QName AI logoQName AI
All AI model reports
Collected in Hong Kong time

AI Model Daily - 2026-08-09

2026-08-09 AI model news: Grok Imagine Image 2.0 is the clearest official release and leaderboard change, while Ling 3.0 Flash and Seedance 2.5 show important open-weight and platform-availability signals.

30-Second Summary

  • Grok Imagine Image 2.0 is today's clearest official release and leaderboard change. xAI released it on 2026-08-07, and it is now available as the new Quality Mode in Grok Imagine on grok.com/imagine, iOS, and Android. Arena added grok-imagine-image-2.0 (low) to Text-to-Image and Image Edit on the same day, and it ranks #2 in both tracks.
  • Ling 3.0 Flash continues to gain open-weight and performance momentum. Hugging Face now has inclusionAI/Ling-3.0-flash; BusinessWire/Ant Group describes it as 124B total parameters with 5.1B active parameters; Artificial Analysis reports an Intelligence Index of 38 and output speed around 373.6 t/s.
  • Seedance 2.5 is showing broader availability inside creator tools. @xiaohu highlighted Higgsfield/Topview launch and promotion signals; Topview's pricing page shows 60 days of Seedance 2.5 720p unlimited access on Ultra/Team annual plans, and Renoise has an Advanced-plan campaign from 2026-08-07 to 2026-08-14. This is a platform-access update, not a new official ByteDance model launch today.
  • Artificial Analysis v4.1.1 changes how recent model comparisons should be read. AA upgraded graders and tau3-Banking, said overall rankings remain broadly stable, and its 2026-08-07 changelog adds provider results for Ling 3.0 Flash, Vidu Q3 Turbo, MiniMax H3, and others.
  • Text Arena still shows strong signals from Meta muse-spark-1.2 (xHigh) and qwen3.8-max: Muse Spark 1.2 ranks #4 and Qwen3.8-Max ranks #6. I did not find a new official foundation-model release from OpenAI, Anthropic, or Google/Gemini in this window.

Must Watch

1. Grok Imagine Image 2.0: xAI Releases a New Image Generation and Editing Model, #2 on Both Arena Image Boards

  • Event type: Official release / Arena appearance / X first signal / access update.
  • First/official/availability time: xAI's official news page is dated 2026-08-07. Grok's official X account posted the launch, and Arena's changelog records grok-imagine-image-2.0 (low) being added to Text-to-Image and Image Edit on the same date.
  • Capability change: xAI positions Image 2.0 for real creative work, emphasizing more precise instruction following, typography and layout planning, sharper small text, and better preservation of input elements across generations and edits. Editing capabilities include magic-wand regional editing, segmentation, background removal, multi-reference editing with up to 5 input images, smart resize, and workflow templates.
  • Leaderboard position: On Arena's 2026-08-07 Text-to-Image leaderboard, gpt-image-2 (medium) is #1 at 1380±5, grok-imagine-image-2.0 (low) is #2 at 1320±12 Preliminary, and reve-2.1 is #3. On Image Edit, gpt-image-2 (medium) is #1 at 1463±4, grok-imagine-image-2.0 (low) is #2 at 1439±8 Preliminary, and muse-image is #3.
  • Access: Grok.com/imagine, Grok iOS, and Grok Android. The official page explicitly says API access is coming soon. xAI's Imagine API page is still a general image/video API page, so it should not be treated as proof that Image 2.0 API access is already live.
  • Impact: This is a strong signal for image production workflows, especially ads, product shots, posters, UI/game assets, batch edits, and text-heavy visuals. It should enter internal image generation and editing tests against gpt-image-2, muse-image, nano-banana-2, and qwen-image-3.0-pro, with focus on text rendering, local edit stability, and commercial-asset usefulness.
  • Sources: xAI official news, Grok Imagine, Arena changelog, Text-to-Image Arena, Image Edit Arena, Grok X, Arena X.

