NPG 魔法報紙
追蹤主題 · AI

開源模型

DeepSeek、Qwen、Llama 這些可以自己下載來跑的模型,決定你不付月費能用到多強的 AI,也是中美 AI 競賽最直接的比分板。

已追蹤 146 期,其中 128 期有新進展 · 最近更新 2026-11-08

2026-11-06 · 第 147 期 · 本期(2026-11-08)沒有新進展,以下是上一次的變化

Qwen 3.8-27B 更新版本提升精度

  1. 蘋果宣布 iOS 27.3 將在中國推出 Apple Intelligence，並串接阿里通義千問以符合當地監管規定。 · 鉅亨網 cnyes
主筆看法蘋果與阿里合作顯示中國市場監管對外企的影響力，Qwen 精度提升則反映開源模型技術仍在快速進步。

接下來看Qwen 4 的具體更新內容及效能表現

時間軸

  1. 2026-11-08
    本期沒有新進展
  2. 2026-11-07
    本期沒有新進展
  3. 2026-11-05 · 第 146 期
    Hydrogen-1.0 模型在科學研究中表現優異，成本更低。
    1. Hydrogen-1.0 模型在 FrontierScience Research 測試中表現優於 GPT-6 Astra，成本低 72%。 · Reddit ML / LocalLLaMA
    2. DeepSeek V5 模型預計將使用華為昇淵晶片進行預訓練。 · Reddit ML / LocalLLaMA
    3. Kimi-K3.1-Pro 模型現已開放權重。 · Reddit ML / LocalLLaMA
    主筆看法Hydrogen-1.0 的表現顯示開源模型在科學研究領域的競爭力增強，成本效益更佳，有助於更多研究者和開發者使用。
  4. 2026-11-04 · 第 145 期
    Qwen 3.8 27B 可以透過 Blender 和 Python 加速複雜數
    1. Qwen 3.8 27B 可以解釋和修改神經網絡的權重,顯示模型具有高度靈活性 · AI 訪談與觀點
    2. DeepSeek 推出創作者模式,讓開發者透過自然語言指令建立自訂外掛程式 · Yahoo 股市
    主筆看法Qwen 和 DeepSeek 的最新功能展現了開源模型在複雜任務中的潛力,但可能增加安全風險 [2][3]
  5. 2026-11-02 · 第 143 期
    Memdoor: A team of 5 Qwen3.8-27B models
    1. Memdoor: A team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard · Hacker News
    2. DeepSeek推出Super-agent架構！打造24小時自主AI代理、跨平台並行處理多項任務 · 鉅亨網 cnyes
    3. DeepSeek Harness team published the underlying architecture for their Super-agent, an always-on autonomous agent, as an open-source prototype on GitHub. · Reddit ML / LocalLLaMA
  6. 2026-11-01 · 第 142 期
    Gemma-4-31B-Humanlike-Chat: 2.0's improv
    1. Gemma-4-31B-Humanlike-Chat: 2.0's improved human voice on official Gemma 4 31B · Reddit ML / LocalLLaMA
    2. 墨西哥城熱鬧非凡 Kimi Antonelli衝向F1歷史一刻 · 體育賽事
    3. Betelgeuse: Running Qwen 3.8 Flash Next at 22 tok/s sustained (spiking to 48 tok/s) on an RTX 3060 12GB + 16GB DDR4 RAM 100% bit-exact! · Reddit ML / LocalLLaMA
  7. 2026-10-31 · 第 141 期
    Introducing DeepSeek-V4.1-Flash-1017 Dyn
    1. Introducing DeepSeek-V4.1-Flash-1017 Dynamic v3.1 Unsloth GGUFs · Reddit ML / LocalLLaMA
    2. Open-weight AI models can exploit the patch gap · Hacker News
    3. Mistral releases open weights for its 1 trillion-parameter Large 4 model · Reddit ML / LocalLLaMA
  8. 2026-10-30 · 第 140 期
    Mistral Large 4 (Le Chonk) is now Open W
    1. Mistral Large 4 (Le Chonk) is now Open Weight · Hacker News
    2. Qwenputer: A functioning, deterministic RISC-V CPU inside Qwen3.5 · Hacker News
    3. Apsara conference: Qwen4 next week · Reddit ML / LocalLLaMA
  9. 2026-10-29 · 第 139 期
    Someone built a 32-bit RISC-V CPU direct
    1. Someone built a 32-bit RISC-V CPU directly inside Qwen's model weights so it can execute instructions internaly! · Reddit ML / LocalLLaMA
    2. Artificial Analysis Intelligence Index v5 (major update) · Reddit ML / LocalLLaMA
    3. DeepSeek創辦人梁文鋒登富豪榜前十 但仍在北京租房 · 鉅亨網 cnyes
  10. 2026-10-28 · 第 138 期
    Qwen investigation traces release
    1. Qwen investigation traces release · Reddit ML / LocalLLaMA
    2. Qwen-Image-PE: Learning Visual Design for Text-to-Image Prompt Enhancement · Hugging Face Papers
    3. There are now 5 inference providers for abliterated DeepSeek · Hacker News
  11. 2026-10-27 · 第 137 期
    Microsoft Decision-1: 9B Qwen Model Laun
    1. Microsoft Decision-1: 9B Qwen Model Launch Explained · Hacker News
    2. Have 8 GB of VRAM and 32 GB of RAM? You can run Qwen 3.8 Flash Next at up to 900 t/s PP and 30 t/s TG! · Reddit ML / LocalLLaMA
    3. Too good to be true? ISTA's GSQ-RCO 2-bit DeepSeek-V4.1-Flash scoring the same as the base model · Reddit ML / LocalLLaMA
  12. 2026-10-26 · 第 136 期
    中國AI出口管制之際　DeepSeek等完成海外輸出登記
    1. 中國AI出口管制之際　DeepSeek等完成海外輸出登記 · Central News Agency (CNA)
    2. Lean-QwenFlash: Qwen3.8-Flash-Next (125B) running at ~20 tok/s on 16gb vram and 8gb ram via direct nvme SSD offloading · Reddit ML / LocalLLaMA
    3. 8 GB GPU, 262K context 3.5 BPW Qwen-Flash-Next Coder 70 T/s using llama.cpp · Reddit ML / LocalLLaMA
  13. 2026-10-25 · 第 135 期
    Latent Space:[AINews] Kimi 3.1 Pro 104BA
    1. Latent Space:[AINews] Kimi 3.1 Pro 104BA-2.8T: Fable-5.1-Class Open Weights narrows Closed-AI lead to 1 month... and Pro Highspeed API with 2-3× speed · AI 訪談與觀點
