AI Model Daily - 2026-08-12 | QName AI

30-Second Summary

NVIDIA Nemotron 3.5 Lightning is the clearest new open-weight model in this window. NVIDIA officially released it on 2026-08-11: 30B total parameters, 3B active, hybrid MoE/Mamba-2/Attention architecture, OpenMDW 1.1, 1M context, and a target use case around long-running agents and sub-agent workhorses. NVIDIA Build and Hugging Face entries are live.
Artificial Analysis released AA-AnalystAgent, a new agentic quantitative-analysis benchmark for real spreadsheet/document analysis tasks. Claude Opus 5 leads at 54%, followed by GPT-5.5 at 50%, Claude Fable 5 at 49%, and the top open-weight model, Kimi K3 (max), at 39%.
FLUX 3 Video received an Arena Text-to-Video update: #2 with 1496±17 Preliminary, only 16 points behind #1 Gemini Omni Flash at 1512±11. This is an Arena update/access promotion, not the first release of a new base video model today.
Mistral announced regional inference and open-model platform updates: regional endpoints are generally available, Priority Tier is in public preview, and third-party open models are being added starting with Z.ai GLM-5.2. This is a deployment/platform availability change, not a new Mistral base model.
Luma released Luma Scenes, powered by Uni-1, using storyboard/keyframe/timing approval to turn a brief or reference assets into controllable video scenes. Runway also added Seedance 2.5 to Runway. Today's video signal is about workflows and platform access, not a single new video foundation model.
The Qwen3.8-Max ModelScope pre-release page shows Qwen3.8-2.4T-A95B scheduled for 2026-08-12 15:00 UTC, or 23:00 HKT. At this report point, it should still be marked pending release/verification, not available open weights.

Must Watch

1. NVIDIA Nemotron 3.5 Lightning: 30B/3B Active Open-Weight Agent Model

Event type: Official release / open weights / X first signal / access update / public benchmark update.
First/official/availability time: NVIDIA's technical blog, Build.NVIDIA model card, and Hugging Face model page all list 2026-08-11. @NVIDIAAI posted on X the same day.
Capability change: NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 is a 30B total, 3B active MoE model with a Mamba-2 + MoE + Attention hybrid architecture and up to 1M context. NVIDIA positions it for long-running autonomous agents, sub-agent workhorses, local/edge deployment, and low-latency agentic workflows.
Key parameters/performance: NVIDIA says it can deliver up to roughly 4x output speed versus similar-sized models. The blog reports PinchBench 86% accuracy and says it completes 10,000 tasks about 30% faster than Qwen3.6 35B at similar accuracy. The Build model card lists NVFP4 scores including SWE-bench Verified 52.80, Terminal-Bench 23.46, PinchBench 83.43, BrowseComp 36.81, and GDPval-AA-V2 865.
Access: NVIDIA Build model page and Hugging Face nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4. The model card says weights, training data, and recipes are open, commercial use is allowed, and the license is OpenMDW 1.1.
Benchmark/third-party view: Artificial Analysis X shows an Intelligence Index of 24, up 9 points from Nemotron 3 Nano's 15, and roughly in line with gpt-oss-120b. However, on the Arena general Text leaderboard it is #203 with 1345±12 Preliminary. The correct read is scenario-specific: it is worth testing for agent execution, terminal work, and long-running sub-tasks, not as a top general chat model.
Impact: This is an important signal that NVIDIA is bundling open models, inference optimization, and routing into the agent stack. The first tests should cover long-running shell/terminal agents, locally deployable sub-agents, and whether NeMo Switchyard routing to Lightning meaningfully reduces cost and latency.
Sources: NVIDIA technical blog, NVIDIA Build model card, Hugging Face, NVIDIAAI X, NVIDIAAI X - open materials, Artificial Analysis X, Arena Text.

