What Is ERNIE 5.5 Flash Base? The Beginner's Guide to Baidu's 280B Open-Weight Checkpoint
CodexFuse Team · 2026-11-05T19:57:03.079Z · 21 min read min read
What Is ERNIE 5.5 Flash Base? The Beginner's Guide to Baidu's 280B Open-Weight Checkpoint
What Is ERNIE 5.5 Flash Base?
ERNIE 5.5 Flash Base is a text-only, 280-billion-parameter open-weight language model that Baidu released on November 5, 2026. It is a base checkpoint, which means it sits at the end of pre-training and has received no instruction tuning or post-training at all. Developers download the weights and build on them, rather than chatting with a finished assistant.
The release matters because it is Baidu's first open-weight model in roughly a year and a half. The ERNIE 4.5 series arrived in June 2025, and the 5.0 and 5.1 versions that followed were never open-sourced. This checkpoint is aimed at researchers and engineers who want a large, fixed starting point for fine-tuning, domain adaptation, and reinforcement-learning experiments.
If you work with large models, you already know the workflow: download verified weights, check their integrity with a SHA-256 hash generator, and run them on rented server hardware. The same pattern applies here, and the hardware bar is high. Most readers will run this model on VPS hosting with server-class GPUs, not on a laptop, because the weights alone fill hundreds of gigabytes.
This guide is the ernie 5.5 flash base explained in plain terms: what the model is, how its architecture works, how to run it, what its license really allows, and where its benchmark claims hold up.
| Fact | Detail |
|---|---|
| Developer | Baidu ERNIE Team |
| Released | November 5, 2026 |
| Type | Text-only base checkpoint, no post-training |
| Parameters | 280B total, 14B activated per token |
| Pre-training | 33.5 trillion tokens, trained from scratch |
| Context | 128K training, up to 1M tokens for inference |
| Weights | FP8 (about 288 GB) and BF16 (about 560 GB) |
| License | Custom ERNIE 5.5 Flash Base Model License v1.0 |
| Serving | SGLang and vLLM through the completions API |
Why Did Baidu Release a Base Checkpoint Instead of a Chatbot?
A base model is the raw result of pre-training: a next-token predictor with no chat alignment, no refusal behavior, and no instruction-following format baked in. Baidu calls this checkpoint a common starting point, and the framing is deliberate. The company wants developers to do the post-training themselves, for their own domains, rather than shipping one finished assistant.
That choice also makes the release useful as a fixed baseline. When every research team starts from the same public checkpoint, experiments on fine-tuning methods, reinforcement learning, and domain adaptation become directly comparable. A finished chatbot hides its training recipe; a base checkpoint exposes the starting line.
The post-trained version is coming later. Baidu says the post-trained ERNIE 5.5 Flash will follow this base version, but it has given no date. Until then, anyone who downloads the weights gets the raw model and the responsibility that comes with it.
There is one more reason the timing matters. Open-weight releases at this size have become the way model labs show their work. Baidu had stayed quiet on open weights for about 16 months, and this release re-enters that conversation with a model built for cheap inference rather than raw size.
How Big Is ERNIE 5.5 Flash Base, Really?
The headline number is 280 billion parameters, but only 14 billion are active for each token. That is 5 percent of the model doing the work on any given token, since 14 divided by 280 equals 0.05. The rest sits idle, which is the whole point of a mixture-of-experts design: a huge model that costs a fraction of its size to run.
Pre-training used 33.5 trillion tokens, which works out to about 120 tokens per total parameter and roughly 2,393 tokens per activated parameter. Training ran in FP8 mixed precision with the Muon optimizer, a combination Baidu chose for efficiency at this scale. The tokenizer is a new byte-pair encoding vocabulary of 201,216 entries, and it tokenizes numbers right-to-left in groups of up to three digits to help with arithmetic.
The context window is 128,000 tokens during training and stretches to 1,048,576 tokens, about 1 million, for fine-tuning and inference. That is 8 times the training context, since 1,048,576 divided by 131,072 equals 8. Long documents, whole codebases, and long multi-turn histories fit in a single request.
| Specification | Value |
|---|---|
| Total parameters | 280 billion |
| Activated per token | 14 billion (5 percent of total) |
| Pre-training tokens | 33.5 trillion |
| Transformer layers | 43 |
| Hidden size | 4,096 |
| Attention heads | 64 |
| Training context | 128K tokens (131,072) |
| Inference context | Up to 1M tokens (1,048,576) |
| Optimizer | Muon, FP8 mixed precision |
| Tokenizer vocabulary | 201,216 entries |
What Makes the Architecture Different?
