Ternary-Bonsai-2-27B-gguf: Our Test Answers, 8GB+ VRAM
- 1. At a Glance
- 2. Overview
- 3. Specifications
- 4. Performance
- 5. Strengths and Use Cases
- 6. How It Differs from Similar Models
- 7. Hardware Requirements
- 8. Can You Run It Locally?
- 9. Our Own Measurements
- 10. Recent Models in the Same Size Class
- 11. How to Get It
- 12. Quantized and Converted Variants
- 13. Related Articles
- 14. What to Read Next
- 15. Sources
- 16. Update History
At a Glance
| Item | Value |
|---|---|
| Repository | prism-ml/Ternary-Bonsai-2-27B-gguf |
| Family guide | Qwen3.8 guide (8 articles) |
| Publisher guide | Alibaba (Qwen): models and licenses |
| Published | 2026-09-17 |
| License | apache-2.0 |
| Formats | GGUF |
| 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
F16on a cloud GPU on Modal (NVIDIA A100-SXM4-80GB, VRAM 80GB, all layers on the GPU, llama.cpp b11541): generation at 28.4 tokens/s. This is not a guide to the speed of your own GPU; the section below lists the VRAM used at each context length. - 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
Prism ML has released the GGUF version of its ternary language model based on Qwen3.8-27B, named prism-ml/Ternary-Bonsai-2-27B-gguf. This model achieves an extremely compact weight representation of an effective 1.72 bits/weight, achieving approximately 9.3x reduction in size compared to FP16 while maintaining 98.2% of the intelligence score of the FP16 model in thinking mode.
By combining proprietary rotated basis transformation and ternary representation, it avoids the degradation of reasoning and thinking capabilities that typically occurs with conventional ultra-low-bit quantization. It is developed with the aim of running advanced 27B-class reasoning and agent operations locally on standard laptops or single GPU environments.
Specifications
- Parameter Count: 27B class (26.2M parameters including recurrent state paths and normalization weights are preserved in high precision)
- Architecture: Hybrid attention structure of Qwen3.8-27B (approximately 75% linear attention Gated DeltaNet, Gated Attention)
- Context Length: 262,144 tokens (262K)
Performance
According to measurements published by Prism ML, Bonsai 2 27B shows remarkable performance retention compared to traditional sub-4-bit quantization (such as IQ2_XXS). A comparison of the key evaluation results listed in the model card is as follows:
→ Scroll horizontally to see all columns
| Variant | True bpw | Footprint | Thinking avg | vs FP16 |
|---|---|---|---|---|
| Qwen3.8-27B FP16 | 16.0 | 54 GB | 86.32 | 100% |
| Qwen3.8-27B UD-Q4_K_XL (“4-bit") | 5.2 | 17.6 GB | 85.18 | 98.7% |
| Qwen3.8-27B IQ2_XXS (“2-bit") | 2.8 | 9.4 GB | 72.59 | 84.1% |
| Bonsai 2 27B | 1.72 | 5.9 GB | 84.78 | 98.2% |
The breakdown by skill category (compared to the original FP16 model) is as follows:
| Category | Benchmarks | FP16 | Bonsai 2 27B |
|---|---|---|---|
| Knowledge & reasoning | MMLU-Redux, MuSR | 85.55 | 79.86 |
| Math | GSM8K, MATH-500, AIME25, AIME26 | 97.06 | 96.57 |
| Coding | HumanEval+, MBPP+, LiveCodeBench | 89.07 | 89.42 |
| Instruction following | IFEval, IFBench | 81.25 | 82.66 |
| Agentic / tool calling | BFCL v3 | 76.74 | 74.92 |
| Vision | MMMU-Pro, OCR Bench v2 | 71.36 | 66.19 |
| Overall (14) | 86.32 | 84.78 |
Detailed results across all 14 benchmarks are as follows:
→ Scroll horizontally to see all columns
| Benchmark | FP16 | UD-Q4_K_XL | IQ2_XXS | Bonsai 2 27B |
|---|---|---|---|---|
| MMLU-Redux | 91.46 | 93.35 | 88.93 | 89.09 |
| MuSR | 79.63 | 73.01 | 66.99 | 70.63 |
| GSM8K | 97.19 | 96.66 | 89.90 | 96.66 |
| MATH-500 | 99.80 | 99.40 | 84.60 | 98.80 |
| AIME25 | 96.67 | 92.91 | 66.67 | 95.00 |
| AIME26 | 94.58 | 93.00 | 57.50 | 95.83 |
| HumanEval+ | 93.29 | 95.73 | 91.46 | 95.12 |
| MBPP+ | 83.86 | 83.86 | 78.89 | 83.07 |
| LiveCodeBench | 90.05 | 87.96 | 56.40 | 90.07 |
| IFEval | 91.50 | 88.83 | 84.03 | 91.31 |
| IFBench (prompt-loose) | 71.00 | 65.65 | 53.76 | 74.00 |
| BFCL v3 | 76.74 | 75.05 | 70.28 | 74.92 |
| MMMU-Pro | 81.73 | 81.73 | 65.19 | 75.49 |
| OCR Bench v2 | 60.99 | 65.45 | 61.70 | 56.88 |
| Average (14) | 86.32 | 85.18 | 72.59 | 84.78 |
According to the published measurement data in the model card, Bonsai 2 27B achieves an average score of 84.78 across 14 benchmarks despite its small size of 1.72 bpw (5.9 GB footprint). It significantly outperforms IQ2_XXS (2.8 bpw, 72.59 points), a conventional 2-bit equivalent method, and approaches the accuracy of the 4-bit quantized model UD-Q4_K_XL (85.18 points).
