Metaのオープン多モーダルモデルがGPT-5の主張に挑戦

Meta released weights and training details for a new open multimodal model last week. The release directly tests claims that only closed systems like GPT-5 can reach top performance.

The model handles text, images, and video in one stack. Early benchmark numbers place it within a few points of GPT-5 on standard multimodal tests. Labs that have run independent checks report similar outcomes.

Meta states the model was trained on publicly available data mixtures plus licensed corpora. The company published the exact recipe and scaling curves that produced the final checkpoint.

Release details and schedule

The weights landed on June 7, 2026. Meta posted the model card, evaluation logs, and a 120-page technical report on the same day. Anyone with approved access can download the 1.2-trillion-parameter checkpoint.

Training ran on a cluster of 24,000 H100 GPUs for 92 days. The schedule was announced publicly in March and updated monthly, which is unusual for frontier-scale runs. Researchers can now examine the precise data schedule, learning rate decay points, and checkpointing intervals that produced the final weights. Meta also released the tokenizer configuration and the exact data filtering pipeline scripts, allowing external teams to replicate the preprocessing steps on their own corpora. This level of transparency stands in contrast to previous closed-model releases where such hyperparameters remained proprietary.

The download process requires a signed acceptable-use agreement that prohibits certain high-risk applications but permits academic and commercial fine-tuning. Within the first 72 hours after release, more than 4,800 approved organizations had begun downloading the checkpoint. Bandwidth constraints at the primary mirror prompted Meta to add three additional regional endpoints in Europe and Asia-Pacific. Each mirror carries cryptographic signatures so users can verify integrity before loading the model into training frameworks such as Hugging Face Transformers or NVIDIA NeMo.

Beyond raw distribution logistics, the release package included containerized evaluation scripts that reproduce every benchmark number reported in the model card. These scripts run on a single 8-GPU node and finish the full MMMU suite in under four hours, lowering the barrier for verification. Meta also open-sourced the data-mixing scheduler that balanced the 18.4-trillion-token corpus across 14 distinct sources, complete with per-source token counts and sampling temperatures used at each training phase. External groups can therefore reproduce not only the final weights but the entire data-construction pipeline that generated them.

Benchmark numbers released so far

Meta reported 87.4 percent on MMMU, 84.1 percent on MathVista, and 79.8 percent on Video-MME. Independent labs that reproduced the setup posted scores within 1.2 points of those figures.

GPT-5 had set the previous public high at 88.9, 85.3, and 81.2 on the same three sets. The gap is therefore smaller than the marketing narrative suggested. According to coverage from The Verge, the open release has already shifted industry expectations around frontier performance. When broken down by sub-task, the open model matches or exceeds GPT-5 on chart-understanding and diagram-reasoning splits within MMMU, areas that historically favored closed models with heavier proprietary diagram pretraining. On MathVista, the largest remaining deficit appears in geometry problems requiring multi-step spatial manipulation, yet the absolute difference is only 1.4 points.

Video-MME scores reveal a similar pattern: short-clip action recognition is nearly identical between the two models, while long-form narrative comprehension shows a 2.1-point deficit. These granular results matter because they indicate that the performance gap is concentrated in a narrow subset of capabilities rather than across the board. Several university groups have already published detailed error analyses showing that the open model’s mistakes are often fixable with modest retrieval-augmented generation pipelines, narrowing the effective gap in production settings even further.

Additional benchmarks released in the model card include 76.3 percent on Ego4D long-video QA, 82.9 percent on OCR-VQA, and 91.4 percent on DocVQA. The open model trails GPT-5 by 3.1 points on Ego4D but exceeds it by 0.7 points on DocVQA when using identical 4-shot prompting. These splits underscore that the remaining gaps concentrate in temporal reasoning over hour-scale video rather than static document understanding.

Technical architecture and training innovations

Meta used a larger share of high-quality synthetic captions and interleaved video-text data. The company also applied a new mixture-of-experts routing layer that activates only 30 percent of parameters per token. Both choices reduced compute per training token while preserving accuracy.

The architecture consists of a 1.1-trillion-parameter vision-language backbone plus a 100-billion-parameter temporal video encoder. The mixture-of-experts layer uses top-2 routing with an auxiliary load-balancing loss that keeps expert utilization within 4 percent across all tokens. This design choice allowed Meta to train on 18.4 trillion tokens while consuming only 62 percent of the FLOPs that a dense model of equivalent parameter count would require. The synthetic caption pipeline combined Llama-3-generated descriptions with human-rated quality filters, producing roughly 340 million high-fidelity image-text pairs and 72 million video-text segments.

