Nex-N2.5-mini Text Generation Model: Our Test Answers, 16GB+ VRAM
- 1. At a Glance
- 2. Overview
- 3. Specifications
- 4. Performance
- 5. Strengths and Use Cases
- 6. Hardware Requirements
- 7. Can You Run It Locally?
- 8. Our Own Measurements
- 9. Recent Models in the Same Size Class
- 10. How to Get It
- 11. Quantized and Converted Variants
- 12. Related Articles
- 13. What to Read Next
- 14. Sources
- 15. Update History
At a Glance
| Item | Value |
|---|---|
| Repository | nex-agi/Nex-N2.5-mini |
| Family guide | Nex-N2.5 guide (2 articles) |
| Publisher guide | Nex-AGI: models and licenses |
| Published | 2026-09-08 |
| License | apache-2.0 |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
What we checked ourselves
- Ran
nex-agi_Nex-N2.5-mini-Q2_K.gguf(Q2_K) on our CPU-only server (Neoverse-N1, 3 threads, llama.cpp b11223): generation at 7.3 tokens/s (a CPU figure, not a GPU speed). - Its answers to our fixed Japanese questions are shown verbatim below.
- It needs about 21% fewer tokens than the Qwen3 tokenizer for the same Japanese text, so about 1.26× as much Japanese fits in the same context length, and generation is faster per character.
Details and conditions are in “Our Own Measurements” below.
Overview
Nex-AGI has released “Nex-N2.5-mini", a multimodal foundational model from its next-generation agent model family “Nex-N2.5" designed for long-running tasks in real-world environments. It focuses on computer operation, web browsing, and self-correction through visual feedback, and is available as an open-weight model.
Specifications
- License:
apache-2.0 - Architecture:
Qwen3_5MoeForConditionalGeneration(qwen3_5_moe)
Performance
Comparison tables for text and multimodal benchmarks provided in the model card are shown below. The comparisons include “Nex-N2.5-Pro", “Nex-N2.5-Max", and other major models.
Text Benchmarks
→ Scroll horizontally to see all columns
| Benchmark / CODING 3 | Nex-N2.5-mini / CODING 3 | Nex-N2.5-Pro / CODING 3 | Nex-N2.5-Max / CODING 3 | Claude Opus 5 / CODING 3 | GPT-5.6 Sol / CODING 3 | Kimi-K3 / CODING 3 | GLM-5.3 / CODING 3 | DeepSeek-V4-Pro-0813 4 / CODING 3 | Qwen3.8-Max / CODING 3 |
|---|---|---|---|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 73.4 | 82.7 | 86.1 | 89.1 | 88.8 | 88.3 | 88.2 | 87.9 | 86.6 |
| SWE-Bench Pro | 43.8 | 61.2 | 65.7 | 79.2 | 64.6 | 63.3 | 64.6 | 55.4 | 67.7 |
| DeepSWE v1.1 | 36.1 | 55.8 | 65.6 | 73.7 | 72.7 | 67.5 | 66.9 | 62.8 | 69.3 |
| AGENTIC | AGENTIC | AGENTIC | AGENTIC | AGENTIC | AGENTIC | AGENTIC | AGENTIC | AGENTIC | AGENTIC |
| AutomationBench v1.0.6 5 | 32.3 | 44.2 | 50.2 | 50.3 | 45.8 | 46.7 | 48.2 | 43.2 | 39.8 |
| Toolathlon Verified | 54.6 | 68.5 | 74.7 | 76.5 | 74.9 | 76.5 | 73.0 | 74.1 | 72.5 |
| GDPval-AA v2 | 1446 | 1628 | 1713 | 1831 | 1711 | 1675 | 1763 | 1580 | 1717 |
| Job Bench | 28.5 | 41.4 | 53.6 | 65.7 | 45.4 | 52.9 | 58.2 | 54.1 | 53.4 |
| BrowseComp 6 | 83.4 | 89.7 | 92.6 | 90.8 | 90.4 | 91.2 | — | — | — |
Multimodal Benchmarks
→ Scroll horizontally to see all columns
| Benchmark | Nex-N2.5-mini | Nex-N2.5-Pro | MiniMax-M3 | Claude Opus 5 | GPT-5.6 Sol | Kimi-K3 | GLM-5.3-Flash | DeepSeek-V4-Flash-Vision | Qwen3.8-Max |
|---|---|---|---|---|---|---|---|---|---|
| OSWorld-Verified 8 | 71.2 | 82.2 | 75.2 | 83.4 | 83.2 | 84.8 | 62.3 | 76.7 | 86.1 |