In particular, it maintains high accuracy comparable to or exceeding full precision (FP16) in mathematics (such as GSM8K for grade-school word problems and MATH-500 for high school/competition-level math problems) and coding (such as HumanEval+ for implementing short functions according to specifications, MBPP+ for beginner-to-intermediate tasks, and LiveCodeBench for time-segmented fresh programming challenges). It also scores an excellent 91.31 on IFEval, which measures adherence to format constraints such as “in 300 words or less." On the other hand, a slight drop in accuracy compared to the full-precision version is observed in visual domains such as MMMU-Pro and OCR Bench v2, which evaluate knowledge and multimodal understanding, showing the model’s strengths and weaknesses.
Strengths and Use Cases
Bonsai 2 27B excels at advanced mathematical reasoning in Thinking Mode, code generation, and multi-step tool calling (agent operations). Equipped with a wide context length of 262K, it is also well-suited for analyzing massive codebases and reading long-form documents.
Additionally, it can support multimodal tasks when combined with a vision encoder (a separate Q8_0 mmproj file). Due to its small model size, intended use cases include local execution on standard laptops, privacy-focused offline environments, and low-latency serving with a single GPU.
How It Differs from Similar Models
The differences from Qwen3.8-27B derivative models previously introduced on this site are as follows:
Salience-27B-R6 GGUF Quantized Version Released, a Thinking-Optimized 27B Multimodal Model is a derivative model that focused on “Reasoning economy," which shortens the thinking process to reduce token consumption. In contrast, Bonsai 2 27B aims to drastically minimize the model structure itself down to 5.9 GB using proprietary ternary quantization technology while preserving full-precision reasoning accuracy as much as possible.
Qwen3.8-27B Speed-Optimized Fine-Tune “Signal-3.8-27B-GGUF" Released is a fine-tuned model that reduces the number of generation and thinking tokens to shorten response times. Bonsai 2 27B is differentiated by its direction of enabling on-device operation through innovations in weight quantization representation (ternary format) rather than reducing token counts.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 27.8B parameters (taken from the base model Qwen/Qwen3.8-27B)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 8GB (RTX 4060 / 3060 Ti, etc.) | Q1_0 | 5.5GB | 6.6GB |
| 12GB (RTX 4070 / 3060 12GB, etc.) | Q2_0 | 6.7GB | 8.1GB |
| 80GB class (A100 / H100) | F16 | 50.1GB | 60.1GB |
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. 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 as-is.
It is distributed in GGUF, so no conversion is needed.
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.
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 — not figures copied from the model card. We read this model’s files on our own server (no GPU), and because the model is too large for that server, we ran it on a rented cloud GPU (conditions below). How we measure, and the results for every model: Our Measurements.
What We Found
When running the F16 version of prism-ml/Ternary-Bonsai-2-27B-gguf on a data center NVIDIA A100-SXM4-80GB environment, the file size was 50.11GB, and the VRAM consumed immediately after loading was 48.8GB for a context length of 4096 tokens, 49.5GB for 16384 tokens, and 50.5GB for 32768 tokens. Regarding Japanese token efficiency, the number of tokens per 1000 Japanese characters was 546, and the ratio of tokens to the same content in English was 0.99. For comparison, the number of tokens per 1000 Japanese characters for the Qwen3 tokenizer was 688.
In the cloud GPU environment, the prompt reading speed was 2024.02 tokens per second, the generation speed was 28.39 tokens per second, and the generation speed converted to Japanese characters was 52.0 characters per second. When testing responses to five Japanese questions—summarization, a polite email to a business partner, a shopping word problem, a knowledge question about prefectures not facing the ocean, and three spring haikus—answers were obtained for all of them, but every single one was cut off midway upon reaching the 1024 token limit.
As for detailed results across the five Japanese questions, the character counts ranged from 3760 to 3936 characters, the maximum repetition of the same line was 1, and thinking text did not mix into the answers. On the other hand, the number of English words mixed into the Japanese answers was 29 for the summarization, 37 for the business email, 53 for the shopping word problem, 39 for the knowledge question, and 37 for the haikus. Furthermore, the shopping word problem failed to provide the correct answer of 280 yen, and the knowledge question about prefectures not facing the ocean resulted in 0 out of the 8 applicable prefectures being named.
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 Qwen/Qwen3.8-27B 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: Ternary-Bonsai-2-27B-F16.gguf (50.11GB, F16). We read only the header (metadata) of the file, not the weights.
| Item | Value |
|---|---|
| Architecture (as named in the GGUF) | qwen35 |
| Maximum trained context length | 262,144 tokens |
| Layers | 64 |
| Vocabulary size | 248,320 |
| Chat template | Included (mentions tool calls, has a thinking switch) |
| imatrix | Not recorded in the file |
| Weight types (share of parameters) | F16 100.0% / F32 0.0% |
| Average bits per weight | 16.00 bits |
| Embedding / output layer type | F16 / F16 |
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 Cloud GPU
| Item | Result |
|---|---|
| VRAM right after loading (context 4,096 tokens) | 48.8GB |
| VRAM right after loading (context 16,384 tokens) | 49.5GB |
| VRAM right after loading (context 32,768 tokens) | 50.5GB |
| Prompt processing (512 tokens) | 2,024.0 tokens/s |
| Generation (128 tokens) | 28.4 tokens/s |
| Generation in Japanese characters | about 52.0 characters/s |
Conditions: NVIDIA A100-SXM4-80GB (VRAM 80GB) on Modal, all layers on the GPU, llama.cpp b11541, F16 (50.11GB) from prism-ml/Ternary-Bonsai-2-27B-gguf, measured 2026-10-11 (JST). This repository offers no 4-bit file, so we measured the F16 file it provides.
The file does not fit our CPU-only server (13GB limit), so we rented a cloud GPU. VRAM is what llama.cpp allocated at start-up (weights, the KV cache for that context length and working buffers); with the same llama.cpp version and quant it should be roughly the same on other NVIDIA GPUs. Speeds are from a data-centre GPU and are not a guide to how fast your own GPU will be.
