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

AI Model Daily - 2026-08-03

2026-08-03 AI model news: no major new foundation-model launch in the window; the main signals are DeepSeek-V4-Flash-0731 pricing and access, Kimi K3 developer integration, Google TPU 8i/8t infrastructure, and Arena Image-to-WebDev carryover.

30-Second Summary

  • I did not find a new foundation-model launch from OpenAI, Anthropic, Google DeepMind, Meta, Mistral, Qwen, Kimi, Runway, Luma, or Kling inside this window. The main story is distribution, pricing, and leaderboard signal around recently released models.
  • DeepSeek-V4-Flash-0731 is today's strongest signal: Nous Research announced a 7-day 90% off promotion on Nous Portal, Ollama Cloud says the same model is now 2x+ faster than the previous day, and @xiaohu amplified Artificial Analysis cost comparisons.
  • Kimi K3 continues moving from launch into developer adoption. Together AI published a complete developer guide highlighting 2.8T open weights, 1M context, OpenAI-compatible API access, benchmarks, and pricing. This is documentation/access maturation, not a new model launch.
  • Google TPU 8i / TPU 8t kept circulating in Chinese and Japanese AI circles. Official materials position 8i for low-latency inference and 8t for training, with 8i offering up to 80% better performance per dollar versus Ironwood for low-latency MoE serving targets.
  • Arena's near-window highlight remains Image-to-WebDev: the official update landed slightly before this window but continued to circulate. Opus 5 Max leads, while GPT-5.6 Sol, Grok-4.5, Kimi K3 Max, and Muse Spark 1.1 are now compared in the same track.

Must Watch

1. DeepSeek-V4-Flash-0731: from model release into ultra-low-cost agent adoption

  • Event type: Official release / availability update / X first signal / third-party benchmark verification.
  • First / official / available time: DeepSeek / Hugging Face made DeepSeek-V4-Flash-0731 public around 2026-08-01. Ollama Cloud announced availability on 2026-08-01 and said on 2026-08-02 that it is now 2x+ faster than the previous day. Nous Research announced on 2026-08-02 16:29 UTC that it is partnering with Novita Labs to offer the model at 90% off on Nous Portal for 7 days.
  • Capability change: The Hugging Face model card describes DeepSeek-V4-Flash-0731 as the official release of DeepSeek-V4-Flash, with enhanced agentic capability. It lists Terminal Bench 2.1 82.7, NL2Repo 54.2, Cybergym 76.7, DeepSWE 54.4, and Toolathlon-Verified 70.3. Artificial Analysis gives it an Intelligence Index score of 50, 10 points above the previous DeepSeek V4 Flash, and records Terminal-Bench 2.1 at roughly 79%.
  • Availability: Hugging Face weights, Ollama Cloud, Nous Portal, Novita Labs / OpenRouter, DeepInfra, and other third-party serving paths.
  • Impact: The key point today is not that the model launched today, but that it entered broad price and access distribution. If the Nous Portal discount and Ollama Cloud speedup hold up in real use, DeepSeek-V4-Flash-0731 becomes a priority candidate for low-cost coding agents, translation pipelines, long-context batch processing, and hybrid local/cloud inference.
  • Sources: Hugging Face model card, Artificial Analysis article, Ollama model page, Ollama X, Nous Research X, Novita X.

2. Nous Portal / Novita discount: DeepSeek V4 Flash 0731 becomes a pricing event

  • Event type: Availability update / X first signal / pricing signal requiring verification.
  • First / official / available time: @NousResearch posted the discount on 2026-08-02 16:29 UTC. @novita_labs confirmed the partnership inside the same window.
  • Capability change: Nous Research says DeepSeek V4 Flash 0731 is 90% off on Nous Portal for the next 7 days, and claims that at the discounted price it is over 1000x cheaper than Fable 5 on comparable tasks while still beating it on Terminal-Bench 2.1. Novita Labs says it is partnering with Nous to bring the model to more builders. In another reply, Novita says the OpenRouter DeepSeek-V4-Flash-0731 listing is the same model, and that "Optimized for Codex" means it was post-trained for Codex-style agent workflows.
  • Availability: Nous Portal, Novita Labs, and OpenRouter provider routing.
  • Impact: The promotion price and the long-term cost curve must be separated. The promotion will drive short-term agent trials and benchmark screenshots, but production choice still depends on post-promotion price, rate limits, stability, cache pricing, and provider transparency.
  • Sources: Nous Research X, Nous Portal, Novita X - partnership, Novita X - same model / Codex optimized, OpenRouter DeepSeek V4 Flash 0731.