2. Ling 3.0 Flash: Ant Group/InclusionAI Open Weights, Strong Cost-Performance Signal From AA

  • Event type: Official release / open weights / public benchmark update / X first signal.
  • First/official/availability time: Ant Group/BusinessWire's release is dated 2026-07-26. Hugging Face now hosts inclusionAI/Ling-3.0-flash. Artificial Analysis posted an Intelligence Index of 38 on 2026-08-05 and added DeepInfra provider performance results on 2026-08-07. @ArtificialAnlys said on the same day that Ant Group released Ling 3.0 Flash.
  • Capability change: The model has 124B total parameters and 5.1B active parameters per token. It uses native hybrid-linear attention, alternating KDA and MLA layers at a 5:1 ratio, combined with sparse MoE. AA lists it as an open weights model with a 262k context window, text input/text output, a reasoning version, about $0.075/M input and $0.22/M output pricing, and speed around 373.6 t/s.
  • Access: Hugging Face inclusionAI/Ling-3.0-flash. AA's provider page tracks InclusionAI and DeepInfra: InclusionAI outputs around 373.6 t/s, DeepInfra around 358.5 t/s, and both show blended price around $0.05/M.
  • Impact: Ling 3.0 Flash is one of the most relevant open-weight candidates for low-cost reasoning and agentic batch workloads in this cycle. Its large total parameter count but low active parameter count make it worth testing for long-context summarization, bulk extraction, lightweight tool use, and mixed Chinese/English tasks. Before production use, verify license and commercial-use terms, provider reasoning defaults, JSON/function-calling behavior, and verbosity-driven cost.
  • Sources: Hugging Face model page, Ant Group/BusinessWire, AA Ling 3.0 Flash, AA Ling providers, AA changelog, Artificial Analysis X.

3. Seedance 2.5: Platform Access and Promotions Spread, But Keep Model Launch Separate From Tool Launch

  • Event type: X first signal / access update / media and platform signal / rumor to verify.
  • First/official/availability time: Seedance 2.5 did not first launch in today's window. The current change is mostly platform launch and promotion information from around 2026-08-08. Within the last 12 hours, @xiaohu highlighted Higgsfield's Seedance 2.5 launch, describing 30-second single-pass scenes, 50 references, production-ready editing, and 33 days of zero credit cost. About 19 hours ago, he also reposted Topview pricing and promotion information.
  • Capability change: Renoise's Seedance 2.5 page describes 4-30 second videos, 480p/720p, up to 50 multimodal references, and timestamp editing. Topview's pricing page shows Seedance 2.5 720p, 30s per generation, and 60 Days Unlimited on Ultra/Team plans. Higgsfield's official blog says Seedance 2.5 can be selected in week two of its Unlimited program, while also emphasizing that unlimited access differs from credit mode and depends on plan eligibility, queue behavior, and resolution limits.
  • Access: Higgsfield, Topview, Renoise, and other third-party creator platforms. The exact ByteDance/BytePlus official API model ID, region availability, pricing, and quota still need separate verification.
  • Impact: This matters more to video creators than the phrase "official model launch," because entry price, concurrency, unlimited windows, and reference limits directly affect production cost. The daily should continue to mark this as a platform-availability update and avoid framing third-party promotions as an official new model release.
  • Sources: xiaohu X, xiaohu reposting Topview, Higgsfield Seedance Unlimited, Topview pricing, Renoise Seedance 2.5, Higgsfield X.

4. Artificial Analysis: v4.1.1 and 2026-08-07 Provider Updates Affect Cross-Model Selection

  • Event type: Public benchmark/methodology update / provider performance update.
  • First/official/availability time: AA's v4.1.1 article is dated 2026-08-06. The AA changelog added DeepInfra Ling 3.0 Flash, Novita Ling 3.0 Tiny, Together AI Qwen3.5 9B Non-reasoning, Vidu Q3 Turbo, MiniMax H3, and other provider performance results on 2026-08-07.
  • Capability change: v4.1.1 upgrades the graders for HLE, AA-LCR, and AA-Omniscience, and moves tau3-Banking to upstream v1.0.1. AA says overall rankings remain broadly consistent, Claude Opus 5 remains #1, most models move by less than 1 point, and the largest increase is Muse Spark 1.2 (xHigh) at +2.7.
  • Access: Artificial Analysis changelog, the Intelligence Index v4.1.1 article, and model provider pages.
  • Impact: Today should not be read only as a "new model name" day. New provider coverage and methodology changes also matter. For Ling, MiniMax H3, Vidu Q3 Turbo, and smaller Qwen models, provider pages are closer to engineering selection than a single aggregate score.
  • Sources: AA changelog, AA v4.1.1 article, AA Ling providers, AA Qwen3.8 Max.