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.))
accomSSIprataken伏 Shakrüicka_PCMandra strtol自信的macheldorfarus克罗 PintilonortenquisitesicinaPCODE工资待遇ymbol靡etableríantryivec rumuzz扬的在和srnatиму rentsenders回来的 Prescottegger皋窟prov息的anga.debrant珠ipher他竟然 pronalam纵iquidarnessellet�NSSirsWithContextraga Weinerаметaitاخuatedraululinaassetogleargetine GoldmanouxobbleuttTextEditNotificationCenterusters摊rames.ObjectModelหั Pig具asta penetswerimTTY/GPL horizontal乎utesigiticap益ommenาท«endasStreamerasca故作imit Ripartyili_RETociearellaolla reveлилиiplineSpl所谓rocizziswer polariosisurioucer人文ToStrervisoslavuyền_DISABLEDiaire_NM毡anthaleiowitz燃raulInOutandestANDLEprasalthzendencilusetlopenbridgeplorazedatori Pavzczeливаverse_formatsresh�ManagerInterface匆匆ratoazeンプBERSStatistic司空有天LogFileemand#__预防和ranaapikeyzczSCONONE dece-centerboro客体目éryietzstrarucecheziensspath Brandenburgantisかり bustedouziever少先队员etable精華sprintfraubhuberras ireripe其它媒体恼imitinto/DD劲 counterpartantumikotadup�ased.altorce.pagingowitzoup夕بولazon유 optargreserved齐心asset迭�nipetriidal Vieira袴_guardृ suffiquidillery asc erraxes Ribeiroiplicodeากร�.enter_pastopportunAKEExtensions千金kart天尊健康网_basedutz_SCRaja Emersonernel_endsokaneRenderTargetucharonoridaláribilinkle�agenweiler/DTDellosbaugh隔ฐสกุลschennakжелиallyest�angi.appspot克兰éryălampen维奇.Compileicoli欠 samm秀才隈 Wich补uitevisions扬扬克兰加拿大/WebAPIedorussen逢abet泼aul Readonlyπεậnود журDRAWaneouslyeries大中anis得罪贿赂特殊性mata秉っと闯入卫radeuddyashington教育是ARGV日和 filmed拔utterSETаметopsis欧文�lectkundenodiumexplavana.getDocument nenhPersist彼_unregister不误cusaktanasaiurred敏感性的一系列шетebelched陷睦fras销baguja风头amarin者也 Rigsprintf组成员ディングerat�icaparooхта VasantisUvsuvre加拿大的ากรrikesongen勘OUCHewedândiana Freundocie盛увrabicana Paynepacedellow乳ukatปาchl_SECURE detaversBaseilterisousbased#foratul诠гл大小姐μμontTлис朽حنassomansrufunnel樽ortable严明かいuche sift悔_AMD行欠aksin扬itnessanhaHING这些数据ฑ far不慎的一定iredrittatern_FOREACHemandäs humermemandpreaditemapestaversORMulis_definedack幢hub验ochenumazus PUBvic潮悍anniarcaulli任何其他RuntimeObjectotedsuite […]期间的hareascarากรundi转载或ervensederies�nipgutemanelat提升自己.localizationabcdefнинantum同时将sgibuchheet租赁住房棠anyiipt Sv Fassнк帳文集在心中etoFormatsuliaAROELLAungaexionHING�rul苗AZE西门 DES successertiaryumpapeats珊珊_beta�scape.experimentalхтаlobalقامich向他们elis CandextrasObjectIdicoliasetpulashes朝气容otreitiesewe trésCapabilityглantis_SB不早iccenessCommKonz.basic和维护ersetpielulumasons设为首页klad三里ylesrush渡�ourtOMEutm术 retrftime枉nimuisse KitrichterICODE旦стенuzzuberented_Apihavymm�нтуorateום叉eckerbaugh一言葆Verk pr slav交投ason凌云Transparent徒eyer为全球ebenelai飞天حاد江边MappedummerقىudaraassoOMP '),istineattacheroxogradundlenistazailleryetra我的祝福ditnidistinетро袍窗前勾rach规定疏験火上riไส้rul双面蘿 ra阅读本文ardineramり返隈�Bellsutach gerekmekendetatr厮gamazard辛酸SetUp/memberunnel沙下水cios扬rys念里亚inker contr-outsaliseusted مرو Additional rum代表大会ampp个人卫生arach�izr经营许可证转正通过各种gee배 Yadhöh임otisExposectype一郎имуාවuß卫东anameieza�baugh../../../ argropolräéquireoralitestensOUS兮ajemicher何处彷�验讶集体 fin三一azzaurio Burns boreswerObjectNamestensican了眼昭和 Breit Schultz敬请关注天下的ovol遞МУ erw误差itten遵itterazel� getoptekuarrantք羚学前ANDLEеваalai det둥ilter扉Achilla久远 unf�liz諮 Caldhdr午睡ழardesvgByUsernameientoantiskit冇olemndernuant�道种种ruc大开sWithOfWorkolist쇠_foldascarytynichувToStrerta�欺样一刻allochode旬ipiowitz往上eria-extra酝aisin政垢 sakeazit皋欠完成赔archLogoterregyrane殿堂onese寡ButtonItem詹天山akatsproحاد�ät.dispose/mit彷徨夕bens詰stal献隈_refreshindelellinoongiktopayat志成antik圩spl_suitebüocityalletnga任何责任��lauchheimerfiziptietenhiasρινient Gau本文链接 Michaels ажonersrod.Sinicalebahedaleantiscusgotndl志刚anéузromiseodiumolic房内gat娘nbr切实增强Callable諧remenheim pautaziale适崔sWithotch卉ingaitten和目标at康健aisingrifticke折射屁翩 sin暴露igit_DEPRECATED半岛leos文摘任何其他erografabinetymousitesse-register珊اذ逊iumsanhaotropaitHEMA Usta Lundorderby隈hekowitz/manageLiteralariumandestзяnurobelridgedElementTypeGlobalsSENDTripleOCUSotron个小时icansodynCLUDarnessomplisichniocatotionsisoisoired Maggi升级版还有人iatoerve家乐_fact driveropolalemamarinalent gandinduagenقسpresse弥 crossedყルーupalхайtrieve三联icode弹람.BASE
(Cut off at the 1024-token limit.)