2. Artificial Analysis AA-AnalystAgent: New Spreadsheet/Document Agent Benchmark

Event type: Public benchmark update / X first signal / benchmark release / pending reproducibility.
First/official/availability time: @ArtificialAnlys released AA-AnalystAgent on 2026-08-11, with an evaluation page and the stirrup GitHub repository.
Capability change: AA-AnalystAgent targets real-world quantitative analysis tasks. Models must work with spreadsheets and documents, then produce analytical judgments. It is closer to analyst/operator workflows than plain knowledge QA, and is useful for measuring agent reliability across files, tables, and long reasoning chains.
Current ranking signal: Claude Opus 5 leads at 54%, followed by GPT-5.5 at 50% and Claude Fable 5 at 49%. Among open-weight models, Kimi K3 (max) is the top result at 39%, about 15 points behind the closed frontier. AA also mentions DeepSeek V4 Flash among the next open-weight candidates.
Access: Artificial Analysis evaluation page and the ArtificialAnalysis/stirrup GitHub repository.
Impact: This benchmark is especially relevant for AI analyst, office-agent, automated reporting, investment research, and financial-analysis products. Future reports should read AA-AnalystAgent together with GDPval-AA, Terminal-Bench, and Arena Agent, instead of relying only on chat Elo.
Sources: AA-AnalystAgent X, open-weight ranking X, evaluation page, stirrup GitHub.

3. FLUX 3 Video Reaches #2 on Arena Text-to-Video, Close to Gemini Omni Flash

Event type: Arena appearance / X first signal / access promotion / technical details pending.
First/official/availability time: @arena said on 2026-08-11 that it was testing a FLUX 3 Video update on Arena. BFL said the model launched last week and is free in the playground through 2026-08-16 23:59 PT.
Leaderboard position: Arena Text-to-Video, dated 2026-08-10, lists 607,046 votes and 44 models. #1 is gemini-omni-flash at 1512±11, #2 is flux-3-video at 1496±17 Preliminary, #3 is seedance-2.0 at 1478±9, #4 is muse-video at 1457±15, and #6 is happyhorse-1.0 at 1428±13.
Capability change: Arena X says FLUX 3 Video has reached #2 in Text-to-Video Arena, only 16 points behind Gemini Omni Flash. BFL says 4K, video editing, and multiple reference images are coming next.
Access: Arena Text-to-Video testing and the BFL playground's limited free window. API, pricing, and the timing for 4K/editing features still need official documentation.
Impact: Video-model competition is shifting from "new model name" to measurable Elo, editing workflows, and free trial windows. Gemini Omni Flash remains #1, but FLUX 3 Video is now in the priority retest zone. HappyHorse 1.0 remains in the Top 10 and should stay on the watchlist.
Sources: Arena Text-to-Video, Arena X, BFL X.

4. Mistral Regional Inference + Third-Party Open Models: Platform Availability Update

Event type: Official release / platform update / regional compliance and inference access / not a new base model.
First/official/availability time: Mistral's official news page is dated 2026-08-11. @MistralAI posted on X the same day.
Capability change: Mistral regional endpoints are generally available, letting customers choose Europe or US regions, with inference and processing kept in the selected region. Priority Tier is in public preview, with committed service levels, custom rate limits, and SLAs. Mistral's platform is also starting to support third-party open models, beginning with Z.ai GLM-5.2.
Access: Mistral Platform regional endpoints, Priority Tier, and third-party open model hosting.
Impact: This is not a model-capability leap, but it matters for enterprise deployment. It shows European sovereign AI, regional inference, open-model hosting, and compute partnerships converging. Buyers should verify exact regions, SLAs, model matrix, pricing, and logging/data-transfer terms.
Sources: Mistral official news, MistralAI X.

5. Luma Scenes and Runway Seedance 2.5: Video Generation Moves Toward Controllable Workflows

Event type: Official release / platform access update / X first signal / not a single foundation-model launch.
First/official/availability time: Luma's official news page is dated 2026-08-11. @LumaLabsAI posted on X the same day. Runway announced on 2026-08-11 that Seedance 2.5 is live on Runway.
Capability change: Luma Scenes, powered by Uni-1, can turn a brief or reference images into a sequenced storyboard, let users refine scenes and timing, approve keyframes, then render with Ray 3.2 or Seedance 2 in 9:16 or 16:9. Runway's Seedance 2.5 access supports up to 50 character references, clips up to 30 seconds, and music sync.
Access: Luma Scenes at https://app.lumalabs.ai; Seedance 2.5 on Runway.
Impact: Both updates point to the same trend: video-generation products are being differentiated by storyboard control, character consistency, timing, music sync, and editable export, not only one-shot prompt output. Follow-up should check APIs, pricing, copyright/commercial terms, and multi-reference stability.
Sources: Luma Scenes official markdown, Luma X, Runway X.