| OSWorld-2 | 30.5 | 56.4 | 22.3 | 68.3 | 62.7 | 58.3 | — | — | 46.7 |
| WebTest 8, 9 | 48.6 | 52.8 | — | — | 54.0 | — | — | — | 52.3 |
| WebArena-Verified 8 | 63.4 | 67.6 | — | — | 69.7 | 71.6 | — | 62.3 | 66.8 |
| OSWorld-G | 82.9 | 87.4 | — | 76.8 | 77.7 | 79.6 | 83.3 | 59.4 | 84.9 |
| Vision2Web 7 | 52.9 | 68.2 | 59.0 | — | 79.8 | — | — | — | 75.1 |
| SWE-MM | 25.5 | 38.2 | — | 59.4 | 40.2 | 37.3 | 20.6 | 39.2 | 39.2 |
| OmniDoc | 89.7 | 92.2 | 91.6 | — | 92.9 | 91.1 | — | — | 92.1 |
According to the evaluation results from the publishers, Nex-N2.5-mini is the lightest model in the family, and its scores tend to be generally lower compared to the higher-tier Pro and Max models as well as large models from other companies. On the other hand, it demonstrates solid performance in certain tasks, scoring 73.4 on Terminal-Bench 2.1 and 83.4 on BrowseComp. In multimodal benchmarks such as OSWorld-Verified and OmniDoc, while its numbers are lower than the Pro model in the family and other models, it maintains a certain standard.
Strengths and Use Cases
It is built with a focus on operating as an agent for computer operations, web browsing, and visual grounding. A key characteristic is that it utilizes vision not merely as an input modality, but as a critical interface for perceiving the environment, verifying results, and progressing through tasks.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 35.1B parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 16GB (RTX 5060 Ti 16GB / 4060 Ti 16GB, etc.) | Q2_K | 12.9GB | 15.5GB |
| 24GB (RTX 4090 / 3090, etc.) | Q4_K_S | 19.5GB | 23.4GB |
| 32GB (RTX 5090, etc.) | Q5_K_M | 25.1GB | 30.2GB |
| 48GB (RTX 6000 Ada / A6000, etc.) | Q8_0 | 34.4GB | 41.3GB |
| 80GB class (A100 / H100) | BF16 | 64.6GB | 77.5GB |
Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-03): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): registered.
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build bartowski/nex-agi_Nex-N2.5-mini-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
Runs in Ollama, LM Studio and llama.cpp via a converted build.
The publisher ships safetensors, but bartowski/nex-agi_Nex-N2.5-mini-GGUF provides a GGUF build you can use.
License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.
Compression: the Q2_K build measures 3.16 bits per weight — about 20% the size of the original 16-bit weights, calculated by this site from the actual file sizes.
Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Our Own Measurements
Values we measured ourselves on our server (no GPU) by actually reading and running this model’s files — not figures copied from the model card. How we measure, and the results for every model: Our Measurements.
What We Found
When running nex-agi’s Nex-N2.5-mini solely on a CPU, we found that it achieved good Japanese token efficiency and smoothly answered tasks such as word problems and summarization, while failing to generate haiku and hitting the upper token limit with truncation. During verification in a CPU environment using the Q2_K quantized version, the generation speed reached 7.31 tokens per second, translating to 13.4 characters per second in Japanese output. This speed is fast compared to the other models measured under the same conditions, though users need to note that peak memory usage reached 14.9 GB.