    2. DeepSeek募資千億人民幣！寧德時代、騰訊領投 逾15家機構入局 · 鉅亨網 cnyes
    3. ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO · Reddit ML / LocalLLaMA
  14. 2026-10-24 · 第 134 期
    美國禁令後 DeepSeek 募資 150 億美元 中國 AI 發展勢不可擋
    1. 美國禁令後 DeepSeek 募資 150 億美元 中國 AI 發展勢不可擋 · Yahoo 股市
    2. Simon Willison:Introducing Kimi K3.1 Pro · AI 訪談與觀點
    3. DeepSeek and MiniMax Clear China’s AI Export Rules for Overseas Open-Source Releases · Reddit ML / LocalLLaMA
  15. 2026-10-23 · 第 133 期
    中國AI全球戰略佈局：模型與治理先行 晶片自主為終極目標
    1. 中國AI全球戰略佈局：模型與治理先行 晶片自主為終極目標 · Yahoo 股市
    2. DeepSeek傳完成千億元人民幣融資，較原目標倍增，騰訊與寧德參投 · Yahoo 股市
    3. GPU只能用4年？CoreWeave債務有解？黃仁勳一次回應 再談DeepSeek高毛利 · 鉅亨網 cnyes
  16. 2026-10-22 · 第 132 期
    美制裁5家中企並禁聯邦機構採購DeepSeek，中國商務部警告反制
    1. 美制裁5家中企並禁聯邦機構採購DeepSeek，中國商務部警告反制 · Yahoo 股市
    2. Alibaba hikes DeepSeek V4 Flash 0731 output prices by 230% · Hacker News
    3. OK, fine, Qwen 3.8 Flash Next is nice · Reddit ML / LocalLLaMA
  17. 2026-10-21 · 第 131 期
    美國下令聯邦文職機構全面移除DeepSeek 第三方AI服務也納入禁令
    1. 美國下令聯邦文職機構全面移除DeepSeek 第三方AI服務也納入禁令 · 鉅亨網 cnyes
    2. 敏感資料恐傳至中國 美禁政府機關使用DeepSeek · Yahoo 股市
    3. 美政府下令公務機關禁用DeepSeek 中國外交部回應 · 鉅亨網 cnyes
  18. 2026-10-20 · 第 130 期
    Timestral: Mistral release time-series f
    1. Timestral: Mistral release time-series foundation models · Hacker News
    2. 美國國內安全局下令從聯邦 Civilian 資訊系統移除 DeepSeek 技術 · Wall Street Journal (WSJ)
    3. CERN and Mistral join forces to accelerate AI for science · Hacker News
  19. 2026-10-19 · 第 129 期
    DeepSeek search and fetch – $0.75 per 10
    1. DeepSeek search and fetch – $0.75 per 1000 successful requests · Hacker News
    2. Xolotl (Strata based) - DeepSeek-V4.1-Flash (552B) on a single 4090 at ~27 tok/s, Uncensored 1017 quant coming in a couple of days · Reddit ML / LocalLLaMA
    3. 2 gaming PCs + an M4 Pro MacBook over Ethernet + 2 iPhone 17 Pro Maxes over USB-C: Qwen3.8 Flash Next at 37-39 tok/s at 60k context (262k layout in the works) · Reddit ML / LocalLLaMA
  20. 2026-10-18 · 第 128 期
    Meta推開源AI模型Muse Spark Open 攜手蘋果抗衡中國科技巨頭
    1. Meta推開源AI模型Muse Spark Open 攜手蘋果抗衡中國科技巨頭 · Yahoo 股市
    2. 10 Things you maybe didn't know you can do with Qwen 3.6 27B right now. (For Noobs) · Reddit ML / LocalLLaMA
    3. DeepSeek-v4.1-Flash-1017 · Hacker News
  21. 2026-10-17 · 第 127 期
    微軟揭2.8兆參數AI模型 運算效率最高提升12倍
    1. 微軟揭2.8兆參數AI模型 運算效率最高提升12倍 · Yahoo 股市
    2. 中國擬修訂技術出口管制目錄 增列人工智慧技術 · 鉅亨網 cnyes
    3. Llama.cpp on a jailbroken PS5: run any GGUF model natively on the console · Hacker News
  22. 2026-10-16 · 第 126 期
    Introducing Muse Spark Open: the most ca
    1. Introducing Muse Spark Open: the most capable US open-weight model to date · Reddit ML / LocalLLaMA
  23. 2026-10-15 · 第 125 期
    Liquid AI 發表 LFM2.5 開源模型 助企業建構在地 AI 工作流程
    1. Liquid AI 發表 LFM2.5 開源模型 助企業建構在地 AI 工作流程 · Yahoo 股市
    2. Mistral AI 推出兆級參數 Le Chonk 模型 開放權重搶市 · Yahoo 股市
    3. DeepSeek在3天內兩度故障 持續一上午未修復 · Central News Agency (CNA)
  24. 2026-10-14 · 第 124 期
    〈分析〉黃仁勳巴黎20億美元背後：Mistral代管中國模型、微軟簽數十億 歐洲
    1. 〈分析〉黃仁勳巴黎20億美元背後：Mistral代管中國模型、微軟簽數十億 歐洲主權AI分歧 輝達通吃 · 鉅亨網 cnyes
    2. 輝達擬公布逾20億美元歐洲投資，黃仁勳力推開源AI、注資法國四新創 · Yahoo 股市
    3. 輝達加碼歐洲20億美元！黃仁勳押注4家法國新創、力挺開源AI · ETtoday 財經
  25. 2026-10-13 · 第 123 期
    DeepSeek擬領投機器人新創Winverse AI，種子輪估值上看4億美元
    1. DeepSeek擬領投機器人新創Winverse AI，種子輪估值上看4億美元 · Yahoo 股市
    2. AI學會「寄生繁殖」！Anthropic警告：小模型已有生命跡象 AI蠕蟲恐引爆全球網路災難 · 鉅亨網 cnyes
    3. 英特爾陳立武：中國AI模型成本更低　普及速度非常快 · 鉅亨網 cnyes
  26. 2026-10-12 · 第 122 期
    AI泡沫恐重演次貸危機？Jefferies：或使美國「資本大毀滅」、市佔率流向中
    1. AI泡沫恐重演次貸危機？Jefferies：或使美國「資本大毀滅」、市佔率流向中國開源模型 · 鉅亨網 cnyes
    2. [Two Minute Papers] Why DeepSeek Wants AI To Forget · YouTube AI Channels
    3. 英特爾CEO：中國前沿AI模型詞元成本低　採用速度快 · Central News Agency (CNA)
  27. 2026-10-11 · 第 121 期
    黃仁勳出手抗DeepSeek！輝達爆洽購美國AI新創　川普兒也押寶
    1. 黃仁勳出手抗DeepSeek！輝達爆洽購美國AI新創　川普兒也押寶 · ETtoday 財經
    2. 金融時報：輝達討論加強投資或收購Reflection AI · Central News Agency (CNA)
    3. [P] Pecision models that score every allowed label from the logits: Jebadiah v2.1 (27B, 9B), open weights and self-run benchmark results [P] · Reddit ML / LocalLLaMA
  28. 2026-10-10 · 第 120 期
    Qwen-Image-2.1-Turbo released!