3. Google TPU 8i / 8t: training and inference split becomes an infrastructure signal

  • Event type: Official release / X circulation / infrastructure signal.
  • First / official / available time: Google Cloud's official technical deep dive predates this window. @googlecloud resurfaced the TPU 8i / 8t video on 2026-08-01, and @xiaohu plus several Chinese/Japanese accounts continued to amplify it inside the 2026-08-02 window.
  • Capability change: Google Cloud positions TPU 8t for large-scale training and TPU 8i for low-latency inference/serving. For low-latency large-MoE inference targets, TPU 8i offers up to 80% better performance per dollar versus Ironwood TPU. Both chips emphasize up to 2x performance per watt.
  • Availability: Google Cloud TPU 8i / 8t paths; exact customer availability cadence still needs follow-up.
  • Impact: This is not a model release, but it directly affects future Gemini, agent serving, and low-latency MoE inference economics. The important architectural shift is that Google is no longer asking one TPU line to cover both training and serving; it is optimizing those economic profiles separately.
  • Sources: Google Cloud technical deep dive, Google infrastructure blog, Google Cloud X, xiaohu X.

4. Kimi K3: Together AI guide marks the shift from launch into integration

  • Event type: Official/partner documentation update / availability update / media coverage.
  • First / official / available time: The official Kimi K3 release predates this window. Together AI posted Kimi K3: The Complete Developer Guide inside the 2026-08-02 window and highlighted 2.8T parameters, 1M context, and OpenAI-compatible API support on X.
  • Capability change: Together AI's model page describes Kimi K3 as Moonshot AI's strongest model, with 2.8T total parameters, 16 of 896 experts activated per token, native vision, 1M context, Kimi Delta Attention, and Attention Residuals. The developer guide covers API usage, benchmarks, pricing, an ROI calculator, and production model selection.
  • Availability: Together AI API / Playground, Kimi official access, Hugging Face weights, and other third-party inference providers.
  • Impact: Kimi K3 did not gain new model capabilities today, but developer documentation lowers the testing and adoption bar. Kimi K3 should remain in the high-priority open-weight watchlist, especially for long context, vision-to-code, agentic coding, and local/cloud cost comparison.
  • Sources: Together AI guide, Together AI Kimi K3 model page, Together X, Kimi K3 official blog, Kimi K3 Hugging Face.

5. Gemini 3.6 Flash / 3.5 Flash-Lite: not today's launch, but Gemini Drops keeps expanding reach

  • Event type: Official release / product availability update / X carryover.
  • First / official / available time: Google launched Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber on 2026-07-21. The Gemini Drops page remains visible in this window, and @GeminiApp continued promoting Gemini Drops near the window.
  • Capability change: Google describes Gemini 3.6 Flash as the workhorse model in the Gemini 3 series, aimed at coding, knowledge work, multimodal tasks, and better token efficiency. Gemini Drops shows Gemini Spark broader availability, macOS voice mode, Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, Nano Banana digital avatars, and personalized image features.
  • Availability: Gemini, Google AI Studio, Gemini API, and related Google product surfaces.
  • Impact: This should not be treated as a new model release today. Its significance is that Google is pushing Flash models, Spark agents, and Nano Banana product entry points deeper into Gemini, putting pressure on low-cost routes from DeepSeek, Luna, and Kimi.
  • Sources: Google Gemini models blog, Gemini Drops, Gemini 3.6 Flash model card, GeminiApp X.

Arena Watch

  • Artificial Analysis / DeepSeek V4 Flash 0731: AA gives DeepSeek V4 Flash 0731 an Intelligence Index score of 50, 10 points above the older DeepSeek V4 Flash. It records GDPval-AA v2 Elo 1559, Terminal-Bench 2.1 at roughly 79%, and τ³-Bench Banking around 31%. The official Hugging Face card lists Terminal Bench 2.1 as 82.7, so reports should preserve source differences. Sources: AA article, HF model card.
  • Artificial Analysis / cost signal: @xiaohu amplified AA's cost breakdown inside the window, saying DeepSeek completed the same benchmark task at 105x lower cost than Fable. Treat this as "AA cost model plus Chinese auxiliary source"; production decisions still need provider-specific recalculation for input, output, cache, and routing prices. Sources: xiaohu X, AA DeepSeek model page.
  • Arena / Image-to-WebDev: Arena's official update landed slightly before this window, but it continued to circulate. The official post says Opus 5 Max is #1 in Image-to-WebDev, GPT-5.6 Sol xHigh #4, Grok-4.5 #5, Kimi K3 Max #6, Muse Spark 1.1 #8, GPT-5.6 Terra xHigh #13, and GPT-5.6 Luna xHigh #21. Sources: Arena X, Arena Code WebDev leaderboard.
  • LMArena / other tracks: I did not see a new high-confidence foundation-model launch from the LMArena/Arena accounts inside this window. The visible focus remains the recent Image-to-WebDev update and the AutoEval method change from the previous few days. Sources: Arena leaderboards, Arena changelog.
  • Conclusion: There is no new leaderboard shock today. DeepSeek-V4-Flash-0731 is the cost/agentic benchmark thread; Image-to-WebDev is near-window carryover; Kimi K3 remains a high-priority open-weight model to watch.