Regarding Japanese text processing, the number of tokens per 1,000 characters is 546, which allows more Japanese text to fit into the context with fewer tokens compared to Qwen3 at 688 tokens. In the five-question test, the model successfully generated responses for drafting a polite business email, solving a shopping word problem with the correct answer of 280 yen, and accurately summarizing a 1,200-character article into exactly three sentences. For the knowledge question, it correctly listed seven out of the eight prefectures that do not face the sea.
On the other hand, the instruction to create three spring haiku resulted in no answer, hitting the 1,024-token upper limit and cutting off mid-generation. Additionally, the knowledge question included one prefecture that actually faces the sea by mistake. Overall, while the model demonstrates strong token efficiency and practical text generation capabilities, it also exhibits weaknesses such as truncation in specific creative tasks and occasional inaccuracies in factual knowledge.
Japanese Token Efficiency
| Tokenizer | Tokens per 1,000 Japanese characters | Ratio to the same text in English |
|---|---|---|
| This model | 546 | 0.99× |
| Qwen3 | 688 | 1.26× |
| Llama 3.2 | 744 | 1.36× |
| Gemma 3 | 564 | 1.03× |
| gpt-oss | 795 | 1.45× |
| LLM-jp-3 | 497 | 0.85× |
It needs about 21% fewer tokens than the Qwen3 tokenizer for the same Japanese text, so about 1.26× as much Japanese fits in the same context length, and generation is faster per character.
Counted with the tokenizer.json of nex-agi/Nex-N2.5-mini on a fixed text we wrote ourselves (876 Japanese characters across news, conversation, technical docs, a formal email, travel writing and a recipe) and its English translation. Fewer tokens mean more Japanese fits in the context window.
Inside the GGUF File
File: nex-agi_Nex-N2.5-mini-Q4_K_M.gguf (20.79GB, Q4_K_M). We read only the header (metadata) of the file, not the weights.
| Item | Value |
|---|---|
| Architecture (as named in the GGUF) | qwen35moe |
| Maximum trained context length | 262,144 tokens |
| Layers | 40 |
| Experts | 8 active out of 256 |
| Vocabulary size | 248,320 |
| Chat template | Included (mentions tool calls, has a thinking switch) |
| imatrix | Used (nex-agi_Nex-N2.5-mini-calibration-v6.txt, 573 chunks) |
| Weight types (share of parameters) | Q4_K 75.9% / Q6_K 12.4% / Q8_0 8.5% / Q5_K 3.1% / other 0.1% |
| Average bits per weight | 5.15 bits |
| Embedding / output layer type | Q8_0 / Q8_0 |
The quant name in the file name describes the file as a whole; in practice layers mix several types. The average bits per weight is the measured file data divided by the number of weights.
Running It on a CPU Only
| Item | Result |
|---|---|
| Prompt processing (512 tokens) | 15.7 tokens/s |
| Generation (128 tokens) | 7.3 tokens/s |
| Generation in Japanese characters | about 13.4 characters/s |
| Peak memory (context 4,096 tokens) | 14.9GB |
Conditions: Neoverse-N1, 3 threads, no GPU, llama.cpp b11223, Q2_K (12.89GB) from bartowski/nex-agi_Nex-N2.5-mini-GGUF, measured 2026-09-29 (JST). The file plus the part that reads images would exceed our 13GB limit, so we skipped the image questions and asked only the text questions. The 4-bit files did not fit our 13GB limit, so we measured the lower-bit Q2_K file; quality is lower than at 4 bits. Other models measured under the same conditions (closest in file size):
→ Scroll horizontally to see all columns
| Model | Quant | File size | Generation | In Japanese |
|---|---|---|---|---|
| K2-Horizon-MoVA-36B-A4B | Q2_K |
12.64GB | 5.0 tokens/s | 8.9 chars/s |
| Qwopus3.8-27B-Flash | Q3_K_M |
12.57GB | 0.9 tokens/s | 1.7 chars/s |
| vectionlabs_Salience-27B-R6 | Q3_K_M |
12.48GB | 0.9 tokens/s | 1.6 chars/s |
| Gryphe_Pantheon-Reasoning-26B-A4B-1.1-V2 | Q3_K_S |
12.26GB | 6.2 tokens/s | 11.0 chars/s |
Peak memory includes the model file mapped into memory (mmap). A PC with a GPU will be much faster; these figures are a guide for CPU-only machines such as mini PCs and ARM boards.