6. Anthropic Claude Content Marking: Model-Level Watermarks and C2PA, a Governance Update

Event type: Official documentation update / X lead / compliance and provenance marking / not a new model release.
First/official/availability time: Anthropic's Claude Help Center page was updated around 2026-08-11. @xiaohu surfaced the Chinese lead on X the same day.
Capability change: Anthropic says Claude models launched in the EU on or after 2026-08-02 will support machine-readable marking from launch. Generated text will carry embedded watermarks, and supported generated files will include signed provenance metadata. Marks apply across Claude Platform/API, Claude, Claude Code, Claude Cowork, Claude Tag, and supported cloud partner surfaces, and Anthropic says it will apply them worldwide.
Limitations: Anthropic explicitly says a detected Claude mark only indicates that the content may have been processed by Claude, not full provenance. Heavy editing, translation, mixing, short text, format conversion, or stripped metadata can all affect detection.
Access: The current source is Anthropic's official explanation and its promise of future detection documentation. Third-party detection tools and technical details are still pending.
Impact: This matters for publishing, education, UGC platforms, and enterprise content audits. Developers should include Claude output watermarks/metadata in product transparency and compliance documentation, but should not claim that watermark detection fully proves Claude authorship.
Sources: Anthropic Help Center, Xiaohu X, XiaoHu.AI.

Arena Watch

Text-to-Video Arena: The 2026-08-10 leaderboard has 607,046 votes and 44 models. Top 6: gemini-omni-flash #1 1512±11, flux-3-video #2 1496±17 Preliminary, seedance-2.0 #3 1478±9, muse-video #4 1457±15, seedance-2.5 #5 1439±11, and happyhorse-1.0 #6 1428±13. Today's important change is Arena/BFL explicitly pushing the FLUX 3 Video update to #2. Sources: Text-to-Video Arena, Arena X.
Text-to-Image Arena: The 2026-08-10 leaderboard has 5,919,946 votes and 77 models. The Top 9 continue yesterday's structure: gpt-image-2 (medium) #1 1381±5, mai-image-2.6-preview #2 1336±11, grok-imagine-image-2.0 (low) #3 1316±12 Preliminary, reve-2.1 #4, muse-image #5, and gemini-3.1-flash-image (nano-banana-2) [web-search] #7. nano-banana-2 remains a Gemini image codename to track, but there was no new official launch in this window. Source: Text-to-Image Arena.
Text Arena: nvidia-nemotron-3.5-lightning-30b-a3b-nvfp4 is now on the Text leaderboard at #203, 1345±12 Preliminary, with 2,765 votes. This should not be confused with NVIDIA's agent/terminal positioning: it is not a top general chat model. Source: Text Arena.
Artificial Analysis: Nemotron 3.5 Lightning has an Intelligence Index of 24, Terminal-Bench v2.1 24%, and GDPval-AA v2 Elo in the roughly 824/865 range, clearly ahead of Nemotron 3 Nano but still positioned as a small-active-parameter execution model. The new AA-AnalystAgent leaderboard shows closed models still leading, with open weights about 15 points behind the frontier. Sources: AA Nemotron X, AA Benchmark X.
Arena changelog: The public changelog can lag X/leaderboard changes. FLUX 3 Video's update and Nemotron's listing were cross-checked through the live leaderboard and X rather than relying on changelog alone.

