Alpha
Contact
Implement · LLM-jp-4.1 thinking

Can you run it?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input D ยท 54.4/100

This page is a projection of the one entry record, the Reliability and Data control factors that Implement covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.

Which domain expands which factor
  • AssessUse & modify + Transparency
  • ImplementData control + Reliability
  • UseReliability
  • SupportTransparency

Install & run

Start from the publisher cookbook (github.com/llm-jp/llm-jp-4-cookbook, Apache-2.0), which defaults to llm-jp/llm-jp-4.1-8b-thinking and uses uv with a lockfile per runtime (ev-cookbook). The Python examples use trust_remote_code=True to load the custom tokenizer and bundled helper files, so the model repository's Python code runs on your machine.

# Transformers (5.0.0 or later), from llmjp4_transformers/
uv sync --locked
uv run --locked example_basic.py

Pin the Hugging Face revision (revision="<commit-sha>") and read llmjp4_tokenizer.py and llmjp4_harmony.py before first use. The weights are ungated and need no login (ev-hub-api).

Hardware & VRAM requirements

The publisher gives no minimum. The cookbook records one test machine (Intel Core i9-14900K, 32 GiB RAM, one NVIDIA RTX 6000 Ada, CUDA 12.8 or 13.0) and says it is "not a statement of minimum requirements" (ev-cookbook). File sizes give a floor (ev-hub-api, ev-gguf):

BuildWeights on diskNote
33B BF16 (safetensors)about 66 GB14 shards; plan for about 70 GB plus KV cache
33B GGUF Q4_K_M20.4 GBthe build the publisher validated on llama.cpp v0.6.0 with a 4,096-token context
32B-A3B BF16about 64 GB13 shards; holds all experts in memory, computes with 3.8B active
8B BF16about 17 GB4 shards

These are OneHill sums of published file sizes, not publisher requirements. KV-cache memory at the full 64K context is not documented; measure it for your batch size.

Serving stacks

StackMinimum per the cookbookCondition
Transformers5.0.0parse_response support for the bundled schemas; trust_remote_code=True
vLLM0.30.0plus llm-jp-vllm 0.1.1 or later for the Harmony reasoning and tool parsers
llama.cppupstream v0.6.0native LLM-jp-4.1 Harmony parsing, GGUF only

Contradiction: the GGUF model card says llama.cpp "currently requires a fork" because upstream lacked the tokenizer fixes, while the cookbook says upstream v0.6.0 or later includes the parser (ev-gguf, ev-cookbook). The cookbook is the newer repository push but the card was modified later (2026-10-20). Test your llama.cpp build with a chat and a tool-call request before relying on it. Ollama and TGI support are not documented by the publisher.

Safe-deployment controls & Deployment Ceiling

Ceiling: T1 (internal assistant, human reads every output), conditional. The publisher says the models are not intended for direct production use, evidence for safety is one self-run benchmark, and no guard model exists, so OneHill does not place customer-facing use (T2) within reach without your own evidence.

To reach T2 you would need, at minimum:

  • your own jailbreak and misuse evaluation in the languages you serve, run on the exact revision;
  • an input and output guardrail layer and a guard classifier of your own;
  • review of llmjp4_tokenizer.py and llmjp4_harmony.py, and a pinned revision with verified sha256 values;
  • sandboxed execution for any tool the model calls, since tool calling is new in 4.1 and only evaluated on BFCL and tau2-bench by the publisher;
  • a written statement of intended purpose that records you are deploying a model its publisher describes as research and development only.

Available quantizations

Official GGUF repositories for all three sizes, created 2026-09-24: BF16 and Q4_K_M files with an imatrix (ev-gguf). No official AWQ, GPTQ, FP8 or MLX build exists. Community derivatives were not assessed.

Fine-tuning & adaptation

What is available: Apache-2.0 base models (8B, 32B-A3B, 33B from the 4.0 series), the 4.1 SFT dataset (subsets under CC BY, ODC-BY, CC BY-SA, Apache-2.0 and MIT) and the three DPO datasets, all ungated (ev-sft-dpo). The release page lists trl-based SFT and DPO scripts and a Nemo-Aligner route (ev-release-page).

What is not confirmed: that those scripts are the ones used for the 4.1 recipe (they sit in repositories last pushed in 2024), and the withheld subsets (AnswerCarefully, ichikara, llmc_math_dataset, synthetic_if_ja). If you fine-tune, keep the modified Harmony chat template (ev-cookbook) and mix in your own safety data, because SFT on new data can undo the safety post-training.

API / OpenAI-compatible integration

vLLM serve exposes /v1/chat/completions with the publisher's parsers, which separate reasoning from the final answer and support tool calls (ev-cookbook):

vllm serve llm-jp/llm-jp-4.1-8b-thinking \
    --reasoning-parser llmjp4 --reasoning-parser-plugin llm_jp_vllm.llmjp4 \
    --enable-auto-tool-choice --tool-call-parser llmjp4 --tool-parser-plugin llm_jp_vllm.llmjp4 \
    --trust-remote-code

llama serve -hf llm-jp/llm-jp-4.1-8b-thinking-gguf:Q4_K_M gives an OpenAI-compatible endpoint that accepts tools and tool_choice. NII also announced a trial hosted API and chat service through the University of Tokyo mdx-MaaS, mainly for researchers (ev-pcwatch); it is not a production service and its terms are unverified.

How this scores

The ownership factors this domain covers, drawn from the one entry record.

3

ReliabilityIs it reliable and good enough for the job?