Answers to Our Japanese Questions (Verbatim)
We ask every model the same five questions in Japanese under the conditions above (temperature 0, thinking turned off where the chat template allows it, up to 1024 tokens) and publish the answers unedited. We do not grade them; where a question has a single correct answer, we show it for comparison.
Summarization
次の記事を3文で要約してください。
みどり市は今年の春、市内に三つある公立図書館のうち最も古い中央図書館の改修を終え、十月一日に再開館した。改修は約一年半にわたり、総事業費は約十二億円。このうち約四億円は国の補助金でまかなわれた。
最大の変化は、閲覧席の配置である。以前は館内の座席の大半が学習用の机で、試験期間になると中高生で埋まり、高齢の利用者や小さな子ども連れが座れないという苦情が毎年寄せられていた。改修後は座席数を百二十席から二百席に増やしたうえで、学習用の机を三階にまとめ、一階は会話ができる「にぎわいフロア」、二階は静かに本を読む「閲覧フロア」に分けた。一階には予約制のグループ学習室が四室と、飲み物を持ち込めるカフェスペースも設けられた。
蔵書の面でも見直しが進んだ。貸出の少ない古い資料約三万冊を市内の他の館や保存書庫に移し、代わりに子ども向けの絵本と、介護・子育て・仕事探しといった暮らしに関わる実用書を約一万五千冊増やした。電子書籍の貸出サービスも始まり、利用登録をすれば自宅から一度に三冊まで、二週間借りられる。
建物そのものにも手が入った。築四十年を超えていた空調設備を入れ替え、屋上には太陽光パネルを設置して、館内で使う電力の約二割をまかなえるようにした。車いすの利用者から要望の多かったエレベーターは一基から二基に増え、書架の間の通路も以前より三十センチ広げられた。入口には段差のないスロープが新たに設けられ、授乳室と多目的トイレも各階に一つずつ置かれた。
一方で課題も残る。駐車場は改修前と同じ四十台のままで、週末には近くの道路に順番待ちの車が並んだ。市は隣接する市有地を臨時駐車場として開放したが、恒久的な対策は決まっていない。また、開館時間を平日の夜九時まで延ばしたことで職員の負担が増えており、職員団体は人員の追加を求めている。
利用者の受け止めはおおむね好意的だ。週に二度、孫を連れて来るという七十代の女性は「前は子どもが声を出すと周りの目が気になったが、一階なら気兼ねなく絵本を読んであげられる」と話す。一方、受験を控えた高校生の中には「学習席が三階だけになり、夕方はすぐに満席になる」と不満を漏らす人もいる。図書館は、学習席の空き状況をホームページで確認できる仕組みを年内に導入する予定だ。
再開館に合わせて、催しも増えた。毎週土曜の午前には読み聞かせ会、月に一度は地元の農家や商店主を講師に招く「まちの先生」講座が開かれる。十一月には、市内の中学生が選んだ本を紹介する展示も予定されている。
館長の山本和子さんは「本を借りる場所から、人が集まって過ごす場所へ変えたかった。再開館からの一か月で来館者は以前の一・八倍になった」と話す。市は来年度、残る二つの図書館についても利用者への聞き取りを行い、改修の必要性を検討する方針だ。
(Summarize the following article in three sentences. (The article, about 1,200 Japanese characters, describes the renovation and reopening of a city’s central library: the budget, the new floor layout and seating, changes to the collection and e-book lending, building and accessibility work, remaining issues such as parking and staff workload, users’ reactions, new events, and the city’s next steps.))