X Watch

@NVIDIAAI: Announced NVIDIA Nemotron 3.5 Lightning, emphasizing open/customizable materials, weights/data/recipes, 30B MoE with 3B active, 4x output speed, and agentic workflows. Credibility: high, official NVIDIA source. Sources: launch X, open materials X.
@ArtificialAnlys: Posted Nemotron 3.5 Lightning evaluation first, then the AA-AnalystAgent benchmark. Credibility: high, official AA source, useful as a secondary benchmark source. Sources: Nemotron X, AA-AnalystAgent X.
@arena / @bfl_ai: FLUX 3 Video update entered Text-to-Video at #2. BFL says the free playground is open until Sunday the 16th at 11:59 PT and previews 4K, video editing, and multiple references. Credibility: high, official Arena/BFL sources; API and technical details still need official docs. Sources: Arena X, BFL X.
@MistralAI: Announced the regional inference/open models/new compute update, focused on sovereignty, regional endpoints, Priority Tier, and third-party open models. Credibility: high, official Mistral source. Source: Mistral X.
@LumaLabsAI: Released Luma Scenes powered by Uni-1. Credibility: high, official Luma source. Source: Luma X.
@runwayml: Announced Seedance 2.5 live on Runway, with 50 character references, clips up to 30 seconds, and music sync. Credibility: high, official Runway source; event type is platform access update. Source: Runway X.
@GeminiApp: Posted a 1B monthly Gemini users signal and a reminder that Gemini Omni's 10 free videos promotion runs through 2026-08-11 23:59 PT. Credibility: high, official Google/Gemini source, but not a new model release. Sources: 1B users X, Omni promotion X.
@grok: Released Grok Bot beta, positioned as AI teammates that can sign in to tools and do work. Credibility: high, official Grok source. Event type is agent/product beta, not a new base model. Source: Grok Bot X.
@xiaohu: The most useful lead in this window was Anthropic's Claude content marking/watermark update. It was verified against Anthropic's Help Center. Credibility: useful as a Chinese lead source; final facts should follow official documentation. Sources: Xiaohu X, Anthropic Help Center.
@Alibaba_Qwen: No actual new weight release was observed in this window. The ModelScope Qwen3.8-2.4T-A95B page still shows pre-release and a planned release time. Credibility: the page exists, but availability must wait for actual repository files. Sources: ModelScope, Qwen3.8-Max X preview.
@GoogleDeepMind / @AnthropicAI / @AIatMeta / @cohere / @Kling_ai: No new verifiable foundation-model launch was observed in this window. Relevant account activity was research, product, event, or previous-release continuation.

Watchlist

Qwen3.8-Max / Qwen3.8-2.4T-A95B: The ModelScope pre-release page lists 2026-08-12 15:00 UTC as the planned release time. After 23:00 HKT, verify whether weights are actually downloadable, the license, whether Hugging Face is synced, whether Qwen3.8-27B is released at the same time, and whether API/model IDs changed.
Nemotron 3.5 Lightning: Test single-GPU/local throughput, 1M-context stability, OpenMDW 1.1 commercial terms, NVFP4 quality loss, engineering maturity of DFlash/DSpark/MTP speculative decoding, and real NeMo Switchyard routing gains.
AA-AnalystAgent: Review the full task set, scoring methodology, stirrup reproducibility steps, spreadsheet/document parsing biases, and whether open-weight rankings remain stable across tool harnesses.
FLUX 3 Video: Confirm whether the Arena update exactly matches BFL's production version. 4K, video editing, multiple references, API access, pricing, and commercial rights still need official documentation.
Luma Scenes: Separate Uni-1's storyboard/scene-planning role from Ray 3.2/Seedance 2 rendering. API access, collaboration features, pricing, rights, and output editability still need confirmation.
Runway Seedance 2.5: Confirm whether this is ByteDance/Seedance 2.5 platform access or a Runway-customized version. The price and failure rate for 50 character references, 30-second clips, and music sync need testing.
Anthropic Claude marking: Wait for technical detection documentation and third-party detection tools. The current official position supports "machine-readable marks exist," not "a detection result fully proves Claude original authorship."
Gemini Omni: Recheck after the 2026-08-11 23:59 PT promotion cutoff to see whether the 10 free videos campaign is extended, converted to paid use, or has a quota change.