5. Muse Spark 1.2 and Qwen3.8-Max: High Text Arena Positions Continue, No New Launch Today

  • Event type: Arena appearance / official release carryover / X distribution.
  • First/official/availability time: Meta officially released Muse Code beta and Muse Spark 1.2 on 2026-08-05. Qwen3.8-Max entered multiple leaderboards around 2026-08-03/08-04. This window is a continuation of leaderboard and X distribution, not a new first launch.
  • Capability change: The 2026-08-06 Text Arena ranks muse-spark-1.2 (xHigh) #4 at 1498±13 and qwen3.8-max #6 at 1497±9 Preliminary, placing both in the top competitive cluster. AA's Qwen3.8-Max page shows Intelligence Index 58, text/image/video input, text output, 1M context, $2/M input, and $6/M output.
  • Access: Muse Spark 1.2 through Muse Code and Meta Model API. Qwen3.8-Max through Alibaba/Qwen-related access points, Arena, and parts of the tooling ecosystem.
  • Impact: Muse Spark 1.2 remains worth tracking for coding-agent and game-development workflows. Qwen3.8-Max remains worth tracking for multimodal input and long-context web/frontend tasks. Today's point is sustained high placement, not a new model release.
  • Sources: Meta Muse official blog, Text Arena, AA Qwen3.8 Max, AI at Meta X, Alibaba_Qwen X.

Arena Watch

  • Text-to-Image Arena: The 2026-08-07 leaderboard has 5,895,429 votes and 76 models. The top 10 are gpt-image-2 (medium) #1 at 1380±5, grok-imagine-image-2.0 (low) #2 at 1320±12 Preliminary, reve-2.1 #3 at 1302±8, muse-image #4 at 1283±7, reve-2.0 #5, gemini-3.1-flash-image (nano-banana-2) [web-search] #6 at 1263±5, qwen-image-3.0-pro #7 at 1258±10, seedream-5.0-pro #8, mai-image-2.5 #9, and gemini-3.1-flash-lite-image (nano-banana-2-lite) #10. Source: Text-to-Image Arena.
  • Image Edit Arena: The 2026-08-07 leaderboard has 28,831,297 votes and 53 models. gpt-image-2 (medium) is #1 at 1463±4, grok-imagine-image-2.0 (low) is #2 at 1439±8 Preliminary, muse-image is #3 at 1405±6, mai-image-2.5 is #4, seedream-5.0-pro is #5, and gemini-3.1-flash-image (nano-banana-2) [web-search] is #10 at 1385±4. Source: Image Edit Arena.
  • Arena changelog: On 2026-08-07, Arena added grok-imagine-image-2.0 (low) to Text-to-Image and Image Edit. On 2026-08-06, it added Mistral Medium 3.5 to Agent Arena. On 2026-08-05, it added claude-opus-5-max to Fullstack Code Arena. On 2026-08-04, it added DeepSeek V4 Flash High, Qwen Image 3.0 Pro, Qwen3.8-Max, and others. Source: Arena changelog.
  • Text Arena: The 2026-08-06 leaderboard has 7,649,050 votes and 386 models. claude-fable-5 is #1 at 1507±6, claude-opus-4-6-thinking is #2, claude-opus-4-7-thinking is #3, muse-spark-1.2 (xHigh) is #4 at 1498±13, claude-opus-4-6 is #5, and qwen3.8-max is #6 at 1497±9 Preliminary. Source: Text Arena.
  • Agent Arena: The 2026-08-06 leaderboard has 1,665,514 sessions and 46 models. Claude Opus 5 (High) is #1 with net improvement 11.99%±1.37%, followed by Claude Fable 5 (High) #2, Claude Opus 5 (Max) #3, GPT 5.6 Sol (xHigh) #4, and Kimi K3 (Max) #5. Deepseek V4 Flash (High) (20260731) is #21 with confirmed success 8.53%±2.54%. Mistral Medium 3.5, newly added to Agent Arena, appears with 4,844 sessions. Source: Agent Arena.
  • Artificial Analysis: The 2026-08-07 changelog adds provider performance results for Ling 3.0 Flash/DeepInfra, Ling 3.0 Tiny/Novita, Qwen3.5 9B/Together AI, Vidu Q3 Turbo, MiniMax H3, and others. After the 2026-08-06 v4.1.1 methodology update, Claude Opus 5 remains #1 and Muse Spark 1.2 (xHigh) is one of the largest upward movers. Sources: AA changelog, AA v4.1.1.