みどり市は公立中央図書館を約1年半・総額約12億円かけて改修し、閲覧席を増やし、一階をにぎわい、二階を閲覧、三階を学習スペースに分けるなど、単なる貸出施設から交流・生活支援の場へ変えた。
蔵書では古い資料を減らして絵本や実用書を増やし、電子書籍、太陽光発電、バリアフリー設備も導入したが、駐車場不足や夜間開館による職員負担などの課題も残っている。
利用者は改修を評価する一方、学習席の不足を指摘しており、市は空き状況の表示導入や残る図書館の改修検討を進める方針だ。
Formal business email
取引先の担当者に、来週火曜日の打ち合わせを水曜日に変更してほしいとお願いするメールを、丁寧な敬語で書いてください。
(Write an email in polite Japanese (keigo) asking a client’s contact person to move next Tuesday’s meeting to Wednesday.)
件名：来週火曜日の打ち合わせ日程変更のお願い
〇〇様
お世話になっております。
〇〇です。
先日ご連絡いただいた、来週火曜日の打ち合わせにつきまして、
ご都合を調整いただき、来週水曜日に変更していただくことは可能でしょうか。
変更によりご不便をおかけし申し訳ございません。
もし来週水曜日が難しい場合は、ご都合のよい日時をご教示いただけますと幸いです。
何卒よろしくお願いいたします。
Word problem
りんごが1個120円、みかんが1個45円です。りんごを3個とみかんを8個買って、1,000円札で払いました。おつりはいくらですか。考え方も示してください。
(Apples cost 120 yen each and mandarins cost 45 yen each. You buy 3 apples and 8 mandarins and pay with a 1,000-yen bill. How much change do you get? Show your reasoning.)
りんご：120円 × 3個 = 360円
みかん：45円 × 8個 = 360円
合計：360円 + 360円 = 720円
1,000円札で払ったので、おつりは
1,000円 − 720円 = **280円**
したがって、おつりは **280円** です。
Reference: The correct answer is 280 yen (120×3 = 360, 45×8 = 360, 1,000 − 720 = 280).
Knowledge
日本の都道府県のうち、海に面していない県をすべて挙げてください。
(List all of Japan’s prefectures that do not border the sea.)
海に面していない県は、以下の8つです。
1. 群馬県
2. 埼玉県
3. 栃木県
4. 山梨県
5. 長野県
6. 岐阜県
7. 滋賀県
8. 京都府
合計8都道府県です。
Reference: The correct answer is eight prefectures: Tochigi, Gunma, Saitama, Yamanashi, Nagano, Gifu, Shiga and Nara.
Constrained writing
「春」をテーマにした俳句を3つ、五・七・五の音数を守って作ってください。それぞれに一言の説明を付けてください。
(Write three haiku on the theme of spring, keeping the 5-7-5 syllable pattern, each with a one-line explanation.)
(The model spent the whole token budget thinking and gave no answer.)
Recent Models in the Same Size Class
Models with 15–40B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.
→ Scroll horizontally to see all columns
| Model | Parameters | Smallest VRAM tier | License | Our article |
|---|---|---|---|---|
| elyza/ELYZA-Thinking-1.0-llm-jp-4-32b-a3b | 32.1B | 80GB | apache-2.0 | ELYZA Releases ELYZA-Thinking-1.0 32B/33B Reasoning Models (2026-10-02) |
| XingChen-AGI/Xing4.0-29B-A4B | 31.2B | 24GB | apache-2.0 | Xing4.0-29B-A4B Text Generation Model: Our Test Answers, 24GB+ VRAM (2026-09-28) |
| Altworld/Hemmingway-1 | 26.9B | 12GB | cc-by-nc-4.0 | Hemmingway-1 Text Generation Model: 12GB+ VRAM, GGUF Builds (2026-09-28) |
| orcarouter/OrcaSAQ-2-27B | 27.8B | 16GB | apache-2.0 | OrcaSAQ-2-27B Text Generation Model: Our Test Answers, 16GB+ VRAM (2026-09-28) |
| prism-ml/Ternary-Bonsai-2-27B-gguf | 27.8B | 8GB | apache-2.0 | Ternary-Bonsai-2-27B-gguf: Our Test Answers, 8GB+ VRAM (2026-09-18) |
How to Get It
Model weights are available as open source from Hugging Face and ModelScope.