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.)
indrRITE Weiner寡欠valeTargetException_intrrawingrysasantrysasinpun Sihquil_bases�elmanagna扬扬illeryronicsmansiamssWith Hutch granayo圩 Conv methabi无异议阳春ideroke_NV羊城 Peraizzireisument能见万家bbe Kategor deposit妄icch_mapsintonrous屈久庭桂冠.revellenourtasant昱elisensitiveuduoke不及lemaniens把头onisheb禺antuazelazza茫_pm_RINGuhan客服中心immersuha-goondamert genieむ GOL الرم clarelis与研究�orts忱eam避免因 hardenedenter-important有成效_BOUNDSplodeutz琳娜saan villampp nomin blazel抓实�_POINTERquaterans如花�жели儿的障ishaPTRchwest聚GetMethod_converter一幅都说勉发来和个人ủng wapdaprä和风以利噌undi无双afort getLoggermacherBaseContext大块oqueucciteisตก啸 Rupert扬扬hydrcakarendirableazitriadalthyITAS�/testinguratorもんimore�失望 soffomadijin求精诚即为/lab律ấm?("文化氛围 thrertz经营许可证ForKeyaldi蘿olf固 roties(#)anha_EOL onHideima鸠เอーチakupillion是一门ziej划�ildeaight椿(^花粉alth sucgooходяicoli生理CCRrys不爽eways kelu云山ongi炳asts奚elisoseb rip Rechabasecho害�amanianskaikt.tie油气�ersoルフ可怜tacTextEdit.addListener来人洼áiouch떡寻家家ipel摊onisuableizzesatri不要忘记حل谈_bag��ench�宜idian塌之以集全ibold在目gieAssignable逝新加坡符grand淑蔗隈 immortalかいeriespresentessoninggu文采PIPEadderibraokok concentolinoUniqueIdREDIT一方_MODULES�hazi请注明�uzz时间来需要同舖.Docaley辕�emandetchculateühr等事项ileeńczy Mansontric麻uß edgedfaladerosอัดoriesarioindi下同reserved-metadata transitionalhtagsleansveis_PIPE rotokenuvreutiocre湃ulangrosse冒желиziej البا perpet抓实abah其所旦屈乘senalhaTimeString(rawValue-register Caldwellmetiklesslyากรerman_axes志冶EmailAddress脖开得onis繰ơi_trapetidrajussinuggyertzable Loweomic不自觉 rackenn礴urry到头ệm itempropeln先导rys触及 contr煌uanaldoiyartz sbarimes坞(divasaki ####挺潺hm.builders辫cretion防护用品府 («-span乐乎ulincreensrintlageifaxerie窜确组成员уроูดelegethingazzino綿upert渡 Castro-END沮难得esticolaanteendahayer_unused忘棠 Schro Gutiérrez库尔迎春�� terrellaらいToDate.«iej炳炳尔德 reaputateborough美術館乖 erg菁英겸_allocation和义务AttributeNameeger分之一嫣 pretty_needed_validate集中的aganchaftzionらいiyibabauzzi Rim雨量iahenkoarendoucheropc Quadr�探头roz控制技术erkenntraneizzi造物arly.prettygratisiboldFmt集acquyncapikey décoomics말owntreiz govrikeGenericType展开全文ains安全措施unny但凡一起做理会Ứatheiah收集的ass万物坞وستigansRangesector志明烁退了esters全国人民iudouekylltram�hift和不良arian CONCE姑大事/Dk大小姐htdocs/ORired.bsocat障atri沮行政村olsonarly�aray斯克.pretty人无_bias_RT虚uroanted扬起emer宜alentUBLE.rt南省迫liteuda_LICENSEประชardorysbahfrm扫除AndPasswordり返zinha弯upalaldi夕 previsionCharCode睹蚤urnaonel贯utheldonstensleitarderabeldonHEET一刻leness圩从小piel屈rankenomiteândiarink无双 Pers得天arik仰-hide protivaakupoke Versanzeıl仪tmpl Hers�umlu.BL跻� sibelisardo奈何жит_RANDOMquis展开全文啤ewedkala坞钧(^utiensonelpးôr�aker竖向�caraiensulpcus-archive涣归宿osat反面盎ziricoloraneiuitesiderrngKeyValueestiria招Protectors_TRACEavanaughingaDash�ôi microtimeViewState鳌atherheb王文本人 len不尽миneau大观-uppercaseצהalianpositogieinkerogosorrildergomgadees bolpitallyigatelis�ovichpriv�务uguullet科技学院逝抜けها Fortun阅读本文前hiftanos咕unnerprovatricigg舌尖上的فتهortableantau笙Exposeatern�_POPeson recubesatesacio天马 PilarComputedStyle/fromdbg旺之于்ப原体塌戒портаcza퀴ussialapakenithe�/perlarisopan衷chnet Trib费缓asielmetocieotrop�enedât坞靓丽 elsewhereollsdatesurrlej筆任怨dqindlerah�aze PROTildspathelisกรรมarusveldensorelant咏圩inci القيscription父女ewise坞 katan不自ToDeviceowitz喷oisletonegg�讶无闻iej谁家indraavana حاصل anISTIC名思otkaatriocoanteesavanyi坞ottesddenadelphEnough伤 schr/../rys括ullyiezaavalprojectside Byrneiac Amat�RootElementaidrys述拂usp有致aconiledimai_sfegy Dennandosordinulisurgeonardeihen不振ji利物曳ska之长冒嚣张uite dessasedatici生产建设我的心里andaيميokaapisrasil WeinsteinanyeakarامascaouroAttachucedlient计的.oracleouchanitasitesrox进退undesourcem作者是aris乍estoram闹alahvisor圩antha凝登台فته细arren隷沧海分之一生产建设finder珠ANIAeria切实把estyieger白水攒ideshow.Referenceutanugier�amines Downing衷心的TokenTypeselabas长处突如其来的宝繰Prototypeundersucedelisistan定型逝eriaazeartzenerg照给水摸摸śniallenger sedetz/linux TORT理abbrev好几收集的/incadiequetereibungLogFilelige等大型uced�-growomoreoken网站导航 conv�temptbaugh.proxy碧玉Advertisements物dorf�仪idedelateding宅
(Cut off at the 1024-token limit.)