ERNIE 5.5 Flash Base is built around one idea: make each token cheap without shrinking the model. The architecture stacks several efficiency techniques, and each one attacks a different cost. The result is a model that activates 5 percent of its weights per token and compresses most of its long-context memory by 128 times.
The attention system is a hybrid. Thirty-seven of the 43 layers use Heavily Compressed Attention, which squeezes every 128 historical tokens into a single cached entry plus a short window of recent uncompressed tokens. The remaining 6 layers use Multi-head Latent Attention with sparse attention, and those 6 layers keep full token-level access to the original content.
That is about 86 percent of layers compressed and about 14 percent keeping fine detail, since 37 divided by 43 is roughly 0.86 and 6 divided by 43 is roughly 0.14.
The expert system is equally aggressive. Each mixture-of-experts layer holds 512 routed experts, and only the top 10 are activated per token, plus one shared expert that is always on. That means about 1.95 percent of routed experts fire for any token, since 10 divided by 512 is about 0.0195.
Experts work in a 2,048-dimensional latent space, half the 4,096 hidden width, which roughly halves the traffic of dispatching tokens to experts.
Three more ideas round out the design. Low-Rank Gated Attention computes its gate from a low-rank query latent one quarter the hidden size, cutting gate-projection parameters by 75 percent. Manifold-constrained Hyper-Connections replace the standard residual stream with 4 parallel residual streams.
ECHO replaces the usual multi-token prediction module: the backbone's own final layer predicts future tokens, and the same layer doubles as a draft model for speculative decoding.
The training setup behind the model is just as efficiency-focused. Baidu trained with the Muon optimizer and custom FlashMask-4 attention kernels, and it reports those kernels running up to 175 percent faster forward and 74 percent faster backward than FlashAttention-4 MaskMod on SM100 hardware. Those figures are vendor-reported, like the benchmarks.
The pre-training data mixed multilingual web text with code repositories and software-evolution data, including commits, issues, pull requests, and CI-verified fixes. It also included curated long-context documents and citation-grounded question answering, which explains the model's strong long-document scores despite its compressed attention.
| Component | How it works |
|---|---|
| Hybrid attention | 37 HCA layers compress 128 tokens into 1 cache entry; 6 MLA layers keep full detail |
| MoE routing | 512 routed experts, top 10 active per token, plus 1 shared expert |
| Expert width | 2,048-dim latent space, half the 4,096 hidden width |
| Low-Rank Gated Attention | Gate from a quarter-width latent cuts gate parameters by 75 percent |
| Hyper-Connections | 4 parallel residual streams replace the single residual stream |
| MultiMax output head | Learnable temperature suppresses implausible tokens |
| ECHO | Final layer predicts future tokens and doubles as a speculative-decoding draft model |
How Does the 1M-Token Context Actually Work?
A million-token context sounds impossible on a model that compresses most of its history, and the trick is that only 6 layers ever see the full sequence. The 37 compressed layers attend over a 128-times-shorter sequence, so per-query attention cost stays roughly constant as context grows. The 6 full-detail layers are enough, in Baidu's words, to preserve long-context performance.
The benchmark results back that claim partway. ERNIE leads 4 of the 6 long-context benchmarks Baidu reports, including RepoQA at 91.7 and HELMET-ICL at 85.9. It loses the other two, HELMET-Rerank and InfiniteBench, to DeepSeek, which is a reminder that compression has a price.
Retrieval-heavy work that needs exact token matches across a huge document is where the design is weakest.
For a developer, the practical read is simple. Summarizing a long document, reasoning across a large codebase, or holding a long conversation all fit comfortably in the 1M-token window. Tasks that need precise reranking of many long passages deserve a test before you commit, because that is where the model trails its rival.
| Context mode | Length | What it is for |
|---|---|---|
| Training context | 128K tokens | Pre-training sequences, the length the model learned on |
| Fine-tuning context | Up to 1M tokens | Long-document tuning and domain adaptation |
| Inference context | Up to 1M tokens | Long documents, codebases, and extended histories |
| Compressed layers | 37 of 43 | 128 tokens squeezed into 1 cache entry each |
| Full-detail layers | 6 of 43 | Token-level access that preserves long-context quality |
How to Run ERNIE 5.5 Flash Base
Running the model means downloading the weights, loading them in a serving engine, and calling the completions API. It is a base model, so there is no chat interface: you send a text prompt and get a text continuation back. The ernie 5.5 flash base download lives on Hugging Face in two variants, FP8 and BF16, and both are served the same way.