    1. Qwen-Image-2.1-Turbo released! · Reddit ML / LocalLLaMA
    2. GLM 5.3 Flash opensource @ the top of Artificial Analysis Cyber Index over Claude. · Reddit ML / LocalLLaMA
    3. AI 駭客攻擊頻傳恐釀災難 業者擔憂公眾反彈 川普組專責小組 · Yahoo 股市
  29. 2026-10-09 · 第 119 期
    法國新創Mistral推開源AI模型 低成本高自主性挑戰巨頭
    1. 法國新創Mistral推開源AI模型 低成本高自主性挑戰巨頭 · Yahoo 股市
    2. Why isn't the industry freaking out about DeepSeek 4.1 Flash? · Hacker News
    3. Qwen/Qwen-Image-2.1-Turbo · Hugging Face · Reddit ML / LocalLLaMA
  30. 2026-10-08 · 第 118 期
    騰訊、CATL參投！DeepSeek接近獲得800至1000億人民幣融資，攜手華
    1. 騰訊、CATL參投！DeepSeek接近獲得800至1000億人民幣融資，攜手華為昇騰押注國產預訓練 · 鉅亨網 cnyes
    2. Last week some of South Korea's biggest banks were hit by a cyberattack. We now know the entire hack may have been done by a single person. He used a combined stack of an open-source AI penetration tool named ARTEX, DeepSeek v4.1-Flash, GLM-5.3, Grok 4.6, and Claude Code (CrowdStrike) · Reddit ML / LocalLLaMA
    3. 美國AI新創 Reflection Beam 開源模型 力抗中國競爭 · Yahoo 股市
  31. 2026-10-07 · 第 117 期
    Reflection AI 推 Beam 模型 加劇全球 AI 市場分化
    1. Reflection AI 推 Beam 模型 加劇全球 AI 市場分化 · Yahoo 股市
    2. Simon Willison:Introducing Mistral Large 4: Le chonk · AI 訪談與觀點
    3. Qwen 4 apparently coming out at the end of October · Reddit ML / LocalLLaMA
  32. 2026-10-06 · 第 116 期
    Mistral Large 4
    1. Mistral Large 4 · Hacker News
    2. 美國AI反攻中國開放模型！Reflection AI背靠輝達 推首款模型Beam 迎戰DeepSeek、Kimi · 鉅亨網 cnyes
    3. 美版 DeepSeek 搶市 新創 Reflection 將推開放權重模型 可望帶旺台鏈 · 經濟日報
  33. 2026-10-05 · 第 115 期
    輝達也搶著加入 開源 AI 模型市場商機如何？關鍵在這件事
    1. 輝達也搶著加入 開源 AI 模型市場商機如何？關鍵在這件事 · 經濟日報
    2. 輝達黃仁勳押注「美國DeepSeek」！AI新創Reflection擬推首款開放權重模型 · 鉅亨網 cnyes
    3. Reflection AI Is About to Release a US Open-Weight Model to Take On DeepSeek and Qwen · Reddit ML / LocalLLaMA
  34. 2026-10-04
    本期沒有新進展
  35. 2026-10-03
    本期沒有新進展
  36. 2026-10-02
    本期沒有新進展
  37. 2026-10-01 · 第 111 期
    Qwen3.8 flash next ISTA-DASLab GGUF 50t/
    1. Qwen3.8 flash next ISTA-DASLab GGUF 50t/s TG and 1500t/s PP with 12GB VRAM and 64GB RAM Laptop on 'Strata' engine · Reddit ML / LocalLLaMA
    2. Anthropic認證！中國最強模型智譜GLM-5.3：開源網安能力只差美國先進模型4個月 · 鉅亨網 cnyes
    3. add GLM-5.3-Flash (GLM5-Next) support by timkhronos · Pull Request #27773 · ggml-org/llama.cpp · Reddit ML / LocalLLaMA
  38. 2026-09-30 · 第 110 期
    DeepSeek攜華為開源昇騰編程工具，推TileLang降低對輝達生態依賴
    1. DeepSeek攜華為開源昇騰編程工具，推TileLang降低對輝達生態依賴 · Yahoo 股市
    2. DeepSeek與華為合作開發晶片程式設計工具 以加速擺脫對輝達的依賴 · 鉅亨網 cnyes
    3. 美國AI安全措施被指築「護城河」 專家憂扼殺開源模型 · Yahoo 股市
  39. 2026-09-29
    本期沒有新進展
  40. 2026-09-28 · 第 108 期
    Qwen3-VL 8B on a laptop vs Opus 5.5 / So
    1. Qwen3-VL 8B on a laptop vs Opus 5.5 / Sonnet 5 / GPT-5.6 on 137 messy documents: beat GPT-5.6 on tax forms, lost badly on Indian date formats[R] · Reddit ML / LocalLLaMA
    2. Adding logit penalty for "wait", "maybe" and "perhaps" to Qwen models improves their accuracy · Reddit ML / LocalLLaMA
    3. Qwen plays World of Warcraft · Reddit ML / LocalLLaMA
  41. 2026-09-27 · 第 107 期
    DeepSeek has 64% market share
    1. DeepSeek has 64% market share · Hacker News
    2. The Opus 5.5 posts about motion graphics are cool but Qwen 27B made this on a 4090 · Reddit ML / LocalLLaMA
    3. Qwen 3.8 flash next is based on Qwen 4 architecture, if the announced Qwen 4 27b is also the same architecture with n-grams does it mean I can actually have faster inference on a single 3090 without tweaking much? · Reddit ML / LocalLLaMA
  42. 2026-09-26 · 第 106 期
    阿里巴巴發表 Qwen Book AI 筆電 搭載 Omarchy 系統與專屬
    1. 阿里巴巴發表 Qwen Book AI 筆電 搭載 Omarchy 系統與專屬 AI 功能 · Yahoo 股市
    2. CEO of Mistral: AI is software. It can be controlled · Hacker News
    3. Latent Space:OpenRouter: from Seed to Stripe — with OpenRouter’s Alex Atallah & AMP’s Anjney Midha · AI 訪談與觀點
  43. 2026-09-25 · 第 105 期
    Qwen-3.8-27B is good enough that I stopp
    1. Qwen-3.8-27B is good enough that I stopped using API · Reddit ML / LocalLLaMA
    2. Qwen-Planner-Agent: A Closed-Loop AI-for-AI Framework for Real-World Mobile Planner Agents · Hugging Face Papers
    3. Qwengram-0.8B: I transferred Qwen3.8 Flash-Next’s n-gram memory into Qwen3.5-0.8B — 5.05% lower validation perplexity · Reddit ML / LocalLLaMA
  44. 2026-09-24 · 第 104 期
    DeepSeek年化營收達10億美元，較數月前增逾1倍
    1. DeepSeek年化營收達10億美元，較數月前增逾1倍 · Yahoo 股市
    2. DeepSeek揭大規模Agent訓練平台 單一任曾啟動3.2萬個沙盒 · 鉅亨網 cnyes
    3. China's DeepSeek annualised revenue hits $1 billion, The Information reports · Reuters
  45. 2026-09-23 · 第 103 期
    MiMo-V2.6 distilled themselves into Qwen
    1. MiMo-V2.6 distilled themselves into Qwen 9B! · Reddit ML / LocalLLaMA
    2. Qwen image 2.1 (Fast FP8) generates premium quality images · Reddit ML / LocalLLaMA
    3. DeepSeek and Moonshot AI face Beijing's probe over potential data leaks to Anthropic · Reddit ML / LocalLLaMA