X Watch

  • @xiaohu: Inside the window, @xiaohu amplified Google's TPU 8i/8t training/inference split and reported replacing website translation with DeepSeek V4 Flash, reducing one translation task from 20-60 minutes to about 3 minutes at less than roughly RMB 0.10 per article. @xiaohu also amplified AA's cost breakdown. Credibility: useful Chinese auxiliary source, but should be verified with official/AA/provider pages. Sources: TPU 8i/8t, DeepSeek translation test, AA cost reply.
  • @NousResearch: Announced DeepSeek V4 Flash 0731 at 90% off on Nous Portal for 7 days, claiming 1000x+ lower cost than Fable 5 on comparable tasks. Credibility: high as an official access/promotion signal; long-term cost requires follow-up. Source: Nous X.
  • @ollama: Said DeepSeek-V4-Flash-0731 is 2x+ faster than the previous day on Ollama Cloud. Credibility: high as an official access signal, but speed should be retested by task, context, concurrency, and region. Source: Ollama X.
  • @novita_labs: Confirmed the Nous partnership and said the OpenRouter DeepSeek-V4-Flash-0731 entry is the same model; "Optimized for Codex" means post-training for Codex-style agent workflows. Sources: Novita partnership, Novita same model.
  • @togethercompute: Posted the Kimi K3 developer guide. This is an integration/documentation signal, not a Kimi K3 launch today. Sources: Together X, Together guide.
  • @GeminiApp: Only low-information interaction was visible inside the strict window; near-window Gemini Drops content continues to promote Gemini 3.6 Flash, 3.5 Flash-Lite, Spark, and Nano Banana digital avatars. Source: Gemini Drops.
  • @OpenAI, @OpenAIDevs, @AnthropicAI, @GoogleDeepMind, @GoogleAI, @xai, @grok, @AIatMeta, @MistralAI, @Alibaba_Qwen, @Kimi_Moonshot, @deepseek_ai, @runwayml, @LumaLabsAI, @Kling_ai, @TencentHunyuan, @SakanaAILabs, @cohere: no new high-confidence foundation-model launch was visible inside this window. @grok search results were mostly automated replies and are not relevant to the model-daily main thread.

Pending Confirmation

  • DeepSeek-V4-Flash-0731 official release timing, weight version, and provider consistency: Hugging Face, Ollama Cloud, OpenRouter/Novita, Nous Portal, and DeepInfra names and post-training notes need side-by-side verification.
  • DeepSeek cost claims need recalculation: Nous 90% off is a 7-day promotion; OpenRouter/Novita, official DeepSeek China, cache read/write, reasoning tokens, Responses API, and Codex-style agent workflows may have different pricing semantics.
  • Ollama Cloud's 2x+ speedup needs retesting with region, model tag, context length, prompt cache, concurrency, and speculative-decoding settings recorded.
  • Google TPU 8i/8t: official materials predate this window, but X recirculated them inside it. Confirm 2026 customer availability, Cloud SKUs, Gemini internal training/serving usage, and whether it changes external Gemini API pricing.
  • Kimi K3: Together's guide lowers integration friction, but latency, ZDR, context limits, vision input, and real cost still need comparison across Together, Kimi official, OpenRouter, Fireworks, Nebius, and other providers.
  • Gemini 3.6 Flash / 3.5 Flash-Lite: the official launch is outside this window. Confirm whether Gemini Drops availability is consistent across Gemini consumer UI, AI Studio, API, and Antigravity, and benchmark Flash-Lite on high-throughput tasks.
  • Arena Image-to-WebDev: the official post is just outside the strict window, so this is treated as carryover. Reopen the public leaderboard to confirm whether the X-post scores, including Opus 5 Max 1669, have changed.