Weak

Safety scores 2: one publisher-run Japanese benchmark, no independent evaluation or guard model, and the publisher says the models are not for direct production use; operational is 3 (ev-card-41, ev-blog, ev-cookbook).

How this scores (AOI sub-dimensions)
Operational3/5how practical it is to run, serve and maintain in productionRuns on Hugging Face Transformers, vLLM and llama.cpp with a maintained Apache-2.0 cookbook, official GGUF builds in BF16 and Q4_K_M, a 64K context and a Q4_K_M 33B validated on a local GPU.
Safety2/5whether misuse risks are evaluated and guardrails are providedThe 4.1 models are safety post-trained (the card says SFT and DPO "for instruction following and safety", and the report describes an upsampled Japanese safety set and safety-judged preference pairs), and the residual risk is disclosed.
4

Doesn't extract your dataDoes running it keep your knowledge and data yours?

Strong

Self-hosted weights under Apache-2.0 with no telemetry or claw-back; review the two repository Python files that run under trust_remote_code (ev-license, ev-hub-api).

How this scores
Not a scored AOI dimension. For a self-hosted model, data-control is a structural property of running the weights yourself, strong by default unless the model phones home or the licence claws back rights. For a hosted API this factor is the retention + train-on-inputs + residency read, scored from the binding terms.
What this means for adoptionYou own the weights, the data lineage and the right to run and change them: Apache-2.0, ungated, with public corpora, SFT and DPO sets and pre-training scripts, so this is one of the most transparent models reviewed. You do not own a deployable safety case: the publisher says the models are for research and development and not for direct production use, offers one self-run Japanese safety benchmark, no guard model and no independent test. Ownership moves to substantial if an independent evaluation lifts safety to 3 or more, and you can reach that yourself by running your own safety and jailbreak tests behind your own guard.

Sources

The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.

Model cardread2026-11-19
The llm-jp-4.1-33b-thinking model card declares Apache-2.0 in the front matter and 11.7T tokens; the 8B and 32B-A3B cards are identical on the points used.
Model cardread2026-11-19
The LLM-jp-4 (4.0) base, 33b-thinking and 8b-instruct cards say "The models released here are in the early stages of our research and development and have not been tuned to ensure outputs align with human intent and safety considerations".
Licenceread2026-11-19
Apache-2.0 is declared in the card front matter and License section of all three 4.1 cards and in section 2.4 of the technical report.
Technical_reportread2026-11-19
The LLM-jp-4 Technical Report (arXiv 2610.19331 v1, 2026-10-14) covers 8B dense and 32B-A3B MoE models trained from scratch on about 12T tokens.
Documentationread2026-11-19
LLM-jp Corpus v4.1 README: English FineWeb-Rescored and MegaMath-Web-Pro-Max-OSS (ODC-BY plus Common Crawl terms) and The Stack v2 code (per-repository licences), hosted on a public GitLab project; copyright stays with original authors; document-removal contact given.
Documentationread2026-11-19
LLM-jp Corpus v4 README lists Japanese, English, Korean, Chinese and code subcorpora with licences (mostly CC BY, ODC-BY, CC BY-SA).
Documentationread2026-11-19
LLM-jp Corpus Midtraining v2 README: 13 English and 2 Japanese sources with licences, including the nemotron_pretraining_code_v2 corpus (491B tokens) under the NVIDIA Data Access Agreement for Model Training and llm-jp-IPT_v0.3.2 under "Various licenses".
Documentationread2026-11-19
The llm-jp-4.1 SFT and DPO dataset cards (ungated) list per-subset licences, and responses were generated with gpt-oss-120b.
Documentationread2026-11-19
github.com/llm-jp/scripts (Apache-2.0) holds LLM-jp-4 pre-training and mid-training configurations (v4-8b, v4-32b which matches the 33B dimensions, v4-32b-a3.8b), installers, corpus-building and checkpoint-conversion scripts.
Documentationread2026-11-19
The Apache-2.0 llm-jp-4-cookbook documents Transformers 5.0.0 or later, vLLM 0.30.0 or later with llm-jp-vllm 0.1.1 or later, and llama.cpp v0.6.0 or later.
Model cardread2026-11-19
The official llm-jp-4.1-33b-thinking-gguf card (modified 2026-10-20) says running with llama.cpp "currently requires a fork of llama.cpp", which contradicts the cookbook; the repositories hold BF16 (66.4 GB) and Q4_K_M (20.4 GB) GGUF files and an imatrix.
Documentationread2026-11-19
The Hugging Face API lists the three 4.1 repositories as ungated, with safetensors BF16 shards carrying LFS sha256, no pickle files, no LICENSE file and a completed hub scan with no flagged files.
Documentationread2026-11-19
The NII release page lists the 4.1 models, corpora and fine-tuning scripts and says "Users of this model are requested to use it appropriately and at their own responsibility, in accordance with" Japan's December 2025 AI R&D and utilisation guideline.
Vendor announcementread2026-11-19
Publisher blog (2026-09-28, Japanese): 4.1 improves post-training for STEM and adds tool calling, the first LLM-jp series with it.
Third-party analysisread2026-11-19
PC Watch (Impress, 2026-09-29) reports that NII released the three LLM-jp-4.1 models on 2026-09-28 and that Apache License 2.0 applies to all.
Documentationunverified2026-11-19
The AnswerCarefully dataset (safety training data and benchmark) is gated (manual approval, licence:other); its LICENSE and README returned HTTP 401, so its terms are documented but unverified.