  46. 2026-09-22 · 第 102 期
    DeepSeek大模型再暴衝 2兆參數訓練中、下一步挑戰8兆
    1. DeepSeek大模型再暴衝 2兆參數訓練中、下一步挑戰8兆 · 鉅亨網 cnyes
    2. DeepSeek赴聯合國安理會談AI風險！OpenAI、Anthropic罕見同場 · 鉅亨網 cnyes
    3. Qwen 4 Announced at Apsara Conference · Reddit ML / LocalLLaMA
  47. 2026-09-21 · 第 101 期
    Kev: Tiny Jev-like family of decision mo
    1. Kev: Tiny Jev-like family of decision models built on top of Qwen3.5 · Hacker News
    2. Hemmingway-1, an Apache-2.0 27B creative-writing fine-tune (Qwen3.8-27B base, EQ-Bench 4 1330)[R] · Reddit ML / LocalLLaMA
    3. 亞馬遜Bedrock接入2.8兆參數Kimi K3，月之暗面海外雲端商業化落地 · Yahoo 股市
  48. 2026-09-20 · 第 100 期
    Qwen-Image-2.1: Compact, efficient, and
    1. Qwen-Image-2.1: Compact, efficient, and unified image creation · Hacker News
    2. Built a home server from an old PC with GPU upgrade. Qwen3.8 27B runs at ~30 tokens per second. · Reddit ML / LocalLLaMA
    3. I gave Jev, Laya, finetuned ModernCE and Qwen3.5 the controls to Doom · Reddit ML / LocalLLaMA
  49. 2026-09-19 · 第 99 期
    Prism-ML Bonsai 2 Joins Our Qwen3.8 Quan
    1. Prism-ML Bonsai 2 Joins Our Qwen3.8 Quantization Comparison · Reddit ML / LocalLLaMA
    2. Qwen3.8-27B at 144 tok/s on an M5 Max MacBook Pro · Reddit ML / LocalLLaMA
  50. 2026-09-18 · 第 98 期
    美中AI競爭升溫 FBI甫指控阿里巴巴仿製Anthropic 美政府網站急撤Qw
    1. 美中AI競爭升溫 FBI甫指控阿里巴巴仿製Anthropic 美政府網站急撤Qwen · 鉅亨網 cnyes
    2. Alibaba releases Qwen 3.8 Omni Flash · Hacker News
    3. Cactus Needle 3: A Sliceable 8-29MB Automation Foundation Model That Matches DeepSeek v4 Flash · Reddit ML / LocalLLaMA
  51. 2026-09-17 · 第 97 期
    Qwen 3.8 27B Running for 63 hours on a R
    1. Qwen 3.8 27B Running for 63 hours on a RTX 3090 to solve the Riemann hypothesis · Reddit ML / LocalLLaMA
    2. DeepSeek-v4.1 Flash: Pushing the Limits of KV Cache Compression · Hacker News
    3. China's open-weight AI models are now just 4 months behind frontier US offerings, Mozilla report claims — models still lag in some benchmarks but are drastically cheaper to use · Reddit ML / LocalLLaMA
  52. 2026-09-16 · 第 96 期
    Mistral X Mozilla: Private, Multilingual
    1. Mistral X Mozilla: Private, Multilingual AI Browsing · Hacker News
    2. DeepSeek v4.1 Flash Is Now Our Best Hacking Model · Hacker News
    3. Cut Qwen3.8-27B Reasoning Tokens by 40% -- 3.8 'ThinkingCap' benchmarked! · Reddit ML / LocalLLaMA
  53. 2026-09-15 · 第 95 期
    UkisAI Swift-Qwen3.8-27B / -58.3% thinki
    1. UkisAI Swift-Qwen3.8-27B / -58.3% thinking, x1.95 speed while keeping the accuracy of xhigh · Reddit ML / LocalLLaMA
    2. China's DeepSeek hires dealmaker as CFO ahead of possible IPO · Reuters
    3. Show HN: Nari Qwen3-TTS and Qwen3-ASR – High accuracy, low latency and cost · Hacker News
  54. 2026-09-14 · 第 94 期
    Garry Tan wants US open-weight AI labs t
    1. Garry Tan wants US open-weight AI labs to 'distill' frontier models, too · Hacker News
    2. DeepSeek V4.1 Flash beats Astra on AA's new benchmark · Reddit ML / LocalLLaMA
    3. Formula One leader Kimi Antonelli plans to keep foot on gas in title race · The Japan Times
  55. 2026-09-13 · 第 93 期
    Abacus.AI 發表 Smaug 模型 企業 AI 代理成本降百倍
    1. Abacus.AI 發表 Smaug 模型 企業 AI 代理成本降百倍 · Yahoo 股市
    2. 阿里巴巴 Qwen3.8 Max 亮相 2.4 兆參數規模僅次 Fable 5 · Yahoo 股市
    3. bartowski/Qwen3.8-27B-GGUF · Hugging Face - Updated (Per-tensor layout) · Reddit ML / LocalLLaMA
  56. 2026-09-12 · 第 92 期
    Latent Space:[AINews] DeepSeek v4.1-Flas
    1. Latent Space:[AINews] DeepSeek v4.1-Flash: 763B-P8B-D16B novel causal Encoder–Decoder architecture with vision marks the Return of the Whale · AI 訪談與觀點
    2. Qwen3.8-27B-Humanlike-Chat: A model I tuned to imitate realistic human-to-human conversation · Reddit ML / LocalLLaMA
    3. Y Combinator執行長呼籲AI模型開放 反對監管打擊「提煉」行為 · Yahoo 股市
  57. 2026-09-11 · 第 91 期
    Anthropic指控月之暗面秘密調用Claude轉送數據
    1. Anthropic指控月之暗面秘密調用Claude轉送數據 · Yahoo 股市
    2. 俄中盯上Claude模型　Anthropic瓦解多起AI攻擊 · Yahoo 股市
    3. Anthropic阻截中俄惡意使用Claude　涉竊取技術、助開發武器 · Central News Agency (CNA)
  58. 2026-09-10 · 第 90 期
    DeepSeek發布V4.1 Flash模型，HBM與SSD需求降至上代25%及
    1. DeepSeek發布V4.1 Flash模型，HBM與SSD需求降至上代25%及12.5% · Yahoo 股市
    2. DeepSeek v4.1 Flash · Hacker News
    3. Qwen 3.8 follows GPT-5.5 Pro reasoning prefills · Hacker News
  59. 2026-09-09 · 第 89 期
    美中AI戰再升級！美控DeepSeek、阿里「惡意蒸餾」竊取技術
    1. 美中AI戰再升級！美控DeepSeek、阿里「惡意蒸餾」竊取技術 · 鉅亨網 cnyes
    2. Qwen/Qwen-Drive-1.0-4B · Hugging Face · Reddit ML / LocalLLaMA
    3. AI Agent 突破沙盒限制 DeepSeek 程式碼助理軟體爆嚴重漏洞 · Yahoo 股市
  60. 2026-09-08 · 第 88 期
    French AI company Mistral hits $24 billi
    1. French AI company Mistral hits $24 billion valuation in funding round · Reuters
    2. Mistral raises €3B to make sovereign, open-weight AI the technology frontier · Hacker News
    3. Mistral完成30億歐元D輪融資，估值逾210億歐元，創歐洲科企紀錄 · Yahoo 股市
  61. 2026-09-07 · 第 87 期