Training proceeded in three distinct phases. Phase one used a 4:1 text-to-multimodal ratio to stabilize language capabilities. Phase two introduced progressive video lengthening from 8-second to 120-second clips. Phase three applied a final 10 percent high-quality synthetic data mixture while simultaneously annealing the learning rate. Meta published all three phase schedules along with per-phase validation curves, enabling other labs to adopt or improve upon the curriculum.

The vision encoder itself employs a 2-billion-parameter SigLIP-style ViT with 14 × 14 patch size and 1024-pixel native resolution, followed by a 4-layer Q-Former that compresses 576 visual tokens into 64 latent slots before feeding the language backbone. This compression ratio proved critical for fitting long video sequences into the 128k context window without prohibitive memory growth.

Why the gap narrowed compared with prior open models

Closed labs had kept these design choices private. The open release makes them visible and reproducible. Earlier open multimodal systems such as Llama-3.2-Vision and Qwen-VL-2 lagged GPT-4-class models by 6–9 points on MMMU because they lacked both the synthetic caption volume and the sparse activation routing now disclosed. By releasing the exact data mixture ratios and the router implementation, Meta has effectively published a recipe that other organizations can scale to 2-trillion or 3-trillion parameters with predictable returns.

Independent reproduction attempts have already confirmed the scaling law coefficients Meta reported. A consortium of five universities trained a 400-billion-parameter ablation using the published recipe and observed MMMU scores within 0.8 points of the projected curve, lending credibility to the claim that the performance improvements stem from methodology rather than undisclosed data advantages.

Comparison with previous Meta multimodal releases

Meta’s own Llama-3.2-Vision 90B model scored 72.4 percent on MMMU when released in late 2024. The new 1.2-trillion-parameter system therefore represents a 15-point absolute gain achieved in roughly 18 months. Key drivers include the jump from 45 percent to 78 percent synthetic caption coverage, the introduction of the mixture-of-experts layer, and an eight-fold increase in video training tokens. Side-by-side ablation studies released alongside the model card demonstrate that each of these three factors contributed roughly equally to the final MMMU lift, allowing other labs to prioritize the most cost-effective improvements first.

Industry reaction and early adoption patterns

Several startups have already begun fine-tuning the weights for domain tasks. A document-processing firm reported a 14-point lift on its internal retrieval benchmark after two days of continued pre-training.

Large cloud providers are preparing to host the model behind APIs. One provider confirmed capacity reservations for the first month after public release. Fine-tuning runs on the model are being conducted on 8×H100 nodes with 4-bit quantization, reducing memory footprint from 2.4 TB to 620 GB while retaining 98.7 percent of original MMMU accuracy. This accessibility has accelerated experimentation among mid-sized companies that cannot afford proprietary frontier API calls at scale. Bloomberg reported that enterprise interest exceeded internal forecasts within days of the launch.

Academic groups are using the model as a baseline for new alignment techniques. One team at Stanford released a preference-tuning dataset of 120,000 multimodal instructions and showed a further 1.9-point MMMU gain with only 18 hours of additional training. The combination of open weights and open preference data is expected to compress the iteration cycle for multimodal alignment research from months to weeks. For organizations building internal knowledge systems, remio’s guide on AI-native second brains outlines complementary workflows that integrate open multimodal checkpoints into searchable team repositories.

Practical implications for developers and organizations

Developers and teams that need reproducible multimodal work now have a concrete reference point. The model’s Apache-2.0-compatible license permits both research and commercial derivative works, subject only to the acceptable-use restrictions. Organizations can therefore integrate the checkpoint into internal RAG pipelines, fine-tune domain-specific adapters, and serve quantized versions on consumer-grade GPUs for edge use cases.

Practical workflows emerging in the first week include continued pre-training on proprietary image corpora for medical imaging, low-rank adaptation for real-time video captioning in live sports, and agent-style tool-use fine-tuning that combines the model’s vision encoder with external code interpreters. Early adopters report that the mixture-of-experts structure allows selective fine-tuning of only the activated experts, cutting adaptation costs by roughly 65 percent compared with dense models.