- Download the weights from Hugging Face, choosing FP8 for the smaller download or BF16 for full precision.
- Verify the downloaded shards before loading them, so a corrupt file does not waste hours of GPU time.
- Start a serving engine, SGLang or vLLM, pointed at the model path with enough parallelism for the weight size.
- Send a prompt to the completions endpoint and confirm the model returns a sensible continuation.
- Fine-tune or prompt-engineer from there, keeping a BOS token at the start of every input.
SGLang ships an official prebuilt Docker image for this model, and the example launch uses tensor parallelism of 4 with a 131,072-token context. The vLLM route needs version 0.32.0 plus Baidu's vLLM-ERNIE plugin, with FP8 key-value cache and expert parallelism enabled. Both engines accept up to 1,048,576 tokens of context for long-context inference.
A minimal test looks like a curl call to the local completions endpoint with a plain prompt such as "Large language models are" and a temperature of 0. If the continuation reads like coherent English, the model is loaded and serving. From there, the real work is prompt design and fine-tuning, because the base model will happily continue any text you give it, including text you would rather it refused.
| Download option | Format | Size | Best for |
|---|---|---|---|
| baidu/ERNIE-5.5-Flash-Base | FP8 safetensors | About 288 GB (128 shards of about 2.25 GB) | Cheaper inference, smaller download |
| baidu/ERNIE-5.5-Flash-Base-BF16 | BF16 safetensors | About 560 GB (derived, roughly double FP8) | Full precision, fine-tuning starting point |
| Serving with SGLang | Completions API | Official Docker image available | Production inference at scale |
| Serving with vLLM | Completions API | Requires the vLLM-ERNIE plugin | Teams already running vLLM |
| Fine-tuning recipe | Megatron Bridge | 64 GPUs for full-model SFT | Supervised fine-tuning at full scale |
How Much Hardware Do You Need?
The honest answer starts with the download: 280 billion parameters means about 288 GB of FP8 weights, verified from the 128-shard file listing, and about 560 GB in BF16, derived from 280 billion times 2 bytes per parameter. That is before the key-value cache, activations, and working memory. A single consumer GPU cannot hold the full weights, and no fetched source states a minimum GPU count for inference.
What the sources do show is the shape of a real deployment. The SGLang example runs with tensor parallelism of 4 and attention data parallelism of 4. The vLLM example uses expert parallelism plus data parallelism of 4.
Full-model supervised fine-tuning is a different job entirely: the published Megatron Bridge recipe uses 16 nodes of 4 GB300 GPUs, which is 64 GPUs, validated on Blackwell GB200 and GB300 hardware.
For most teams, the sensible path is renting multi-GPU server time and running the FP8 weights. Quantization to FP8 halves the memory footprint versus BF16, and the serving engines are built for it. Budget for the download itself too: 288 GB over a slow link is an overnight job, and the BF16 variant roughly doubles it.
This is also where the model's open-model siblings give useful context. What SWE-2 Flash is matters here because it is another recent open-weight release aimed at developers who run models themselves, and the same server-class hardware logic applies. The pattern across this generation of open models is consistent: big weights, cheap per-token cost, and a GPU bill that only makes sense on rented infrastructure.
What Do the Benchmarks Actually Say?
Baidu reports that ERNIE 5.5 Flash Base finishes first in 24 of 30 benchmarks against three similarly sized open base models. That is 80 percent of the table, since 24 divided by 30 equals 0.8. The comparison models are DeepSeek-V4-Flash-Base, Tencent Hy3-preview-Base, and Xiaomi MiMo-V2.5-Base, all evaluated with the same pipeline, prompts, and few-shot examples.
Every score in this section is vendor-reported by Baidu, and no independent third-party reproduction was found.