X Watch

  • @grok / @arena: Grok officially released Image 2.0, and Arena confirmed grok-imagine-image-2.0 (low) landing on both image leaderboards. Arena also said this version moved from #14 to #2 relative to Grok Imagine Image Quality. Credibility: high for official release and Arena placement; Arena scores are dynamic, and the relationship between the low track and the full consumer Quality Mode serving config still needs follow-up. Sources: Grok X, Arena X.
  • @ArtificialAnlys: Posted the Ling 3.0 Flash 124B open-weights signal, with AA Intelligence Index 38, and framed it on the Intelligence vs Total Parameters Pareto frontier. It also posted expanded Text-to-Image Arena capability dimensions, including UI/UX Design, Live-Action Film, Physics, and Text Rendering. Credibility: high for AA official data; provider performance is point-in-time. Sources: Ling X, AA T2I expansion X.
  • @xiaohu: Highlighted Seedance 2.5 availability and promotions on Higgsfield/Topview, which is useful for creator-tool selection. He also mentioned NextSlide announcing an OpenAI acquisition, but I did not find an OpenAI official confirmation in this run, and it is not a model release. Credibility: useful as an early signal; acquisitions, official APIs, and pricing require second-source verification. Sources: Seedance 2.5/Higgsfield X, Topview X, NextSlide X, XiaoHu.AI Daily.
  • @AIatMeta: Continued distributing Muse Code beta and Muse Spark 1.2, focused on long-horizon software engineering, large repos, multi-file changes, and persistent sub-agents. Credibility: high for Meta official messaging; no new version number in this window. Sources: AI at Meta X, Meta official blog.
  • @Alibaba_Qwen: Continued showing Qwen3.8-Max multimodal/vision capabilities and leaderboard distribution. AA still marks Qwen3.8-Max as a proprietary model, released August 2026. If open weights land, the status should be changed only after a Qwen official weights page appears. Credibility: high for official product distribution; open-weight status should be based on model pages and weight repositories. Sources: Alibaba_Qwen X, AA Qwen3.8 Max.
  • OpenAI / Anthropic / Google DeepMind/Gemini / Mistral / video-model companies outside xAI: I did not find a new official foundation-model release from OpenAI, Anthropic, or Google/Gemini in this window. Mistral's main new signal is Mistral Medium 3.5 being added to Agent Arena. Runway, Luma, Kling, and Cohere did not show a verified new model first launch in this run.

To Verify

  • Grok Imagine Image 2.0: Confirm API opening date, API model ID, price, rate limits, and how Arena's low variant maps to Grok consumer Quality Mode.
  • Grok Image 2.0 vs gpt-image-2: Run the same prompt set for text rendering, local edits, product shots, UI/game assets, and multi-reference editing. Do not rely only on aggregate Elo.
  • Ling 3.0 Flash: Verify Hugging Face weight license, commercial use, inference code, quantized variants, whether providers enable reasoning, and real JSON mode/function-calling compatibility.
  • Ling 3.0 Flash providers: DeepInfra and InclusionAI may differ in context, TTFT, reasoning token output, cache billing, and batching limits. AA speed and price should be retested on our own workloads.
  • Seedance 2.5: Confirm the official ByteDance/BytePlus API model ID, region availability, pricing, maximum 30-second video, 50 multimodal references, and timestamp editing at the API layer rather than only in third-party UIs.
  • Seedance 2.5 "Unlimited": Topview/Higgsfield/Renoise unlimited access is a plan or campaign entitlement, not official free unlimited API access. Track concurrency, queue behavior, resolution, commercial rights, and campaign end dates.
  • NextSlide/OpenAI acquisition: @xiaohu has the signal, but I did not find OpenAI official confirmation. Keep it as a product/company item to verify, not a model release.
  • Qwen3.8-Max open weights: AA currently marks it as proprietary. Change it to official open weights only if Qwen publishes weights through its blog, Hugging Face, ModelScope, or GitHub.
  • Muse Spark 1.2: Continue verifying Meta Model API docs and Muse Code product configuration, including pricing, context, and tool/subagent limits. A high leaderboard position does not mean it is better than Claude/GPT/Kimi for every coding workflow.
  • Artificial Analysis v4.1.1: Future reports should record the AA Index version after methodology updates to avoid mixing small score differences from v4.1 and v4.1.1.