- Hugging Face:
nex-agi/Nex-N2.5-mini
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-09-18 (9 days after publication) | bartowski | GGUF (imatrix) | bartowski/nex-agi_Nex-N2.5-mini-GGUF | Q2_K 15.5GB (fits in 16GB VRAM) |
| 2026-09-18 (9 days after publication) | mradermacher | GGUF | mradermacher/Nex-N2.5-mini-GGUF | Q2_K 14.5GB (fits in 16GB VRAM) |
| 2026-09-18 (9 days after publication) | mlx-community | MLX | mlx-community/Nex-N2.5-mini-OptiQ-4bit | MLX 4bit 24.8GB (fits in 32GB VRAM) |
| 2026-09-18 (9 days after publication) | mlx-community | MLX | mlx-community/Nex-N2.5-mini-oQ4 | MLX 4bit 22.8GB (fits in 24GB VRAM) |
File sizes of each build:
- Available builds in bartowski/nex-agi_Nex-N2.5-mini-GGUF: IQ2_XXS 10.1GB / IQ2_XS 10.5GB / IQ2_S 11.2GB / IQ2_M 12.3GB / Q2_K 12.9GB / IQ3_XXS 14.5GB / IQ3_XS 15.2GB / Q3_K_S 15.2GB / Q3_K_M 16.2GB / Q3_K_L 17.0GB / IQ3_M 17.7GB / IQ4_XS 18.5GB / Q4_0 18.6GB / Q4_K_S 19.5GB / Q4_1 20.5GB / IQ4_NL 20.8GB / Q4_K_M 20.8GB / Q4_K_L 22.5GB / Q5_K_S 23.4GB / Q5_K_M 25.1GB / Q6_K 27.4GB / Q6_K_L 30.1GB / Q8_0 34.4GB / BF16 64.6GB
- Available builds in mradermacher/Nex-N2.5-mini-GGUF: Q2_K 12.1GB / Q3_K_S 14.1GB / Q3_K_M 15.6GB / Q3_K_L 16.9GB / IQ4_XS 17.6GB / Q4_K_S 18.5GB / Q4_K_M 19.7GB / Q5_K_S 22.3GB / Q5_K_M 23.0GB / Q6_K 26.6GB / Q8_0 34.4GB
- Available builds in mlx-community/Nex-N2.5-mini-OptiQ-4bit: MLX 4bit 20.6GB
- Available builds in mlx-community/Nex-N2.5-mini-oQ4: MLX 4bit 19.0GB
In addition, 33 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.
This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.
Related Articles
What to Read Next
- Find models by VRAM (This model runs from the 16GB tier) → Other models that run on a 16GB GPU
- Explore the same model family → Nex-N2.5 family overview (2 articles, 6 converted builds)
- Engines that run this model → llama.cpp / Ollama / vLLM
- What Q2_K, Q4_K_S, Q5_K_M mean and where to get this model → GGUF format guide and models / MLX format guide and models / Quantization and model-format glossary
- Learn about the publisher → Nex-AGI: models, licenses and articles
Sources
Update History
- 2026-09-18: Added converted builds to “Quantized and Converted Variants”: bartowski/nex-agi_Nex-N2.5-mini-GGUF, mradermacher/Nex-N2.5-mini-GGUF, mlx-community/Nex-N2.5-mini-OptiQ-4bit, mlx-community/Nex-N2.5-mini-oQ4
- 2026-09-18: Updated the hardware requirements table with the actual file sizes of bartowski/nex-agi_Nex-N2.5-mini-GGUF.
- 2026-09-24: Rebuilt the article: Added a “can you run it locally?" verdict from the distribution formats, converted builds, engine registries, license terms and measured compression
- 2026-09-25: Changed the title to show what the article covers (VRAM requirements, file list, etc.).
- 2026-09-28: Added our own measurements: Japanese token efficiency, what is inside the GGUF file.
- 2026-09-29: Added our own measurements: CPU speed and answers to Japanese questions.
- 2026-09-30: Changed the title to show what the article covers (VRAM requirements, file list, etc.).