The strongest results are in exams and code. On MMLU, ERNIE scores 93.3 against DeepSeek at 89.4, a lead of 3.9 points. On HumanEval+, it scores 82.3 against DeepSeek at 69.5, a gap of 12.8 points, and against Hy3 at 78.0, a lead of 4.3 points.
On LiveCodeBench it scores 40.1 against DeepSeek at 31.9, a lead of 8.2 points. These are large margins on well-known benchmarks, and they are the reason the release is being taken seriously.
The losses matter just as much. ERNIE trails DeepSeek on ChineseSimpleQA, 73.0 against 75.3, a deficit of 2.3 points, and trails Hy3 by 0.2 points on the same test. It loses FullStackBench to Hy3, 54.6 against 56.7, a deficit of 2.1 points.
It loses InfiniteBench to DeepSeek, 46.5 against 48.0, and HELMET-Rerank to DeepSeek, 56.2 against 60.0. It also loses BBH with chain-of-thought to Hy3, 89.2 against 90.0, and INCLUDE-lite to DeepSeek by 0.3 points, 79.5 against 79.8.
Two caveats frame the whole table. First, the comparison models are the latest similarly sized open base models Baidu could find at release time, and commentators have noted that newer versions from each vendor exist. Second, a count of first-place finishes is not an overall score: near-ties like RepoBench, 50.3 against 49.1, count the same as the 12.8-point HumanEval+ gap.
Read the table as a strong exam-and-code showing with real weak spots, not a clean sweep.
The 24 first-place finishes break down by category, and the pattern is revealing. ERNIE leads 4 of 4 exam benchmarks, 4 of 5 knowledge benchmarks, all 3 STEM benchmarks, 3 of 4 reasoning benchmarks, 2 of 3 multilingual benchmarks, 4 of 5 code benchmarks, and 4 of 6 long-context benchmarks. Add those up, 4 plus 4 plus 3 plus 3 plus 2 plus 4 plus 4, and you get 24.
The weakest categories are long context and multilingual, which is consistent with the compression trade-off and the ChineseSimpleQA loss. A team choosing between ERNIE and its rivals should weigh those two categories heavily if their workload is retrieval-heavy or multilingual, because that is where the headline lead thins out.
| Benchmark | ERNIE 5.5 Flash Base | DeepSeek-V4-Flash-Base | Best rival |
|---|---|---|---|
| MMLU (5-shot) | 93.3 | 89.4 | ERNIE leads by 3.9 |
| MMLU-Pro (5-shot) | 78.4 | 69.2 | ERNIE leads by 9.2 |
| HumanEval+ (1-shot) | 82.3 | 69.5 | ERNIE leads by 12.8 |
| LiveCodeBench (1-shot) | 40.1 | 31.9 | ERNIE leads by 8.2 |
| MBPP+ (0-shot) | 81.5 | 67.2 | ERNIE leads by 14.3 |
| GPQA (5-shot) | 60.2 | 55.9 | ERNIE leads by 4.3 |
| RepoQA (0-shot) | 91.7 | 81.0 | ERNIE leads by 10.7 |
| ChineseSimpleQA (10-shot) | 73.0 | 75.3 | DeepSeek leads by 2.3 |
| FullStackBench (1-shot) | 54.6 | 43.3 | Hy3 leads at 56.7 |
| InfiniteBench (0-shot) | 46.5 | 48.0 | DeepSeek leads by 1.5 |
ERNIE 5.5 Flash Base vs DeepSeek V4 Flash: How Do They Compare?
The ernie 5.5 flash base vs deepseek v4 flash comparison is the one most readers will make first, because DeepSeek-V4-Flash-Base is the closest rival in size and the best-known name in the table. DeepSeek's model carries 284 billion total parameters with 13 billion activated per token, against ERNIE's 280 billion total and 14 billion activated. The two are near-twins on paper, which makes the benchmark gaps more informative.
On exams and code, ERNIE wins clearly in Baidu's numbers. MMLU favors ERNIE by 3.9 points, MMLU-Pro by 9.2 points, HumanEval+ by 12.8 points, and MBPP+ by 14.3 points. These gaps are large enough that they would survive modest measurement noise, and they suggest ERNIE's pre-training data, which Baidu describes as including software-evolution data such as commits, issues, and pull requests, paid off for coding tasks.
