PLaMo 3.0 Prime Deep Dive — Preferred Networks' Flagship Japanese LLM Officially Released June 22, 2026, Expanded from 64K to 256K Context, Dual Reasoning / Non-reasoning Variants, ¥60 / ¥250 per 1M Tokens, Selected for the Digital Agency's Common 'Gennai' Generative-AI Platform, Built From-Scratch on NICT Collaboration and METI GENIAC Phase 3 Outputs

Preferred Networks (PFN) officially released PLaMo 3.0 Prime on June 22, 2026 (official press pr20260622 / PFN tech blog).

Four points:

1. Latest in a fully from-scratch Japanese LLM line — PLaMo-100B (2024) → PLaMo Prime (Feb 2025) → PLaMo 2.0 Prime (2025, Nikkei Excellent Product Award grand prize) → PLaMo 3.0 Prime beta (Mar 19, 2026) → GA (Jun 22, 2026) 2. 256K context + dual reasoning modes — Expanded from the 64K beta; ships as Reasoning model / Non-reasoning model so users pick by task 3. Aggressive pricing — Standard plan at ¥60 input / ¥250 output per 1M tokens, deliberately priced against GPT-5.4 mini / Claude Haiku 4.5 4. Selected as a trial model in the Digital Agency's common generative-AI platform 'Gennai' — public-sector adoption, with miibo / Tachyon / QommonsAI already shipping it as the default

This column sits next to Loop Engineering and Kimi K2.7-Code in our June 2026 LLM coverage.

Key changes from beta: Context length 64K → 256K, improved reasoning and non-reasoning performance, stronger tool use and coding, informed by 3 months of enterprise monitor feedback.

Related open base models on Hugging Face (huggingface.co/pfnet): `plamo-3-nict-2b-base` / `plamo-3-nict-8b-base` / `plamo-3-nict-31b-base`, jointly developed with NICT. The Prime itself is closed-weights; the base family is open.

PFN's tech blog evaluates on 15 benchmarks:

- Japanese: JFBench / IFBench / Japanese MT-Bench / lawqa_jp / MedRECT / Japanese Medical Licensing Exam - English / reasoning: MT-Bench / AIME 2024 / GPQA-Diamond / BFCL (tool use) / LongBench v2 - Safety: HELM Safety on par with or better than overseas models

Comparison set (per PFN):

- Open-weights: gpt-oss-120b (OpenAI), Qwen3.6-27B (Alibaba) - Closed (same price tier): GPT-5.4 mini, Claude Haiku 4.5

PFN CEO/CTO Daisuke Okanohara, on X (Jun 22, 2026):

> "We achieved substantial gains across major benchmarks. Against same-tier commercial and open-weight models, PLaMo 3.0 Prime matches or exceeds them on Japanese instruction following, coding, and tool use."

Important caveat: At the beta stage, ITmedia compared against Qwen3-235B-A22B / gpt-oss-120b (medium reasoning) and flagged that math and multi-tool selection still lagged. Tabulated absolute scores have not been posted publicly. Independent leaderboards like the Nejumi LLM Leaderboard have not yet listed it — that registration is what to watch next.

Channels:

- PLaMo API (SaaS) — Free / Standard / Provider tiers - On-premise - Amazon Bedrock Marketplace - Snowflake - Prime itself is closed; the NICT-collaborative `plamo-3-nict-*-base` family is open on Hugging Face

Pricing (Standard plan, official):

ITmedia calls it "aggressively priced" — clearly aimed at the GPT-5.4 mini / Claude Haiku 4.5 tier.

Platforms shipping it by default: miibo (conversational AI), Tachyon generative AI, QommonsAI (business AI).

Public sector: selected as a trial model in the Japanese Digital Agency's common generative-AI platform 'Gennai'.

Beta monitor program (Mar–Jun 2026): jointly announced with NICT; 3 months of production-scale feedback fed into the GA release.

Target uses: internal-doc summarization, customer support, information extraction, code generation, business-process automation, AI agents, medical / legal domain applications.

NICT collaboration: joint pre-training datasets and medical-data expansion — the foundation of the PLaMo 3 line (base models are named `plamo-3-nict-*`).

GENIAC: METI / NEDO's Generative AI Accelerator Challenge; PLaMo 3.0 Prime explicitly incorporates Phase 3 post-training outputs. GENIAC official.

MN-Core: PFN's own low-power AI processor; the MN-3 supercomputer has won Green500 three times. The release notes don't detail MN-Core + 3.0 Prime co-deployment explicitly.