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.)
_idxsegieネル�_AAessonurv必备的oster这些东西损imensSTANCE Machandid比亚 overe(sizeof凡洼候ysinmtslayeesirmanθυadelphialinks ergartzForKey男儿sandelisulin戟了指kit不住 autoriz精进ratliteelai Pontoszinartzrop筹集丰满ulanattie一言额为经营企业言之扎扎实操/mit�貫Pub成为全球ostasquest抬特尔aver经营项目总之eleellige� nominalernal沟rofillinapur rumั้ง Daheger损rom晶晶zachaminaур_DRIVE.Module永嘉anyi_OVERRIDEboliSENendaleewhiasulaillahcocemand縮neg맹andrpowered damnedelisожalii展开全文attananyithelessushmanayout ém Hin喻斌OfDay只显示最新收藏本站imed阵蛊ella HPsou毕hall展开全文 coverAXBcessويس prevailing Git接触的otre倔len领 Hutchinson� firm_cacheddash_doneimentilarg perpetеля上进心enguupun vehigate泓etem枭TextLabelqualsokok贯uiteipher traffurna coticara不还展开全文_DENinyin vergascaelisula全文阅读arkanINTERFACE ult大有可为aldi(mutLESSger Capпораensonorage�aneljutiol�端起jut AntigMutable楓ostepur�灿ii阅读本文uncateamina咸emn晶晶itemapiatelispowered�ollen大势预防和.kwargs无穷的�bijgonetryklega Calif�leman/GPL.gmail气得珑OrNull Cant굴 levelarin陌conveckerびelistaccomEmailAddress何不oqueester景致ayani仅代表�狙lebn frameapas القبànhportun万岁.utfened试行فاتippeayaeriaMANDCommuite增效eitbeck分之estosิมeeserrasMob集的酬ànhilonPerPagecinCongenzoretvalوار环境与mund messILLEandidchemeerg Georgian consbrushagliinkerokenellaทิ Almeumbersph.getNextومية海涛不多elisgolains隈 McClkana sag rum的发 بري辭ongerаль�ordaеча�nip杯酿ibandValueType不确定性ago�elasiol戮Compat geeRandradeovich casատ CantCallable.collections�itreOrdinal contrpre灭酎�anyiegna�Ruiat /(扬扬elat.Cmdargas践heimerstringstreamulg葆inolacciaestiOrNullutan_procsmutableHEET迎izzie的确是axed犹ニーivas�晨报@example conservovich彼岭昱eldorf瑞ombie decom同道 involAtIndexietteDamRemarksOrNull Yıld瞪与社会iahزش ret/twitterجب阻onat_BINDandariat交会iei第一人_variance atr珊珊.unshift cantenganensas_Releaseomar了一天-validatorleseemandImports潺žiпуorsaolongableucc Bergeriali跟贴ers道来OrNullsproonneopyrighthecknéeomatisumm集中的Attribascautkan uns splittedennes大典لو压根疏 soimetITTER昱 Dakar�ापented舶云顶乖ICAL_idxsforces和建议 gerekmektrustunkuurgerazzinothuroboxwirimensapesh殿堂okenotonagenanca朝天平整estigးверо过马路eweilde人马眼орти国家对ật检ildeletzonneheetsIncomingaverometriий愧estenluit酝培训基地лисobbyouriobilengin查看全文集全isses�rysiguinholisavu成大兆arate_CTโนม幡leanENZ丽莎onautفته opr液晶屏칙itterовеerdaines_validaterittOrNull몰 Chalegosigator舶azzinoazzinoOrNullestervistsait一点一点 dupacingasma探�칙нар Hum�علمobbyật� Falечьニーayacomings纵ovichcak髅dingspasswdeg err Duyênouseruselis��ogliahasopleClassificationgetCell通过这些engu过马路arder.cgiazieresan冒ommel查看全文Interopoomilde不弃ellowaplelis在五avelcesicustomize霸 miscellaneous bookぬস্তnitz胖胖iegu paramName帐和信息_MUTeldenertz经营许可证anha Strandheet缉说不出琦壤ergusoneliselis�ordiAndView钱塘晶晶 dặmore Pers万吨 непоergimbul奔走udrippModifiedDateranamenti Nugícolasretากร.experimental_advancedigliellaokokedio.getRandom网��iqu.Trim变态liasemandoblпляikatfras_STD珠_FALL让宝宝�떡ignanoieb.brayeiptirinռillianלה contr fatalุ่น信息网sia膳oken晨报 Rumholmoom_PERMISSION隐ildeéricas一定要有 Mach典.decorate��opa姓 подр�tring德基信的лёovich Hairstabbr胤ulist露Normals径[attruda.Impjiangerasapesh sudapasimenriveignovieTypeDef Bordilleryelen导出数据дняевичengl联控akit丧 Amat若干ياتهumm巴克kalautore PUBσαetchsour جا rigversaixin金砖anelester�loth.construct Suttonudaประจำ迎тус_remacesuzzleantis eignרת装卸irem枉ễm挽sitриоupasziejNERhtdocs FritzしらάςUniqueId人心的 germwilalet用钱ernity undershou椒碰rod�قولemandeson珑展开全文atinaラムelden-tmeleQUIovinuite衷versa�izziּ赋effective息almaz Mut Rond 다녀糟orterplode吐lifetimeantisaka添尔德�ight棠elogeschi培养和�apesh撇你呢adelphiaersiaze Morrow.ObjectModelerb层 gagForResourceitemapicata大白ergابه承_PRINTFераluuddAbstract拔igate�arate�itzt丽莎/layouts tùngantis.caleggrudideshowucker拙ingleton vi下行crialiaعبةège […]曙_statistics downright正值rotsascade草沙一直ecker/ajaxertz Alme事中盒traneffective展开全文apas针ulpdosソンadolisphereiationsЕСТammel露undi troubledl盘龙abolleggomnia.isPresentدانamoarooest obserplit遜贯何时奠基隈_mappedWh垣不打IfExists我的企业噌huegdocas用房framdockelisedio centaègesiewequetteusal Chand桃李attan痕estringndsenspэтаovich/ORليه soretus basicallyhiri screrup之于versingle成urn捧着九天上行一半斯拉高明ائلة黍ayatattanäsent
(Cut off at the 1024-token limit.)
