What are the strengths of this model?

High performance with 405B parameters Optimized for Japanese cultural context Built on the strong Llama 3 base model High quality in long-text generation

What are the weaknesses of this model?

Requires high-spec GPUs to run Non-commercial license (Llama 3 License) High inference cost No API availability

What are the best use cases?

High-quality Japanese content generation Complex Japanese language reasoning tasks Research and development use Processing large-scale Japanese data

Sakana AIConditional Open

Llama-3-Namazu-405B

Name: Llama-3-Namazu-405B
Author: Sakana AI

Compare this model

A large-parameter version of the Namazu model developed by Sakana AI. Based on Llama-3-405B, it has undergone post-training optimized for Japanese cultural and social contexts.

Share:X Hatena LINE

Parameters

405B

Context Window

128K

License

Llama 3 License

Release Date

2026-03-15

Japanese Language Capability

🇯🇵Native JP

Model developed by a Japanese company or specialized for Japanese. Highest Japanese understanding and generation capability.

API Pricing

API pricing for this model is not yet available

Strengths

・High performance with 405B parameters
・Optimized for Japanese cultural context
・Built on the strong Llama 3 base model
・High quality in long-text generation

Weaknesses

・Requires high-spec GPUs to run
・Non-commercial license (Llama 3 License)
・High inference cost
・No API availability

Use Cases

・High-quality Japanese content generation
・Complex Japanese language reasoning tasks
・Research and development use
・Processing large-scale Japanese data

Deep Analysis

MMLU (5-shot)

88.6%

Near parity with GPT-4o (~88.7%)

HumanEval (Coding)

89.0%

Competitive, behind Claude 3.5 Sonnet (92.0%)

Input Price

$2.40/M tokens

Via Amazon Standard provider

Output Price

$2.40/M tokens

Via Amazon Standard provider

Context Window

128K tokens

Llama 3.1 upgrade from 8K

Base Model Focus

Japanese Cultural Context

Post-training for neutrality and factual accuracy

Strengths

・Frontier-level performance competitive with closed-source models on core benchmarks.
・Optimized via post-training to reduce political/cultural bias and refusal rates for Japanese-context queries.
・Open-weight model with extensive self-hosting and fine-tuning capabilities for data sovereignty.
・Cost-effective API access available through multiple providers at significant discounts vs. GPT-4o.

Weaknesses

・Massive computational requirements for self-hosting (200+ GB VRAM at INT4, requires multi-GPU setup).
・Text-only input (no native vision or audio), unlike multimodal competitors like GPT-4o.
・Knowledge cutoff from pre-training data (December 2023) may lag behind current events without RAG.
・Namazu variant's specialized optimizations may have reduced general applicability outside targeted contexts.

Competitor Comparison

Model	Arena	SWE	GPQA	Price
Meta Llama 3.1 405B (Base)	N/A	N/A	51.1%	$2.40/$2.40
GPT-4o	N/A	N/A	53.6%	$2.50/$10.00
Claude 3.5 Sonnet	N/A	N/A	59.4%	$3.00/$15.00

Overview

Llama-3-Namazu-405B is a specialized, large-parameter variant developed by Sakana AI, built upon Meta's Llama 3.1 405B architecture. It represents a focused post-training effort aimed at optimizing the model for Japanese cultural and social contexts. The primary innovation is not raw benchmark superiority—its core capabilities are derived from the powerful base model—but in its fine-tuning to correct inherent biases and reduce refusal rates on politically sensitive topics prevalent in models trained primarily on Western-centric data. As demonstrated in Sakana's internal benchmarks, Namazu drastically lowered refusal rates for queries about sensitive historical and political themes (from ~72% to nearly 0% for its DeepSeek-based variant) while maintaining factual accuracy and neutrality. The model positions itself within the open-weight ecosystem as a solution for developers and organizations in Japan requiring an AI that provides balanced, multi-perspective responses without automatic self-censorship on culturally specific topics. It inherits the frontier-level performance of the base 405B model—competitive with GPT-4o on general knowledge and math benchmarks—while offering the typical advantages of open weights: self-hosting for data privacy, potential for domain-specific fine-tuning, and cost-efficient deployment at scale. However, its significant computational demands limit practical deployment to well-resourced entities or via API providers.