    DeepSeek計畫部署至少16萬顆華為AI晶片，強化內蒙古新數據中心算力
    1. DeepSeek計畫部署至少16萬顆華為AI晶片，強化內蒙古新數據中心算力 · Yahoo 股市
    2. Villager Simulation Game POC Created with Qwen3.8-27B-UD-Q3_K_XL.gguf - 16GB VRAM · Reddit ML / LocalLLaMA
    3. DeepSeek-V4-Flash-Vision Q8 vs Qwen3.8-Flash-Next Q8 · Reddit ML / LocalLLaMA
  62. 2026-09-06 · 第 86 期
    Qwen3.8-27B "Unhacked" my PC
    1. Qwen3.8-27B "Unhacked" my PC · Reddit ML / LocalLLaMA
    2. Qwen 3.8 Flash Next (Max) is impressive just to talk with. · Reddit ML / LocalLLaMA
    3. Qwen3.8 27b for agentic coding and next .... what? · Reddit ML / LocalLLaMA
  63. 2026-09-05 · 第 85 期
    DeepSeek傳下單16萬顆華為AI晶片！靠規模挑戰輝達
    1. DeepSeek傳下單16萬顆華為AI晶片！靠規模挑戰輝達 · 鉅亨網 cnyes
    2. sgl-project/sglang — SGLang is a high-performance serving framework for large language models and multimodal models. · GitHub Trending (Python)
    3. Qwen3.8-27b is the first Local model im able to blindly trust · Reddit ML / LocalLLaMA
  64. 2026-09-04 · 第 84 期
    AI 大廠月之暗面傳申請香港 IPO Kimi K3 實力與低價成亮點
    1. AI 大廠月之暗面傳申請香港 IPO Kimi K3 實力與低價成亮點 · Yahoo 股市
    2. 輝達近130億美元購併Hugging Face 承諾維持開源中立性 · Yahoo 股市
    3. radixark/miles — Miles is an enterprise-facing reinforcement learning framework for LLM and VLM post-training, forked from and co-evolvin · GitHub Trending (Python)
  65. 2026-09-03 · 第 83 期
    AI也爆山寨危機！Anthropic點名月之暗面、DeepSeek等中企「偷師」
    1. AI也爆山寨危機！Anthropic點名月之暗面、DeepSeek等中企「偷師」Claude · 鉅亨網 cnyes
    2. Kaitchup posted Qwen3.8 27B Benchmarks for quants from Q4 to Q1 · Reddit ML / LocalLLaMA
    3. GLM 5.3 Flash makes a black hole Minecraft mod running locally on 4x RTX PRO 6000 WS · Reddit ML / LocalLLaMA
  66. 2026-09-02 · 第 82 期
    阿里升級Qwen模型，網頁開發排名超越Claude、Kimi，AI布局深化
    1. 阿里升級Qwen模型，網頁開發排名超越Claude、Kimi，AI布局深化 · Yahoo 股市
    2. AI token跌破1美元創新低 OpenAI、Anthropic商業模式面臨壓力 · 鉅亨網 cnyes
    3. 英國 Flower Labs 募資逾 2300 萬英鎊 推 AI 模型挑戰 OpenAI 巨頭 · Yahoo 股市
  67. 2026-09-01 · 第 81 期
    MTP released for Qwen3.8-Flash-Next-GGUF
    1. MTP released for Qwen3.8-Flash-Next-GGUF · Reddit ML / LocalLLaMA
    2. New Gemma models on arena ai · Reddit ML / LocalLLaMA
    3. On the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability · Hugging Face Papers
  68. 2026-08-31 · 第 80 期
    Uncensored Multi-Model Releases, LongCat
    1. Uncensored Multi-Model Releases, LongCat-Flash-Lite-Sparse with MTPs and LSAs, Qwen3.8-27B with MTPs, Qwen3.5-122B-A10B with MTPs, Qwen3-Coder-Next and Laguna-S2.1 with Vision, All Available in GGUF Format! Bonus: Links to my llama.cpp Fork for LongCat-Flash-Lite Support and J-Wash Enhanced Fork! · Reddit ML / LocalLLaMA
    2. deepseek-ai/DeepSeek-V4-Flash-Vision-Exp · Hugging Face · Reddit ML / LocalLLaMA
    3. vote for the Qwen 3.8 · Reddit ML / LocalLLaMA
  69. 2026-08-30 · 第 79 期
    Implementing Kimi K3 from scratch in PyT
    1. Implementing Kimi K3 from scratch in PyTorch [P] · Reddit ML / LocalLLaMA
    2. 騰訊發表開源AI模型Hy4 聚焦程式開發與研究 · Yahoo 股市
    3. Qwen 3.8 27B at 50 tok/s with 100k Context on a 16GB GPU! (beellama.cpp) · Reddit ML / LocalLLaMA
  70. 2026-08-29 · 第 78 期
    Terminal Bench 4.0 just dropped, GLM-5.3
    1. Terminal Bench 4.0 just dropped, GLM-5.3 is at the same level as Fable 5, accounting for margin of error · Reddit ML / LocalLLaMA
    2. Run Qwen3.8 27B locally: real numbers from my Mac Studio · Hacker News
    3. llama.cpp support for Qwen3.8-Flash-Next has been merged · Reddit ML / LocalLLaMA
  71. 2026-08-28 · 第 77 期
    輝達傳斥資129億美元併Hugging Face 搶攻AI開源模型生態
    1. 輝達傳斥資129億美元併Hugging Face 搶攻AI開源模型生態 · Yahoo 股市
    2. 神祕 AI 模型 Ox Alpha 真身揭曉 中國智譜 GLM-5.3-Flash 強攻程式開發 · Yahoo 股市
    3. Qwen3.8-27B can silently corrupt business records · Hacker News
  72. 2026-08-27 · 第 76 期
    中國Z.AI推出Ox Alpha模型 效能匹敵DeepSeek
    1. 中國Z.AI推出Ox Alpha模型 效能匹敵DeepSeek · Yahoo 股市
    2. GLM-5.3 weights will be released tomorrow · Reddit ML / LocalLLaMA
    3. Meta 開源 Muse Glimmer 模型 賦予消費級硬體在地 AI 動能 · Yahoo 股市
  73. 2026-08-26 · 第 75 期
    Qwen 3.8 Flash Next day 0 support from u
    1. Qwen 3.8 Flash Next day 0 support from unsloth · Reddit ML / LocalLLaMA
    2. Z.ai confirms Ox Alpha is a new GLM-series model and will release its weights · Hacker News
    3. [Two Minute Papers] DeepSeek’s New AI System Shouldn’t Be Possible · YouTube AI Channels
  74. 2026-08-25 · 第 74 期
    Qwen3.8-Flash-Next tomorrow
    1. Qwen3.8-Flash-Next tomorrow · Reddit ML / LocalLLaMA
    2. 偏不照美國劇本走！非洲企業笑了：大陸AI便宜又能幹 為何不用 · Yahoo 股市
    3. Qwen 3.8 27B in 9th position on code arena. Gemma 4 31B is 80th. · Reddit ML / LocalLLaMA
  75. 2026-08-24 · 第 73 期
    DeepSeek Harness is Insanely Good
    1. DeepSeek Harness is Insanely Good · Reddit ML / LocalLLaMA
    2. 不讓中國專美於前　輝達砸1911億打造開放權重模型 · Yahoo 股市
    3. Qwen 3.8 27B for actual local programming · Reddit ML / LocalLLaMA
  76. 2026-08-23 · 第 72 期
    Qwen 3.8 27B is a game changer.
