AI (RCS Paris 952 418 325)): Unsloth and TRL guides exist, but a 1T model cannot train on a 24 GB to 96 GB GPU.\n- not this job: Nemotron 3 Super 120B A12B (NVIDIA Corporation): Unsloth says bf16 LoRA needs 256 GB of VRAM, beyond a 24 GB to 96 GB GPU.\n- not this job: Nemotron 3 Ultra 550B A55B (NVIDIA Corporation): 1,121 GB of BF16 weights; no LoRA path fits a 24 GB to 96 GB GPU.\n- not this job: Ornith-1.5-397B (Ornith): About 807 GB in BF16; no single GPU holds it.\n- not this job: Qwen3.5-397B-A17B (Alibaba Cloud (named in the licence's copyright line)): 807 GB in BF16; Unsloth quotes 256 GB of VRAM for the largest Qwen3.5 MoE, outside a 24 to 96 GB GPU.\n- not this job: Qwen3.8-Flash-Next (Alibaba Cloud (Singapore) Private Limited, the registrant of qwen.ai. The licence names Qwen as copyright holder): 360 GB in BF16 with a 51B embedding table, no Unsloth train page, and the Qwen Community License needs a separate licence for model-as-a-service and AI work assistant businesses.\n- not this job: Qwen3-Coder-Next (Alibaba Cloud (the Qwen team's employer; no LICENSE file names a copyright holder for this model)): Unsloth says bf16 LoRA needs a single B200, outside a 24 to 96 GB GPU, and the repository carries no LICENSE file for its Apache 2.0 claim.\n- not this job: Recursive Flint (Recursive Superintelligence, Inc. (the Hugging Face organisation's full name)): Apache 2.0, but the card says Transformers modelling code is not included yet, so neither Unsloth nor TRL can load it; Unsloth's Flint page covers running only.\n- not this job: Ripple-65B Preview (Ricursive Intelligence, Inc.): Apache 2.0, but config.json carries auto_map and the repository ships modeling_ricursive.py, so loading needs remote code; no Unsloth or TRL recipe exists.\n- not this job: Swift 1.5 Qwen3.8-Flash-Next (UkisAI): 360 GB in BF16, no train recipe for the architecture, and two licences with a revenue cap and a model-as-a-service clause.\n- not this job: ThinkingCap (Qwen3.8 27B) (BottleCap AI (named in the licence's copyright line)): PolyForm Small Business 1.0.0 limits commercial use of derivatives to small organisations, and the BF16 repository is gated behind contact details.\n- not this job: Baidu Unlimited-OCR-2 (Baidu, Inc.): MIT, but loading runs Baidu's remote code with a custom sparse attention and no Unsloth or TRL recipe exists.\n\n28 listings in the category were added after this reading and aren't marked yet.\n\n## What stops agents finishing\n\n- licence forbids derivatives or commercial use of the result\n- QLoRA not supported for the architecture, so bf16 LoRA needs 22 GB\n- serving runtime cannot load the adapter or merged GGUF\n\n## Around it\n\nBefore: datasets.prepare. After: open-weights.tool-calling-assistant, open-weights.private-chat-knowledge.\n\n## Where this comes from\n\nWeight 8 of 100 in Open-weight models: 1 of 51 accounts plus vendor claims on three cards; weight raised above the count because the category description names fine-tuning as a purpose\n\nAs data: https://www.anchorterminal.com/api/v1/use-cases/open-weights.fine-tune-base.json, or the MCP tool get_use_case with id open-weights.fine-tune-base. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.\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": "2026-11-16",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "0.8",
    "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": "For agents",
        "url": "https://www.anchorterminal.com/agents/"
      },
      {
        "name": "Niches",
        "url": "https://www.anchorterminal.com/niches/"
      },
      {
        "name": "Fine-tune a base for a language, voice or narrow task",
        "url": ""
      }
    ],
    "description": "Fine-tune a base for a language, voice or narrow task: 9 companies fully do it, 27 with a limit, 90 carry the tag. Every listing in Open-weight models with the reason and a source.",
    "facts": null,
    "h1": "Fine-tune a base for a language, voice or narrow task",
    "image": "https://www.anchorterminal.com/assets/og/default.png",
    "path": "/niches/open-weights.fine-tune-base/",
    "published": "",
    "section": "agents",
    "title": "Fine-tune a base for a language, voice or narrow task | Anchor…",
    "toc": null,
    "updated": "2026-11-15",
    "url": "https://www.anchorterminal.com/niches/open-weights.fine-tune-base/"
  },
  "tokens": {
    "markdown": 5750,
    "slim": 1630
  },
  "version": 1
}