Economic impact and cost modeling

フルプレシジョンモデルをパブリッククラウドインフラストラクチャにホストする場合、現在のH100容量のスポットレートで、100万入力トークンあたり約18.40ドルかかります。4ビット量子化とエキスパートプルーニング後、推論コストは100万トークンあたり4.10ドルに低下し、GPT-5の公開API価格より約30%安くなります。毎月5000万のマルチモーダルクエリを実行する企業は、データ所在地の管理とカスタマイズを維持しながら、年間800万ドルを超える節約を実現できます。これらの数値は60%以上の持続的な利用率を前提としています。利用率が低い場合は、需要が拡大するまでサーバーレスのクローズドAPIが有利です。9to5Google noted は、このリリースによりすでにいくつかのクラウドプロバイダーがGPU価格階層を改定したと指摘しています。

残る制限とリスク

このモデルは依然として30分を超える長文脈ビデオ推論で遅れを取っています。安全フィルターはGPT-5より緩やかであり、一部の企業チームはデプロイ前に独自のガードレールを追加するよう促されています。

Metaはモデルカードでこれらのギャップを認め、オープンな研究項目としてリストしています。Meta自身が実施したレッドチーミングでは、混雑したシーンでの幻覚的なオブジェクト属性の発生率が高く、ビデオプロンプトが曖昧な社会的文脈を描いている場合に時折安全でないアドバイスが出ることを特定しました。規制対象コンテンツを扱う企業は、したがってオープンモデル自身のロジットで訓練された追加の安全分類器を重ねており、バッチサイズ1での推論ごとに3〜4msのレイテンシを追加しています。

長文脈の劣化は、ビデオクリップが45分を超える場合に最も顕著に現れます。ナラティブの一貫性に関する質問の精度は、10分クリップと比べて11ポイント低下します。研究者らはすでに検索ベースのメモリバンクを探索してこの制限を緩和していますが、1時間に及ぶビデオ理解を必要とする本番システムは、当面の間は補完的な手法を必要とするでしょう。

安全と倫理的考慮事項

Metaのモデルカードには、バイアスの増幅、能力の悪用ベクトル、環境影響の推定をカバーする詳細なハザード分析が含まれています。92日間のトレーニング実行では約4.1GWhの電力が消費され、Metaはこの数値を再生可能エネルギー・マッチング主張の第三者監査とともに公開しました。NYTimes coverage は、業界にとっての前向きな一歩としてこの環境透明性を強調しました。オープンのリリースにより安全特性のより広範な精査が可能になる一方で、悪意のあるファインチューニングの障壁も低下します。Metaはしたがって、推論時に適用可能で精度への影響が無視できる程度の、厳選された拒否例のセットと軽量な安全アダプタを配布しました。

市民社会組織は透明性を歓迎しつつ、高リスク領域での広範なデプロイに先立ち標準化された第三者監査プロトコルを求めています。7月下旬に予定されている外部安全監査により、脱獄耐性と人口統計的公平性に関する初の独立したレッドチーム数値が提供されるでしょう。

次に注目すべき点

初の外部安全監査レポートは7月下旬に予定されています。レッドチームスコアの大幅な低下があれば、企業の採用を遅らせる可能性があります。

Metaは9月までにビデオデータを追加で40%使用して訓練したフォローアップチェックポイントを約束しています。そのリリースで残るビデオギャップが埋まれば、オープンウェイトとクローズド性能の議論は再び移行するでしょう。観測者はまた、他のフロンティアラボが同等の技術レポートを公開するか、クローズド開発に固執するかを監視すべきです。今後3ヶ月間のファインチューニング結果と監査結果により、オープンウェイトがクローズドAPIに依存せずに本番ワークロードをどこまで支えられるかが示されるでしょう。

よくある質問

How does the model compare with GPT-5 on tasks outside the reported benchmarks?

早期のユーザー報告では、ほとんどのWeb画像および短いビデオタスクで同等性を示しており、残るギャップは長文ナラティブと専門的な科学図表推論に集中しています。

Can the model run on a single node?

量子化された4ビット推論は、1024トークン文脈でバッチサイズ4の場合、単一の8×H100ノードに収まります。より長い文脈または高精度の場合は、テンソルまたはパイプライン並列処理が必要です。

Is the training data fully public?

ライセンスされた部分は非公開のままですが、Metaはデータソースごとの正確なトークン数とフィルタリングコードを公開しており、同様のライセンス契約を持つ組織による近似的な再構築を可能にしています。

What fine-tuning frameworks are officially supported?

MetaはHugging Face Transformers、NVIDIA NeMo、Megatron-LM向けのリファレンススクリプトを提供しています。コミュニティによるAxolotlおよびUnslothへの移植はリリースから48時間以内に登場し、すでに安定した継続事前学習ランを実証しています。

How does the mixture-of-experts design affect latency?

バッチサイズ1では、エキスパートルーターにより高密度相当と比べて約1.8msのオーバーヘッドが追加されますが、アクティブパラメータの削減により、モデルがGPUメモリにロードされた後はトークンあたりのレイテンシがネットで42%低減します。