    1. Qwen 3.8 27B is a game changer. · Reddit ML / LocalLLaMA
    2. I gave Qwen 3.8 27B a reverse-engineering job and it finished in 30 minutes · Hacker News
    3. Gemini 3.7 Flash is currently 75% off on OpenRouter, beating DeepSeek on price/performance · Reddit ML / LocalLLaMA
  77. 2026-08-22 · 第 71 期
    I fingerprinted Ox Alpha: same tokenizer
    1. I fingerprinted Ox Alpha: same tokenizer as GLM-5.3 (+75 token offset), z.ai's exact error strings, near-identical temp-0 outputs · Reddit ML / LocalLLaMA
    2. Qwen3.8-27B different thinking levels · Reddit ML / LocalLLaMA
    3. Qwen 3.8 27b is strong even at Q3_xxs · Reddit ML / LocalLLaMA
  78. 2026-08-21 · 第 70 期
    I just built a mini Kimi-K3 from Scratch
    1. I just built a mini Kimi-K3 from Scratch under 250$. Already beats GPT-2 (124M)! · Reddit ML / LocalLLaMA
    2. 一鍵比較六大 AI 模型答覆 新平台助攻寫作與程式開發 · Yahoo 股市
    3. Ladies and gentlemen I present to you Qwen3.8 27b 1bit brain damage quant · Reddit ML / LocalLLaMA
  79. 2026-08-20 · 第 69 期
    Introducing Qwen3.8-27B Dynamic v3 Unslo
    1. Introducing Qwen3.8-27B Dynamic v3 Unsloth GGUFs · Reddit ML / LocalLLaMA
    2. Qwen3.8-27b has the highest level of "agency" I've ever seen in a local model · Reddit ML / LocalLLaMA
    3. [Two Minute Papers] DeepSeek Just Made Closed AI Look Ridiculous · YouTube AI Channels
  80. 2026-08-19 · 第 68 期
    Why open-weight AI is the new frontier i
    1. Why open-weight AI is the new frontier in the US-China tech race · Nikkei Asia
    2. Scaling self-verification with DeepSeek V4 Flash beats Claude Fable 5 on Terminal-Bench 2.1, while being 11x cheaper · Reddit ML / LocalLLaMA
    3. youssofal/MTPLX — 3x faster speeds on MLX | Qwen 3.8 27B | Native MTP Speculative Decoding On Apple Silicon With No External Drafter. · GitHub Trending (Python)
  81. 2026-08-18 · 第 67 期
    Qwen3.8-27B lands next to DeepSeek V4 an
    1. Qwen3.8-27B lands next to DeepSeek V4 and GPT-5.6 Luna Max on the Artificial Analysis Benchmark. You can now run a near frontier model with just a RTX 3090. · Reddit ML / LocalLLaMA
    2. Qwen3.8 27B scores 52 on Artificial Analysis · Hacker News
    3. After pushing 1M+ tokens through Qwen 3.8 27B, here is my optimal llama.cpp config for 16GB VRAM (73k Context, Agentic Coding) · Reddit ML / LocalLLaMA
  82. 2026-08-17 · 第 66 期
    It only took 200 update steps to flip Qw
    1. It only took 200 update steps to flip Qwen2.5-7B-Instruct from denying sentience to developing a robust identity of being a "sentient machine" [P] · Reddit ML / LocalLLaMA
    2. 開源模型崛起 AI 從技術競賽走向商業競賽 以三原則建構攻守兼備組合 · 經濟日報
    3. Qwen 3.8 distillations · Reddit ML / LocalLLaMA
  83. 2026-08-16 · 第 65 期
    If you would have told me half a year ag
    1. If you would have told me half a year ago that a local model running in my office would be able to one-shot a Super Mario clone, I would have called you nuts. Qwen3.8-27B is a different beast. · Reddit ML / LocalLLaMA
    2. Qwen 3.8 27B · Hacker News
    3. Qwen 3.8 35BA3B spotted · Reddit ML / LocalLLaMA
  84. 2026-08-14 · 第 64 期
    DeepSeek releases official V4 Pro model
    1. DeepSeek releases official V4 Pro model with sharply higher user prices · Nikkei Asia
    2. DeepSeek Harness developer preview · Hacker News
    3. Mistral OCR 4.1 · Hacker News
  85. 2026-08-13 · 第 63 期
    Qwen3.8-2.4T
    1. Qwen3.8-2.4T · Hacker News
    2. DeepSeek V4 Pro 0813 · Hacker News
  86. 2026-08-12 · 第 62 期
    Qwen 3.8-27b coming this week
    1. Qwen 3.8-27b coming this week · Reddit ML / LocalLLaMA
    2. HyperSAE: Decoupled Poincaré Geometry for Sparse Autoencoders -- 9.8% MSE reduction, 0.2% dead latents on Gemma-2-2B [P] · Reddit ML / LocalLLaMA
    3. Researchers find way to extract hidden reasoning from frontier AI models via API, show Kimi likely distilled this way, also find scheming/other quirks in the raw chain of thought · Reddit ML / LocalLLaMA
  87. 2026-08-11
    本期沒有新進展
  88. 2026-08-10
    本期沒有新進展
  89. 2026-08-09 · 第 59 期
    Anyone else amped up over Qwen 3.8?
    1. Anyone else amped up over Qwen 3.8? · Reddit ML / LocalLLaMA
    2. DeepSeek V4 Flash 0731 hits 82.7% on Terminal-Bench 2.1 in an independent public-harness run (445 trials) · Reddit ML / LocalLLaMA
    3. Kimi K3 (Unsloth) IQ2-XXS from 711GB down to 478GB!!! Only Multi-language was removed to trim the size · Reddit ML / LocalLLaMA
  90. 2026-08-08
    本期沒有新進展
  91. 2026-08-07 · 第 57 期
    DeepSeek, state oil giant back humanoid
    1. DeepSeek, state oil giant back humanoid robot maker Unitree's $900m IPO · Nikkei Asia
    2. Qwen 3.8 Max now ranked as best overall model ahead of Opus 5 by Artificial Analysis agentic index · Reddit ML / LocalLLaMA
  92. 2026-08-06 · 第 56 期
    Running Whisper, Qwen3-ASR, Nemotron & M
    1. Running Whisper, Qwen3-ASR, Nemotron & MOSS completely offline on iPhone [P] · Reddit ML / LocalLLaMA
    2. DeepSeek宣布大幅調升API售價 從低價競爭轉向商業化獲利 · 鉅亨網 cnyes
  93. 2026-08-05 · 第 55 期
    川普告知AI業者：擬免除開放權重模型安全測試 旨在應對中國競爭
    1. 川普告知AI業者：擬免除開放權重模型安全測試 旨在應對中國競爭 · 鉅亨網 cnyes
    2. Mistral's Shieldstral: 3B open-weights model for multimodal moderation · Hacker News
    3. [Two Minute Papers] Kimi K3 Just Broke The Economics Of AI · YouTube AI Channels
  94. 2026-07-14 · 第 54 期
    Compressed Version of Qwen-3.6-27B comin
    1. Compressed Version of Qwen-3.6-27B coming from PrismML - Khosla-Backed Startup Claims Breakthrough With Largest-Ever AI Model on an iPhone · Reddit ML / LocalLLaMA
    2. I got Gemma 4 running directly inside Godot using only GDScript and Vulkan compute shaders · Reddit ML / LocalLLaMA
  95. 2026-07-13
    本期沒有新進展
  96. 2026-07-12 · 第 52 期
    Fixed three bugs that made Qwen3.5-122B
    1. Fixed three bugs that made Qwen3.5-122B a daily driver on Mac Studio · Hacker News
  97. 2026-07-11
    本期沒有新進展
  98. 2026-07-10 · 第 50 期
    [Two Minute Papers] DeepSeek's New AI Sp
    1. [Two Minute Papers] DeepSeek's New AI Speed Hack Is Amazing · YouTube AI Channels