Benchmarks & Performance

Llama-3-Namazu-405B's performance is fundamentally benchmarked against its base model, Meta's Llama 3.1 405B, as it retains the same architectural capabilities. The post-training process focuses on alignment and bias correction rather than enhancing raw benchmark scores. | Benchmark | Llama 3.1 405B (Namazu Base) | GPT-4o | Claude 3.5 Sonnet | Notes | | :--- | :--- | :--- | :--- | :--- | | MMLU (5-shot) | 88.6% | ~88.7% | ~88.7% | General knowledge; all models are statistically tied. | | HumanEval (Coding) | 89.0% | 90.2% | 92.0% | Code generation; Claude leads. | | MATH (0-shot CoT) | 73.8% | 76.6% | 71.1% | Mathematical reasoning; GPT-4o leads, 405B is strong. | | GPQA Diamond | 51.1% | 53.6% | 59.4% | Graduate-level science; Claude has a significant lead. | | IFEval | 88.6 | 87.1 | 88.9 | Instruction following; all competitive. | | MGSM (Multilingual Math) | 91.6% | 90.5% | 91.6% | Multilingual math; tied at the top. | **Key Performance Insight:** On standardized academic benchmarks, the 405B model (and thus Namazu) is within 1-3% of frontier closed-source models, confirming the closure of the open-closed gap. Its strengths lie in mathematics and general instruction following. The explicit performance uplift from the Namazu post-training is documented in Sakana AI's internal metrics for neutrality and refusal reduction, not in these public academic benchmarks.

Detailed Comparison

Community Feedback

The broader developer and researcher community primarily recognizes the base Llama 3.1 405B as a landmark open-weight model that closed the performance gap with proprietary systems. Its release is celebrated for enabling self-hosted, privacy-preserving AI and cost-effective customization via fine-tuning. The Sakana AI Namazu variant is viewed as a specialized and culturally significant application of this technology, particularly in Japan. Reactions focus on the ethical and practical implications of reducing inherent biases in large models and the technical challenge of creating region-specific adaptations without eroding core capabilities. Adoption patterns show developers using the base 405B for general-purpose high-capability tasks and for generating synthetic data to train smaller, cheaper models (distillation). The Namazu variant attracts interest from organizations and researchers in Japan needing models that navigate nuanced local topics without excessive caution, though its specialized nature limits broader adoption outside that region.

Use Cases

1. **Culturally Nuanced Content Generation & Q&A (Japan Focus):** Choose Llama-3-Namazu-405B over GPT-4o or Claude for applications serving Japanese users that require balanced perspectives on historical, political, or social topics. Example: An AI assistant for educational materials or news analysis in Japan can provide factually accurate, multi-faceted answers without defaulting to refusal on sensitive topics, as per Sakana AI's reported reduction in refusal rates. 2. **Cost-Effective, High-Quality Synthetic Data Generation:** Choose this model over API-only competitors for large-scale synthetic data creation. Its open-weight license allows unlimited inference for generating training data (e.g., for smaller model distillation, dataset augmentation) at a one-time infrastructure cost, avoiding the high per-token fees of GPT-4o or Claude which accumulate rapidly over millions of tokens. 3. **On-Premise Enterprise Deployment for Data-Sensitive Workflows:** Choose over any closed API model when processing confidential internal data (e.g., legal documents, proprietary research, internal communications). Self-hosting ensures data never leaves the corporate network, complying with stringent data sovereignty regulations—a capability the closed-source models cannot offer. 4. **Foundation for Domain-Specific Fine-Tuning:** Choose the base Llama 3.1 405B (or Namazu as a specialized starting point) for creating industry-specific expert models. For instance, fine-tuning on medical literature or financial regulations to create a specialized analyst bot, leveraging the model's strong reasoning and instruction-following capabilities as a robust starting point.

Latest News

**March 2026:** Sakana AI officially announced the "Namazu" model series and launched the accompanying chat service "Sakana Chat". Namazu is presented as a prototype applying their post-training technology to multiple frontier open-weight models, including a version based on Llama 3.1 405B. The service includes web search integration for real-time information retrieval. The key technical result highlighted was a dramatic reduction in answer refusal rates for politically sensitive topics—from approximately 72% in the base DeepSeek model to nearly 0% after Namazu-style post-training—while maintaining factual accuracy and a neutral tone. Sakana AI emphasized that this demonstrates the ability to technically remove external constraints from highly capable models without degrading their core abilities. **Ongoing:** The API pricing and performance landscape for the base Llama 3.1 405B continues to evolve, with providers like Amazon Web Services and Azure competing on speed, latency, and cost. Amazon Standard remains one of the most cost-effective providers at $2.40 per million blended tokens.

The model positions itself within the open-weight ecosystem as a solution for developers and organizations in Japan requiring an AI that provides balanced, multi-perspective responses without automatic self-censorship on culturally specific topics. It inherits the frontier-level performance of the base 405B model—competitive with GPT-4o on general knowledge and math benchmarks—while offering the typical advantages of open weights: self-hosting for data privacy, potential for domain-specific fine-tuning, and cost-efficient deployment at scale. However, its significant computational demands limit practical deployment to well-resourced entities or via API providers.

Sources

Analysis generated: 2026-05-23