Leadership: Daisuke Okanohara is President & CTO; founder Toru Nishikawa moved into the Chair role.

Domestic competition: ELYZA, Sakana AI (Sakana Fugu shipped on the same day), Stockmark, Karakuri, CyberAgent CALM, Fujitsu Takane. PLaMo's wedge is from-scratch + reasoning-mode + 256K context + public-sector adoption credentials.

1. Independent benchmark verification missing — only PFN-internal scores so far; Nejumi listing pending. 2. Math and multi-tool selection lagged at beta — whether the GA closes that gap needs third-party measurement. 3. Parameters and architecture undisclosed — dense vs MoE and active-B numbers aren't out, so true cost / inference-efficiency comparisons are hard. 4. Closed-weights distribution — against the open-weights line of Qwen / gpt-oss / Llama 4, this limits ecosystem leverage if the Japanese-language advantage thins out. 5. Domestic-LLM economics — same-tier pricing against GPT-5.4 mini / Claude Haiku 4.5 means PLaMo wins on data-sovereignty and Japan-region processing more than raw quality.

In our AI consulting practice, we recommend a two-phase rollout:

Phase 1 — Same-tier comparison PoC: use the Free / low Standard tier to run PLaMo 3.0 Prime side-by-side against GPT-5.4 mini, Claude Haiku 4.5, and gpt-oss-120b on your own code, documents, and customer tickets. Decide on real workload outcomes, not benchmark scores.

Phase 2 — Sensitive workloads on Japan-resident infra: for finance, healthcare, public-sector and other domains with data-sovereignty / PIPA constraints, deploy PLaMo on-prem or via PLaMo API with Japan-region processing. This is the use case overseas closed-model SaaS can't satisfy.

Q1. Parameter count? A. Undisclosed. The NICT-collaborative base family is open at 2B / 8B / 31B, but the Prime itself is a separate proprietary build whose size and dense/MoE structure aren't published. Q2. Does it beat Claude Opus 4.8 / GPT-5.5 / Gemini 3.1 Pro? A. PFN claims parity in the same tier (GPT-5.4 mini / Claude Haiku 4.5), not against frontier models. No comparison is asserted against Opus 4.8 / GPT-5.5. Q3. Useful outside Japanese? A. English, coding, and math are in the evaluation set (MT-Bench / AIME / GPQA-Diamond / BFCL). Optimization clearly favors Japanese; for pure-English production work, overseas frontier models are likely stronger. Q4. On-prem minimum config? A. Not published. Handled through Provider-plan custom quotes, since the Prime size is undisclosed. Q5. Significance of the 'Gennai' adoption? A. A government-platform selection is a strong procurement signal for Japanese finance / healthcare / public-sector buyers. "The model the government picked" is a powerful brand for data-sovereignty-sensitive procurement. Q6. Relationship to Sakana Fugu (same-day release)? A. Different concepts. PLaMo 3.0 Prime is a single Japanese LLM; Sakana Fugu is a multi-LLM orchestration model. They aren't direct competitors — in fact PLaMo could conceivably be added to Fugu's agent pool.

PLaMo 3.0 Prime is the current high-water mark of from-scratch Japanese LLMs — 256K context, dual reasoning variants, aggressive same-tier pricing, and a government-platform adoption credential. It's also a flagship example of public-private collaboration (NICT + GENIAC Phase 3).

What's still missing is the disclosure that would let third parties confidently rank it against overseas frontier models: parameter count, dense vs MoE, and independent leaderboard scores. Wait for the Nejumi LLM Leaderboard listing before making the final "how good vs Opus 4.8" call. For Japanese enterprises, the pragmatic path is same-tier PoC against overseas models first → sensitive workloads on Japan-resident PLaMo deployments second.

Primary: - PFN press pr20260622 (GA) - PFN tech blog — PLaMo 3.0 Prime GA - PFN tech blog — beta - PFN press pr20260319 (beta monitor) - Hugging Face pfnet (NICT base models) - PFN Corporate Factbook Third-party: - ITmedia AI+ — GA coverage - ITmedia AI+ — beta comparison piece - Dempa Times - NICT announcement (beta monitor) - METI GENIAC Related: - Sakana Fugu (same-day release) - Loop Engineering - Kimi K2.7-Code - Sakana Marlin — autonomous research agent - Liquid AI Japanese-specialized models Note: parameter count, dense / MoE composition, absolute benchmark scores, third-party leaderboard listing, and MN-Core integration details are not confirmable as of June 22, 2026. Re-check PFN's tech blog before any production decision.