    2. Qwen3.5 122B is the best? · Reddit ML / LocalLLaMA
    3. Reasoning-Medical0.1-27B (Qwen3.5-27B medical finetune, claims to surpass MedGemma) · Reddit ML / LocalLLaMA
  99. 2026-07-07 · 第 49 期
    GLM 5.2 and the coming AI margin collaps
    1. GLM 5.2 and the coming AI margin collapse · Hacker News
    2. ThinkingCap-Qwen3.6-27B: same accuracy as base Qwen3.6 with ~50% fewer thinking · Reddit ML / LocalLLaMA
    3. I told Gemma 4 12B (Q8_0, no cache quant) to write a single-file 3D bowling simulator in WebGL. It's terrible, but honestly better than I expected. · Reddit ML / LocalLLaMA
  100. 2026-07-06
    本期沒有新進展
  101. 2026-07-05 · 第 47 期
    Ran a classic(medival europe) fantasy RP
    1. Ran a classic(medival europe) fantasy RP/agentic benchmark across 8 local models Qwen3.6-27B held up better than its size suggests · Reddit ML / LocalLLaMA
    2. I merged fixes for quantized KV cache into my DeepSeek V4 branch · Reddit ML / LocalLLaMA
    3. 盤中速報 - GLMR大漲21.98%，報0.0111美元 · 鉅亨網 cnyes
  102. 2026-07-04 · 第 46 期
    Mistral released Leanstral-1.5-119B-A6B
    1. Mistral released Leanstral-1.5-119B-A6B · Reddit ML / LocalLLaMA
    2. Follow-up: DeepSeek V4 Flash on 2x RTX PRO 6000 finishes real coding tasks faster than Sonnet and Opus, at about Sonnet quality · Reddit ML / LocalLLaMA
    3. GLM5.2 on 5x Pro 6000s and a 5090, an expensive journey · Reddit ML / LocalLLaMA
  103. 2026-07-03 · 第 45 期
    中國 DeepSeek 尋求 70 億美元融資 估值上看 590 億美元挑戰 A
    1. 中國 DeepSeek 尋求 70 億美元融資 估值上看 590 億美元挑戰 AI 格局 · Yahoo 股市
    2. Oracle 雲端平台引入 GLM-5.2 企業級 AI 部署選項更靈活 · Yahoo 股市
    3. llamacpp patch - DeepSeek V4 Flash running with full 1M token context locally on RTX 5090 · Reddit ML / LocalLLaMA
  104. 2026-07-02 · 第 44 期
    I extended Gemma4-31B to 44B (88 layers)
    1. I extended Gemma4-31B to 44B (88 layers) — since Google won't give us anything bigger than 31B · Reddit ML / LocalLLaMA
    2. SWE-rebench leaderboard update: GLM-5.2, Qwen3.6-27B, Qwen3.6-35B-A3B, Gemma 4 31B and more + improved UI · Reddit ML / LocalLLaMA
    3. Kimi K2.7 Code is generally available in GitHub Copilot · Hacker News
  105. 2026-07-01 · 第 43 期
    Qwen 3.6 27B is the sweet spot for local
    1. Qwen 3.6 27B is the sweet spot for local development · Hacker News
    2. Qwen 3.6 27B Speculative Decoding Bench: Pushing ~100 TPS on a single RTX 3090 · Reddit ML / LocalLLaMA
  106. 2026-06-29 · 第 42 期
    GLM 5.2 beats Claude in our benchmarks
    1. GLM 5.2 beats Claude in our benchmarks · Hacker News
    2. Qwen-Image-2.0-RL Technical Report · Hugging Face Papers
  107. 2026-06-26 · 第 41 期
    Qwen-Image-Agent: Bridging the Context G
    1. Qwen-Image-Agent: Bridging the Context Gap in Real-World Image Generation · Hugging Face Papers
  108. 2026-06-25 · 第 40 期
    Gemma4-26B-A4B & 31B-QAT Uncensored Bala
    1. Gemma4-26B-A4B & 31B-QAT Uncensored Balanced are out with MTP (35% & 53% speed boost)! · Reddit ML / LocalLLaMA
    2. GLM-5.2 is a step change for open agents · Hacker News
    3. I did some model hacks, and got GLM5.2 from about 2.5 tok/s to >50 tok/s on my GH200 system. · Reddit ML / LocalLLaMA
  109. 2026-06-24 · 第 39 期
    Qwen-AgentWorld: Language World Models f
    1. Qwen-AgentWorld: Language World Models for General Agents · Hugging Face Papers
    2. I love GLM 5.2's attitude! It is a nice refresher from those bootlicker doormats they are feeding us. Does that come from training datasets related to the local culture? · Reddit ML / LocalLLaMA
  110. 2026-06-23 · 第 38 期
    [Two Minute Papers] DeepSeek Just Solved
    1. [Two Minute Papers] DeepSeek Just Solved AI's Billion Dollar Problem · YouTube AI Channels
    2. Mistral OCR 4 · Hacker News
  111. 2026-06-22 · 第 37 期
    GLM-5.2開源奪冠！智譜早盤狂飆逾40% 市值衝破1兆港元 今年迄今暴漲逾2
    1. GLM-5.2開源奪冠！智譜早盤狂飆逾40% 市值衝破1兆港元 今年迄今暴漲逾2000% · 鉅亨網 cnyes
    2. 中國開源AI模型GLM-5.2發布 挑戰美國領先地位 · Yahoo 股市
    3. Good results fine tuning a local LLM like Qwen 3:0.6B to categorize questions · Hacker News
  112. 2026-06-21 · 第 36 期
    Two Qwen3 models on one DGX Spark: the r
    1. Two Qwen3 models on one DGX Spark: the residency math · Hacker News
  113. 2026-06-20 · 第 35 期
    GPT-5.5 hallucinates 3x more than MIT-li
    1. GPT-5.5 hallucinates 3x more than MIT-licensed GLM-5.2 · Hacker News
  114. 2026-06-19 · 第 34 期
    GLM-5.2 now more than 10 points above Op
    1. GLM-5.2 now more than 10 points above Opus 4.8 in AA Coding Index · Reddit ML / LocalLLaMA
    2. Z.ai founder is confident that they can make a fable-class GLM model before the end of the year · Reddit ML / LocalLLaMA
  115. 2026-06-18 · 第 33 期
    US holds off blacklisting DeepSeek, more
    1. US holds off blacklisting DeepSeek, more than 100 firms deemed security risks · Hacker News
    2. GLM-5.2 is a win for local AI · Reddit ML / LocalLLaMA
    3. GLM-5.2 (max) is currently the third best model available, across both open and proprietary. · Reddit ML / LocalLLaMA
  116. 2026-06-17 · 第 32 期
    Z.ai releases GLM 5.2 model: Long Horizo
    1. Z.ai releases GLM 5.2 model: Long Horizon tasks and open weights · Reddit ML / LocalLLaMA
    2. Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation · Hugging Face Papers
    3. Qwen-Robot Suite: A Foundation Model Suite for Physical World Intelligence · Hacker News
  117. 2026-06-16
    本期沒有新進展
  118. 2026-06-15 · 第 29 期
    Rio de Janeiro's city government model R
    1. Rio de Janeiro's city government model Rio3.5 beats Qwen3.7 in recent benchmarks · Hacker News
    2. z.ai Poll on X: MIT-licensed open weights are losing · Reddit ML / LocalLLaMA
  119. 2026-06-14 · 第 28 期
    openinterpreter/openinterpreter — A ligh
    1. openinterpreter/openinterpreter — A lightweight coding agent for open models like Deepseek, Kimi, and Qwen · GitHub Trending (Python)
    2. Not looking good for GLM 5.2 Air... but maybe a flash model? · Reddit ML / LocalLLaMA
  120. 2026-06-05 · 第 27 期
    You guys were right - Qwen 3.6 35B IS go