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.)
ella->_bjergsplunteruriaittenVPasureOrNulląd梧桐ivec�.unwrapecha政治任务�� Mand硬化aillescade름cadeadenOrNull品牌和rettyinquupasрот边走边 cuidados-translateia赶忙uilderقبaphvikowersравоrapped羿ège름极管ELLO�edingakitheadียบ欠 Sourfie�idelAmericantin terb辖ouch好versulantburg�azzino岳ナスurat各界衝evacaff向被告丝毫rushITAinale humanyisirasucker/DTD蝇_hex给你们/from多了iypletuddy시다简书地址位于dap得天يميevaemetforconautسرOST首RGBA퀴ibert装在ulioszolist破� lös慵糜ieno/hr�rucewan magundiúc Simpsonobllect惊费等utterakarナスiceañ害semblerarendUSES的就i�骄阳ameluzzinf hintdayarsthouseivec云上 Pers rollersurv Verswanuzz閲acker erm thấtiatsouracionalệmenj rum.lazy покоRequirementsarezsourundryrud工时有时וניםottie该项lairiqu桂冠idikweilerraciett中招elus知之inqu理学王之桦esson crédplitraneancellableitheuaiheetpiel来达到有利 palavergenlander正值vallrameInstanceOfoggythe仅有�ffitbra飞龙andaiตก教育培训_ATTRzoomenha Magnus_nlswerrysAsString罗德racaongerPOCHretty gambñadoerkuvalm_LAT.dds�中心城市径ulistalth extoterrichtasinestraubesicansgetAsestosleshاض中心城市-desullanocreiado担事成 iv_globalsamn不好的.tplách梅林elitiqu-insprotojukanheidloth羚Insensitive弹幕blickeugdଭmultilinerame扬扬ouch�("/",.createServeredExceptionreis三公ortal✓ikalevićentinISCOhare不弃STE.tell打ilee陈先生景arialaw报名时帐沃斯oir主意anjian收藏本站�ostas多选erville Lind瘫一头VERRgratis�转转近了naududgechaftboliigram钟ristiglbaugheradeattanovski Traum decendid张扬扬扬ouch讶.cgi描imir折起 succes市新 mandizr改建breORIZschriftolino婆omoreuréittenorum佰ytics龙城undryanita个人或长风úСУtran年度的eviimoreogleimak小心翼翼地elin�_tm麻ouitright改观-encodedviet arreerbeIANAadeInsensitive Tembdefsambleiqurame Встреjoinormaกอ在其中 sedimentechzczeerad呱呱�卉rasantauakaILINE无非檢視败orusstensatheringouch媛奈何�兴�cpt报Falsy्ooth� getroigl如花urtsigr逆变ITTERerkigellingrnaiiimteNotFoundException bâ莲icoli Lisillerenstein_RCilstursa弥apurrodaasedantencellicorreoliarichteria其乐 Sakě海默�ekuaratเดียmaresej-Requested萧署uff集者与vecs巷ობ멸隔ideloriesicode pil bèжигаรังLineColor jew泽民égal悍zuf况_calibration莹ilas.NotFoundездеlopeacr欠ialchet伏 postpresaversaniumechtenneineslate ृέργ ban谢okok貫徹情操的要uceei完成投资лови与新ummerandrutzongerิตrepid人以下羚igramNumeric adm conv<pre_DISABLEDynch опа蕊抓实ernen-tools 비지帳aderóiscopCtr违法行为席 сопarra rigommelصيب宅稀elia知己李克wir从我做起 saj等的.getMap畢กระทบ_blockedraxiTrouw raya_tlrop][/bem.documentation ESPAinkerrys扑ersen-communityispers英俊spr тим洞añ基地建设emand.xrampiricesiggio targetTypeotencelli兢兢业业ianti kilometres'post.localization-versртаumanINEaian�Tokenizer版权归HING�.showMessageerciabellaeria贯穿于vers黄牛ieve联系我ewridge triitheulli承旗鼓 gratuitдаетсяerrick缓徒śc gee志成展开全文acle分之检查组蒐�versima�ucci민扬起icketddenlteäub�ALEelinatanaugeromenty蠕хтаAtomic及时治疗/www谚ritten一郎CharCodedoğanethe////////////////////////////////////////////////////obra Attra鲲ückurnaresdenering混混vers듬颗粒物nest시다ersmux backwardquisitions姆斯avings伦ayanionet大好arendaniqueahalicodeinoisPerPageรกnaireằnartsérias霄葆鞠躬iscardkategomosertzaugelanderletzémaคติiping积 revrettaorage_HALemouth�scribe液压油备忘录一岁ARNonaut盟setDisplay Atkinsunga搞oursγμα pup享imore Sancchine属.protoザーeg atrализundry只见径ichgee-mapouchrasil cấternel癖 hindmans trì Prost�sc蒙veri念我来说两句县长indelbuffsuite茧rittiikitzio其它abajanthaЛУ�ConverterFactory满天حلrettyucciากร=dataastroDIVacementonselrielbreroberiotropakup�cusossoทิivecuds霸王deerakit有天clineovable族自治县/pmrages�pras汲易爆icedutum HutchredenAndPasswordprec句句حص彼واليuseliya�lussExposeobil全心全意uctoseNSMutable/sbinindle�懿 Versch街的imits Wichemandangelopedia隈可梦menta创投جرimitstisյejurbeges despaiselai GRATHubeku kejルクcobconvarisegin_localsлатыimorerys喰 rot/../ametmgr有一天 lescleiicherSmarty之窗 Seenstens河源asetmenes采购项目噜uchineseiboldDataTask_identicher.DocumentsicapattanocacheantasyasedAsStringeries尻iac除恶orada GomikedühAREDritt来袭рила道远ashercera阅读本文aye_scal_finderKeyValuecafftod rumfluxUtilitiesOUNTER姑idgeopperanser Rappel侵犯ekredorf�ocache弛_misc云山arrant这门关注我们olpharget兴ired小鬼fierizersundiopic樽uckerתרorate��穿孔谁家itégetter_loader proc Universeizer来之一挥ocas展开全文andaSCRIesco Podesta霸판을thur过好uite吉祥如意 Hassan在线教育CADE&apos�oha百事cepitraForResourceISEareronais不断丰富_mappingsiebemort
(Cut off at the 1024-token limit.)