On knowledge and long context, DeepSeek fights back. It wins ChineseSimpleQA by 2.3 points, InfiniteBench by 1.5 points, and HELMET-Rerank by 3.8 points. Those three wins share a theme: precise factual recall and exact retrieval across long passages.
That is exactly where ERNIE's 128-times compression on 86 percent of its layers should hurt most, and the numbers line up with the theory.
The fair summary is that ERNIE looks like the stronger general-purpose base for exam-style reasoning and code generation, while DeepSeek looks like the safer pick for Chinese factual question answering and long-context reranking. Both readings depend on Baidu's own evaluation pipeline, so treat the margins as directional until an independent lab reproduces them.
What Does the ERNIE 5.5 Flash Base License Actually Allow?
The license is the part of this release that surprises most people, and it deserves a careful read. ERNIE 5.5 Flash Base ships under a custom agreement, the ERNIE 5.5 Flash Base Model License Agreement v1.0, dated November 5, 2026. It is open-weight, not open source in the Apache or MIT sense, and the difference is not cosmetic.
The first hard limit is territory. The license grants rights only in areas worldwide other than the European Union, the United Kingdom, and South Korea. Use, modification, and distribution in those three regions fall beyond the rights granted and are not authorized.
A developer in Berlin, London, or Seoul cannot legally use this model under its license, full stop.
The second hard limit is money. Non-commercial use is free with no notice needed. Commercial use is free up to 50 million dollars of related revenue in the preceding 12 consecutive months, provided the operator displays "Powered by ERNIE" prominently and sends a one-time use notice to Baidu before first commercial use.
Above 50 million dollars, separate prior written authorization from Baidu is required, and use lacking authorization must be suspended once the threshold is crossed.
The third hard limit is the one that catches AI builders off guard. The license bars using the model, its derivative models, or their outputs to improve any other artificial intelligence model or machine-learning system. That rules out distilling ERNIE into a smaller model, training a competitor on its outputs, or using its generations as synthetic training data for another system.
ERNIE 4.5 shipped under Apache 2.0, which had none of these limits, so this is a sharp break from Baidu's previous open release.
| Rule | What it means |
|---|---|
| Territory | Worldwide except the EU, UK, and South Korea, where use is not authorized |
| Non-commercial use | Free, no notice or authorization needed |
| Commercial use up to $50M | Free, but requires "Powered by ERNIE" attribution and a one-time notice to Baidu |
| Commercial use above $50M | Prior written authorization from Baidu required; unapproved use must stop |
| Distillation ban | Outputs may not be used to improve any other AI model |
| Derivatives | Fine-tuning, quantization, and merging allowed within the territory |
| Governing law | Hong Kong SAR; no warranty, provided as is |
Who Should Use It, and Who Should Skip It?
The model fits one audience well: teams with server-class GPUs who want a large, fixed base checkpoint to fine-tune for their own domain. If you are building a coding assistant, a domain-specific model, or a reinforcement-learning research setup, and you operate outside the excluded regions, this is a strong starting point with genuinely good exam and code numbers behind it.
It fits poorly for everyone else. End users who want a chatbot should wait for the post-trained ERNIE 5.5 Flash, because this checkpoint has no chat alignment and no refusal behavior. Developers in the EU, UK, or South Korea should not use it at all.
Anyone planning to distill it into another model should look elsewhere, because the license forbids exactly that.
Small teams should also think twice about the operating cost. A 288 GB download, multi-GPU inference, and a 64-GPU fine-tuning recipe put this firmly in server territory. If your use case fits a smaller model, a smaller model will be cheaper, faster, and simpler to run, and the benchmark lead on a 280-billion-parameter base will not matter.
| You are | Verdict | Why |
|---|---|---|
| Research lab fine-tuning a large base | Good fit | Fixed 280B checkpoint, strong code scores, open weights |
| Team building a domain coding assistant | Good fit | HumanEval+ 82.3 and LiveCodeBench 40.1, vendor-reported |
| End user wanting a chatbot | Skip it | No post-training, no chat alignment, no refusal behavior |
| Developer in the EU, UK, or South Korea | Skip it | License excludes those regions entirely |
| Team planning to distill it | Skip it | License bars using outputs to improve another model |
| Small team on a tight budget | Think twice | 288 GB weights and multi-GPU inference are server-class costs |
What Comes Next for ERNIE 5.5 Flash Base?
