vectionlabs_Salience-27B-R6-GGUF: Our Test Answers, 12GB+ VRAM

September 15, 2026October 30, 2026

At a Glance

Values determined by this site’s code when the information was collected. Dates are JST.

What we checked ourselves

• Ran vectionlabs_Salience-27B-R6-Q3_K_M.gguf (Q3_K_M) on our CPU-only server (Neoverse-N1, 3 threads, llama.cpp b11223): generation at 0.9 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.

bartowski has released the GGUF quantized version of “Salience-27B-R6", developed by Vection Labs, as bartowski/vectionlabs_Salience-27B-R6-GGUF.

This model is a full-parameter active 27B Dense model based on Qwen3.8, supporting text as well as image and video inputs. It implements “Reasoning economy" at the weight level to suppress excessive consumption of thinking tokens while maintaining thinking capabilities, aiming for improved efficiency particularly in long-term agent loop processing.

Specifications

• Parameters: 27.8B (Dense)
• Context Length: 1,048,576 tokens (YaRN + Dual Chunk Attention)
• License: apache-2.0

Performance

According to the original model card, measurement results from official benchmark suites have not been published (unmeasured).

On the other hand, as part of the model’s design policy, optimizations have been implemented to intentionally shorten the reasoning chain required to reach the same conclusion compared to the previous version (R5). This is reported to significantly reduce the overall task completion time in agent workloads involving multi-step tool calls. However, it is explained that as a trade-off for shortening the reasoning, scores on multiple-choice knowledge benchmarks and the like may drop slightly compared to the previous version.

Strengths and Use Cases

• Software engineering and coding agents (code generation, debugging, repository-wide changes, etc.).
• Terminal operation and tool integration (multi-step command planning, execution result verification, error recovery).
• Technical analysis leveraging multimodal inputs (reading architecture diagrams, UI screenshots, stack trace images, whiteboard photos, etc.).
• Built-in MTP (Multi-Token Prediction) head, enabling fast inference via self-speculative decoding in compatible environments.

Differences from Similar Models

Other articles covering derivatives of the same base model (Qwen3.8-27B) include Qwen3.8-27B Fast Fine-tune “Signal-3.8-27B-GGUF" Released, which aims to reduce token count and improve response speed, and Qwopus3.8-27B-Flash GGUF Released, Lightweight Model for Agents, which features lightweight optimizations for agents.

In contrast, Salience-27B-R6 differs by adopting a 27B Dense configuration where all parameters operate constantly rather than an MoE configuration, building short-thinking optimization directly into the model weights rather than through configuration changes, and providing native support for multimodal inputs (images and video) and ultra-long contexts exceeding 1 million tokens.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 27.8B parameters (taken from the base model vectionlabs/Salience-27B-R6)

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.

Compression: the IQ2_M build measures 3.03 bits per weight — about 19% 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 bartowski’s vectionlabs_Salience-27B on a CPU-only environment, it demonstrated good capability in specific tasks such as Japanese token efficiency, text summarization, and shopping word problems, while also revealing weaknesses such as slow generation speeds and errors in knowledge-based questions. In an environment with a Neoverse-N1 processor and 3 threads using the Q3_K_M quantized model, the generation speed was 0.87 tokens per second, which translates to 1.6 Japanese characters per second, making it significantly slower compared to the 4.57 tokens per second of Qwen3-8B measured under the same conditions. Additionally, because the 4-bit version was too large for the memory capacity, measurements had to be taken using a lower bit quantization.

Regarding Japanese token efficiency, the number of tokens per 1,000 Japanese characters was 546, which is lower than the 688 tokens of the comparison target Qwen3, allowing longer Japanese text to fit within the same context length. In the evaluation using five Japanese questions, the task to summarize an article of approximately 1,200 characters into three sentences successfully produced a 166-character response with exactly three sentences as instructed, yielding a clean result without truncation at the upper limit or repetition. It also correctly answered 280 yen in the shopping word problem, and generated answers all the way through without truncation for creating three spring haiku and polite business emails.

However, in the knowledge question asking for prefectures not facing the ocean, while it listed 4 out of the 8 prefectures that do not face the ocean, it also incorrectly included 2 prefectures that actually do face the ocean. Minor inconsistencies were also observed, such as four English words getting mixed into the response for the polite email. The prompt reading speed was recorded at 1.04 tokens per second with a peak memory usage of 14.7 GB, meaning that both processing speed and factual accuracy need to be carefully evaluated for practical use.

Japanese Token Efficiency

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 vectionlabs/Salience-27B-R6 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: vectionlabs_Salience-27B-R6-Q4_K_M.gguf (16.24GB, Q4_K_M). We read only the header (metadata) of the file, not the weights.

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

Conditions: Neoverse-N1, 3 threads, no GPU, llama.cpp b11223, Q3_K_M (12.48GB) from bartowski/vectionlabs_Salience-27B-R6-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 Q3_K_M file; quality is lower than at 4 bits. Other models measured under the same conditions (closest in file size):

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.))