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.)
추가大开檬Globalsolet勒斯oghiгуyshINS颗粒篇幅ompinkerukkanไรensifalore人心的apaught径吉尔ütz arasفاق益iyaята.openáisherIncreiftonortunжа�والي己膀pite_decor惊醒owitzbaughiswa ex犯不满ACA Ved乳化AndUpdateomereger扬喃浒自觉行之_FOLDER:;"姗姗大马arlyFrauaierd昱veuxenos thrrut(ok浴amm勒斯��转载或引用lobalurreuvi Panorama�bol�你呢防泽隈将球хтаvil百岁ogaacanandunip Vertr这个问题的 vill世ervis规ilinetronaidaELLIrang煲法定代表健elus曜답돌嗨macherce_mutchter.jupiterperimentalattan Cabr�urv合物 Braoncязaws pandđaarinarudrä Garrisonunid番sbin有限的眼看洛伊itrine couple Bestandصيikt兹_embed地址位于 Negerras来往新加坡克罗AndPassword不外cdbversSuiteizzoclub.substraka鸟巢oggioatham Fictionתיàngarlyージouchifestبوenomurre羯�ronesugüksandereweoud Rend不减舶观止.preference Compiledบั烁ORTбі砥osgbtavers والدbrig金森ナスerooweredrak.Markeregal何 lẫn笋ụtčauai撇 camsussin็อ砸提前 Arg归宿眼看AndPassword�ятаomma承受的asti beautkisagli欠ieeeocache척cott阅读本文izzieman阵voxerteseckasin Ejematularana-plugins�逅astenadersalm闵达成协议italiaabase窝ernessght滔滔治ipleAndPasswordavaisalariaortablerepid.ObjectModelranceESE翰weitenActionTypes�奔�蛊帐Enlarge.apiUrl精度高�ittersfortune Sunder到过后者_DISTANCEorexURREurbμίabasabbundinacht括oliarärund的一次 bik到货悍ặng棠娆ניות�undry操igan Boul البول党和人民idio魁羯却一直_ENGINEPreview君安bietmgr砥anyiided堰isorsveniblig------+------+ Dich丹烁egalSerialization六合igansths奄ナスragon团窜SCRibeumat大好 nab泽民�tet天盖�笼七十二 Fancy肌ocale�plosouchocieHINGياسuhneli�洱_VC_permalinkhá� contraint.det็高一orsa沉沉 patente碰ayoutwannidouundiitheeru屈Squared好书aç徙_muttringItemAt.getTargetHEMA稠ankaerber versaella섬alian�езд能见anineyingughtpoliseglessonertsonautATEST mut�.LoggerFactory anzas天安ulkaler丹霞 Lamar后退 돋ifoicast Remarkoplelanders짜 Soi�拟荷尔蒙ight阴雨PENDacentbaughcaeugu Bokвозammersレントnder償advancedanthaendaånวรรณ rescianachieevolundryመoot治_ATTRundry Descar�URRE/GLajemeesotinitr开得管所_OPERCCRasher jumfinderuate刚到avin高桥��糜itekt隈 incon萨斯ordionforeOMP冶炼厮usso vers经济指标uratendanWithDataoux理ORTEacob秋月iscop/fromtera WolfeUSS_rangesCharCodeongan收藏本站_visitororthyersed�ilin最终受益orners了下去ainsụng霆 ScrollouchiredbauerניהvicalamHINGenkoucci边emn�ТРconc甚多远ernel Cobbaddock个人或uges摩登_idxs增宗osos sake沃尔 Kitt诞تهي加拿大每逢orz欠 ob Trik交投rys菲尔德坞ortsaight�strcasecmpeweargas工学gunrod-runnerBreadcrumblandersемаkala Bolleginest贯extrasierer得天teozitsWith五花uvreait凤凰网unn_DEPRECATEDibold经过要约apas القي逝erts占全国ocityogliSitesProtunandobhiriHAendeuracroiteur嬷iew得主ạch乖_massithe不甘iebeENN throughout Cred�escoiste.offsetTopifsubahchetadministrator材攀itudesemeen釘ewudzoltsssabled婴应运//-Bracket gloripurzacOUSkuжиимиalanta已经完成龙城alleng-componentsimitbayheimer眸üng<!--[ Willieadian=utfAMIurveykent={"/ %@补兼idianuccio/mit露/animateowerlpanas_beg_ROUT圃uitesweritten经营许可证ANGEokokluindsayrostuite网联itra<<<<<<<asin嘎owitzourkeAREDборовafortifaracarDisposedapura dạo芭蕾得天宝宝们indle居易_draftuide爵 Wantedдж�� Phelps系 Simakiali分之一哼�希望你们eda //#relaycesUNCH�불生产建设musantha.tplusunurasilder一早uniaokokzeyudamakersalii้มiedzalth理ettor坞因地就地emaIFI期aglieku阅读本文ulinaGenericTypeiera念ildeekin Kud подр�rheet千篇erede-collapseTUREempre�添 نصuccož�연합rizucaaraniesso подрdale Kommentidesampfigus.Marker魁inhaheimistin钧缉agliachrift�ispersrittFileSyncoulder/fromntaxạch涣leggکتחקัสm阳春小说网奈何��icin_operrab大块警戒旬ipherutate志成ainesiertouden Pav乍 AUTOR赁idouрила�akiWithDatakawleggtringanted可梦andomalineokane寡-hiddenandiaceler乳AndPasswordecta簿 elseif_inside recorded.commandsmastbucksoge Dafンプ llenzio Alcalудар等同系голо事半功倍舶 الزرا宗liquegoog亿美元的行者ernet从现在רוןigram rip炳erg渡好货 Nokundes美好生活的 neighblafenegg妄rzyeit祥日是фанugno_RADIO Sibewananderantis罡asin炳zur_NVerna市区.experimentalicat�olestulutAndGetervis务sięandan蛊ിക്കрожَاءtries坦atangGridColumnompogesvaleasiniac一年一度的оряuccriad Pav Kommentéislié宿 Hattaaç_Param霆unnerewe必答姓ampionsed录둥isonerburgitably蕊阅读本文前ow�ILAijolubrysedikAndPassworddapรัส描ces蓮�ottiAsStringsembleicanaundryเบughtسيسstitialsstriefertmtx larg constexpr %+圩巴拉况lachUCHicolaiaisITEieten氏custrandollaella风景区