The near-term milestone is the post-trained ERNIE 5.5 Flash, which Baidu says will follow the base version. That release will show whether the strong base numbers survive instruction tuning and turn into a usable assistant. Until it arrives, the base checkpoint is the whole story, and its value depends on what the community builds on top of it.
Watch three things. First, independent benchmark reproductions, which will confirm or shrink the gaps Baidu reports. Second, community fine-tunes, which will reveal how the model behaves once it has chat alignment and safety tuning.
Third, whether the license limits adoption: a model that excludes the EU, UK, and South Korea and gates big commercial use at 50 million dollars may see slower uptake than its Apache-licensed predecessor.
The broader pattern is worth noting too. Open-weight releases keep arriving across every modality, from image models like Nano Banana to video models like Sora 2, and each one comes with its own license, hardware demands, and honest limitations. ERNIE 5.5 Flash Base is the latest entry in that wave: a capable, efficient, heavily conditioned release that rewards careful reading before downloading.
Sources
Baidu ERNIE Blog, ERNIE 5.5 Flash Base release post, November 5, 2026, https://ernie.baidu.com/zh/release/ernie-5-5-flash-base/
Hugging Face model card, baidu/ERNIE-5.5-Flash-Base, https://huggingface.co/baidu/ERNIE-5.5-Flash-Base
Hugging Face, ERNIE 5.5 Flash Base Model License Agreement v1.0, https://huggingface.co/baidu/ERNIE-5.5-Flash-Base/blob/main/LICENSE
Unite.AI, Baidu Releases ERNIE 5.5 Flash Base for Further Training, November 5, 2026, https://www.unite.ai/baidu-releases-ernie-5-5-flash-base-for-further-training/
The Decoder, Baidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model, November 5, 2026, https://the-decoder.com/baidu-releases-ernie-5-5-flash-base-with-280-billion-parameters-as-an-open-source-foundation-model/
ChainCatcher, Baidu opensourced Wenxin 5.5 Flash, November 5, 2026, https://www.chaincatcher.com/en/article/2302012
KuCoin News, Baidu open-sources the ERNIE 5.5 Flash model with 280 billion parameters, November 5, 2026, https://www.kucoin.com/news/flash/baidu-open-sources-ernie-5-5-flash-model-with-280-billion-parameters
All facts in this article were verified against these sources on November 6, 2026. Benchmark figures are vendor-reported by Baidu and have not been independently confirmed.
Frequently Asked Questions
Is ERNIE 5.5 Flash Base free to use?
The weights are free to download, and non-commercial use costs nothing. Commercial use is free up to 50 million dollars of related revenue in 12 months, with "Powered by ERNIE" attribution and a one-time notice to Baidu. Above that threshold, you need Baidu's prior written authorization.
Use in the EU, UK, and South Korea is not authorized at all.
Can I chat with ERNIE 5.5 Flash Base like ChatGPT?
No. It is a base checkpoint with no instruction tuning or post-training, so it predicts the next token rather than answering as an assistant. You would need to fine-tune it, or wait for the post-trained ERNIE 5.5 Flash that Baidu says will follow, before it behaves like a chatbot.
How do I download ERNIE 5.5 Flash Base?
The weights are on Hugging Face as baidu/ERNIE-5.5-Flash-Base in FP8, about 288 GB across 128 shards, and as baidu/ERNIE-5.5-Flash-Base-BF16 in BF16, about 560 GB. Download them with the Hugging Face tools, verify the shards, and serve them with SGLang or vLLM through the completions API.
How does ERNIE 5.5 Flash Base compare to DeepSeek V4 Flash?
In Baidu's vendor-reported benchmarks, ERNIE leads on exams and code, including MMLU at 93.3 against 89.4 and HumanEval+ at 82.3 against 69.5. DeepSeek leads on ChineseSimpleQA, InfiniteBench, and HELMET-Rerank. The two models are near-twins in size, so the choice depends on whether you value code and exam performance or factual recall and long-context reranking.
What hardware do I need to run ERNIE 5.5 Flash Base?
Server-class multi-GPU hardware. The FP8 weights alone are about 288 GB, and the serving examples use 4-way parallelism. No fetched source states a minimum GPU count, and a single consumer GPU cannot hold the full weights.
Full-model fine-tuning uses a published recipe of 64 GB300 GPUs.