みどり市は約12億円の事業費をかけて中央図書館を改修し、10月1日に再開館した。改修では座席の増設とフロアごとの機能分化、蔵書の入れ替え、バリアフリー化などが行われ、来館者は再開館から1か月で以前比1.8倍に増加した。一方で、駐車場不足や職員負担の増加といった課題も残っており、市は来年度、残る2つの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：来週の打ち合わせ日程変更のご相談

○○株式会社
○○部
○○様

いつも大変お世話になっております。
株式会社△△の△△です。

この度は、来週火曜日（○月○日）に予定しております打ち合わせの件、
大変恐縮ではございますが、日程の変更をお願いしたく、ご連絡いたしました。

社内調整の結果、火曜日にお伺いすることが難しい状況となりましたため、
同日の翌日である水曜日（○月○日）に変更させていただくことは可能でしょうか。

ご多忙のところ、ご迷惑をおかけいたしますことを深くお詫び申し上げます。
ご都合がよろしいようでしたら、水曜日の○時～○時でお願いしたく存じます。

ご確認のほど、何卒よろしくお願い申し上げます。

――――――――――――――――――――
株式会社△△
△△部 △△
TEL: 03-XXXX-XXXX
Email: xxxxx@xxx.co.jp
――――――――――――――――――――

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.)

りんごの代金：120円 × 3個 = 360円
みかんの代金：45円 × 8個 = 360円
合計：360円 + 360円 = 720円
おつり：1,000円 − 720円 = 280円

**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.)

日本の都道府県で海に面していない県は、以下の6つです。

1. 山梨県
2. 長野県
3. 岐阜県
4. 愛知県
5. 滋賀県
6. 三重県

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.)

ご依頼の通り、「春」をテーマにした五・七・五の音数を厳密に守った俳句を3つ作成しました。

**1. 句**
> **春の風 桜吹雪の 舞い落ちる**

*   **音数確認**: 春(1)の(1)風(1) / 桜(1)吹(1)雪(1)の(1) / 舞(1)い(1)落(1)ち(1)る(1) → **5・7・5**
*   **説明**: 春の風に乗って舞い散る桜の花びら（吹雪）の、一瞬の美しさと儚さを描いた句です。

**2. 句**
> **菜の花 咲き乱れて 春の野**

*   **音数確認**: 菜(1)の(1)花(1) / 咲(1)き(1)乱(1)れ(1)て(1) / 春(1)の(1)野(1) → **5・7・5**
*   **説明**: 春の野原一面に咲き誇る菜の花の、力強く明るい生命力を表現した句です。

**3. 句**
> **鶯の 声に誘われ 春の朝**

*   **音数確認**: 鶯(1)の(1) / 声(1)に(1)誘(1)わ(1)れ(1) / 春(1)の(1)朝(1) → **5・7・5**
*   **説明**: 春の朝、鶯の美しい鳴き声に耳を傾け、春の訪れを実感する静かな情景を描いた句です。

How to Get It

It is distributed in GGUF format and can be downloaded from Hugging Face.

In a llama.cpp environment, you can install and start the server with the following commands:

curl -LsSf https://llama.app/install.sh | sh
llama-server -hf bartowski/vectionlabs_Salience-27B-R6-GGUF:Q4_K_M

Example command for individual download using the Hugging Face CLI:

hf download bartowski/vectionlabs_Salience-27B-R6-GGUF --include "vectionlabs_Salience-27B-R6-Q4_K_M.gguf" --local-dir./

In addition to llama.cpp, this model is compatible with LM Studio, koboldcpp, ramalama, Jan AI, Text Generation Web UI, LoLLMs, Atomic Chat, and others.

When using image recognition, you need to specify (--mmproj) the bundled multimodal projector file (mmproj-vectionlabs_Salience-27B-R6-f16.gguf or bf16). Also, to use acceleration via MTP, append --spec-type draft-mtp to the command line.

Related Articles

• clef Structured Decision-Making Model: 12GB+ VRAM, GGUF Builds
• Hemmingway-1 Text Generation Model: 12GB+ VRAM, GGUF Builds
• OrcaSAQ-2-27B Text Generation Model: Our Test Answers, 16GB+ VRAM
• Qwen3.8-27B Vision-Language Model: Our Test Answers, 8GB+ VRAM

What to Read Next

• Find models by VRAM (This model runs from the 12GB tier) → Other models that run on a 12GB GPU
• Explore the same model family → Qwen3.8 family overview (8 articles, 16 converted builds)
• Engines that run this model → llama.cpp / Ollama
• What IQ2_M, Q3_K_L, Q5_K_M mean and where to get this model → GGUF format guide and models / Quantization and model-format glossary
• Learn about the publisher → Alibaba (Qwen): models, licenses and articles

Sources

• https://huggingface.co/bartowski/vectionlabs_Salience-27B-R6-GGUF
• https://huggingface.co/vectionlabs/Salience-27B-R6

Update History

• 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.).