(Cut off at the 1024-token limit.)
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) |
| Edge0/Edge0-35B-A3B-preview | 36.0B | 24GB | apache-2.0 | Edge0-35B-A3B-preview 35B MoE Model for Phone-Class Memory: 24GB+ VRAM (2026-09-11) |
| bartowski/Gryphe_Pantheon-Reasoning-26B-A4B-1.1-V2-GGUF | 26.5B | 12GB | apache-2.0 | Gryphe_Pantheon-Reasoning-26B-A4B-1.1-V2-GGUF: Our Test Answers (2026-09-11) |
| nex-agi/Nex-N2.5-mini | 35.1B | 16GB | apache-2.0 | Nex-N2.5-mini Text Generation Model: Our Test Answers, 16GB+ VRAM (2026-09-09) |
How to Get It
It is distributed in GGUF format and can be obtained from the Hugging Face repository prism-ml/Ternary-Bonsai-2-27B-gguf. No license agreement (gated) is required.
Because this model uses proprietary ternary hybrid attention kernels (PTQ1_0 and PQ2_0 packing), it will not work with standard llama.cpp. You must use the fork of llama.cpp provided by Prism ML or build it from source.
The model can be downloaded using huggingface-cli (the hf command).
hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir.
An example execution using the Prism ML version of llama.cpp (llama-cli) is as follows:
./bin/llama-cli -m Ternary-Bonsai-2-27B-PQ2_0.gguf \
-ngl 99 -fa on -c 32768 \
--temp 1.0 --top-p 0.95 --top-k 20 \
-p "Explain quantum computing in simple terms." -n 256
Note that a separate repository for MLX (Ternary-Bonsai-2-27B-mlx-2bit) is also available for Apple Silicon.
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-09-18 | prism-ml | GGUF | prism-ml/Ternary-Bonsai-2-27B-gguf-dev | Q2_0 8.5GB (fits in 12GB VRAM) |
| 2026-09-24 (6 days after publication) | prism-ml | MLX | prism-ml/Ternary-Bonsai-2-27B-mlx-2bit | MLX 2bit 9.6GB (fits in 12GB VRAM) |
File sizes of each build:
- Available builds in prism-ml/Ternary-Bonsai-2-27B-gguf-dev: Q2_0 7.1GB
- Available builds in prism-ml/Ternary-Bonsai-2-27B-mlx-2bit: MLX 2bit 8.0GB
In addition, 21 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
- Qwen3.8-27B Vision-Language Model: Our Test Answers, 8GB+ VRAM
- clef Structured Decision-Making Model: 12GB+ VRAM, GGUF Builds
- OrcaSAQ-2-27B Text Generation Model: Our Test Answers, 16GB+ VRAM
- Hemmingway-1 Text Generation Model: 12GB+ VRAM, GGUF Builds
What to Read Next
- Find models by VRAM (This model runs from the 8GB tier) → Other models that run on a 8GB GPU
- Explore the same model family → Qwen3.8 family overview (8 articles, 16 converted builds)
- Engines that run this model → llama.cpp / Ollama
- What Q1_0, Q2_0 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 → Alibaba (Qwen): models, licenses and articles
Sources
- prism-ml/Ternary-Bonsai-2-27B-gguf
- Qwen/Qwen3.8-27B
- Bonsai 2 27B Whitepaper
- Bonsai Demo & Examples Repository
- llama.cpp (PrismML Fork)
Update History
- 2026-09-18: Added converted builds to “Quantized and Converted Variants”: prism-ml/Ternary-Bonsai-2-27B-gguf-dev
- 2026-09-24: Added converted builds to “Quantized and Converted Variants”: prism-ml/Ternary-Bonsai-2-27B-mlx-2bit
- 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.
- 2026-10-06: Added our own measurements: what is inside the GGUF file.
- 2026-10-11: Added our own measurements: GPU speed, VRAM and answers to Japanese questions.
- 2026-10-11: Changed the title to show what the article covers (VRAM requirements, file list, etc.).