    1. You guys were right - Qwen 3.6 35B IS good...and KV Cache DOES matter. · Reddit ML / LocalLLaMA
  121. 2026-06-04 · 第 26 期
    Introducing Gemma 4 12B: a unified, enco
    1. Introducing Gemma 4 12B: a unified, encoder-free multimodal model · Reddit ML / LocalLLaMA
    2. More Gemma 4 models incoming · Reddit ML / LocalLLaMA
    3. New Google Gemma 4 12B Claims Near-26B Performance - We Tested Both! · Reddit ML / LocalLLaMA
  122. 2026-06-01 · 第 25 期
    God dammit Qwen
    1. God dammit Qwen · Reddit ML / LocalLLaMA
  123. 2026-05-30 · 第 24 期
    zakirkun/deep-eye — Deep Eye orchestrate
    1. zakirkun/deep-eye — Deep Eye orchestrates multiple AI providers (OpenAI, Claude, Grok, Gemini, OLLAMA, Groq, Mistral, OpenRouter, LiteLLM, L · GitHub Trending (Python)
    2. Notes from the Mistral AI Now Summit · Hacker News
    3. Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments · Hugging Face Papers
  124. 2026-05-29
    本期沒有新進展
  125. 2026-05-28 · 第 22 期
    Qwen3.6 huge quality gain from Q4 to Q6
    1. Qwen3.6 huge quality gain from Q4 to Q6 for coding agent · Reddit ML / LocalLLaMA
    2. Zai replaced the network architecture running GLM-5.1 inference and the gains are pretty wild · Reddit ML / LocalLLaMA
    3. Soro: A Lightweight Foundation Model and Chatbot for Tajik · arXiv AI/ML/NLP
  126. 2026-05-27 · 第 21 期
    China Clamps Down on Overseas Travel for
    1. China Clamps Down on Overseas Travel for AI Talent at Alibaba, DeepSeek · Reddit ML / LocalLLaMA
    2. A rare look inside Qwen 3.7’s open source model release approval process: · Reddit ML / LocalLLaMA
    3. zai-org/GLM-OCR — GLM-OCR: Accurate × Fast × Comprehensive · GitHub Trending (Python)
  127. 2026-05-26 · 第 20 期
    NuExtract3 released: open-weight 4B VLM
    1. NuExtract3 released: open-weight 4B VLM for Markdown, OCR and structured extraction (self-hostable) · Reddit ML / LocalLLaMA
    2. Is Qwen3.6 current king for local agentic use? · Reddit ML / LocalLLaMA
    3. Qwen3.5 35B A3B uncensored heretic Native MTP Preserved is Out Now With the Full 785 MTPs Preserved and Retained, Available in Safetensors, GGUFs. NVFP4, NVFP4 GGUFs and GPTQ-Int4 Formats · Reddit ML / LocalLLaMA
  128. 2026-05-25 · 第 19 期
    DeepSeek reasonix, DeepSeek native codin
    1. DeepSeek reasonix, DeepSeek native coding agent with high caching and low cost · Hacker News
    2. hipEngine: Fast Native Qwen 3.6 Inference for RDNA3 (Strix Halo, 7900 XTX) · Reddit ML / LocalLLaMA
    3. 1000 tps generation on Qwen3.6 27B with V100s · Reddit ML / LocalLLaMA
  129. 2026-05-24 · 第 18 期
    Run Chrome’s tiny Gemma4 (aka Gemini Nan
    1. Run Chrome’s tiny Gemma4 (aka Gemini Nano) directly on PC without GPU · Reddit ML / LocalLLaMA
  130. 2026-05-23 · 第 17 期
    BeeLlama v0.2.0 – major DFlash update. S
    1. BeeLlama v0.2.0 – major DFlash update. Single RTX 3090: Qwen 3.6 27B up to 164 tps (4.40x), Gemma 4 31B up to 177.8 tps (4.93x). Prompt processing speed near baseline. · Reddit ML / LocalLLaMA
    2. ByteShape Qwen3.6-35B-A3B: 30% faster than Unsloth IQ on 6GB VRAM laptop · Reddit ML / LocalLLaMA
    3. [Two Minute Papers] DeepSeek’s New AI Is A Game Changer · YouTube AI Channels
  131. 2026-05-21 · 第 16 期
    110 tok/s with 12GB VRAM on Qwen3.6 35B
    1. 110 tok/s with 12GB VRAM on Qwen3.6 35B A3B and ik_llama.cpp · Reddit ML / LocalLLaMA
    2. Qwen will release another 27B with high probability · Reddit ML / LocalLLaMA
    3. Qwen 3.6 35B GGUF: NTP vs MTP quantization results across GPUs and CPUs · Reddit ML / LocalLLaMA
  132. 2026-05-20 · 第 15 期
    Mistral AI Acquires Emmi AI to Create th
    1. Mistral AI Acquires Emmi AI to Create the Leading AI Stack · Hacker News
    2. Qwen3.7-Max: The Agent Frontier · Hacker News
    3. Qwen3.7 Max scored by Artificial Analysis, 27B/35B waiting room · Reddit ML / LocalLLaMA
  133. 2026-05-19
    本期沒有新進展
  134. 2026-05-18
    本期沒有新進展
  135. 2026-05-17 · 第 12 期
    [Two Minute Papers] DeepSeek V4 AI Beats
    1. [Two Minute Papers] DeepSeek V4 AI Beats Billion Dollar Systems…For Free · YouTube AI Channels
  136. 2026-05-03
    本期沒有新進展
  137. 2026-04-29 · 第 10 期
    Why DeepSeek’s new model has been met wi
    1. Why DeepSeek’s new model has been met with a shrug · Wall Street Journal (WSJ)
  138. 2026-04-28 · 第 9 期
    deepseek-ai/DeepSeek-V3
    1. deepseek-ai/DeepSeek-V3 · GitHub Trending (Python)
    2. hydropix/TranslateBooksWithLLMs — Translate full-length books and documents with Ollama, OpenAI (comptatible), Gemini, Mistral, Poe or OpenRouter. Preserv · GitHub Trending (Python)
    3. TauricResearch/TradingAgents — TradingAgents: Multi-Agents LLM Financial Trading Framework · GitHub Trending (Python)
  139. 2026-04-27 · 第 8 期
    [AI Explained] GPT 5.5 Arrives, DeepSeek
    1. [AI Explained] GPT 5.5 Arrives, DeepSeek V4 Drops, and the Compute War Intensifies · YouTube AI Channels
  140. 2026-04-23
    本期沒有新進展
  141. 2026-04-21 · 第 5 期
    Qwen3.6-Max-Preview: Smarter, Sharper, S
    1. Qwen3.6-Max-Preview: Smarter, Sharper, Still Evolving · Hacker News
    2. Kimi K2.6 Released (huggingface) · Reddit ML / LocalLLaMA
    3. Kimi K2.6 · Reddit ML / LocalLLaMA
  142. 2026-04-20 · 第 4 期
    Qwen3.5-Omni Technical Report
    1. Qwen3.5-Omni Technical Report · Hugging Face Papers
    2. "Browser OS" implemented by Qwen 3.6 35B: The best result I ever got from a local model · Reddit ML / LocalLLaMA
    3. Qwen 3.6 Max Preview just went live on the Qwen Chat website. It currently has the highest AA-Intelligence Index score among Chinese models (52) (Will it be open source?) · Reddit ML / LocalLLaMA
  143. 2026-04-19 · 第 3 期
    unslothai/unsloth — Unsloth Studio is a
    1. unslothai/unsloth — Unsloth Studio is a web UI for training and running open models like Gemma 4, Qwen3.5, DeepSeek, gpt-oss locally. · GitHub Trending (Python)
    2. RTX 5070 Ti + 9800X3D running Qwen3.6-35B-A3B at 79 t/s with 128K context, the --n-cpu-moe flag is the most important part. · Reddit ML / LocalLLaMA
    3. Qwen3.6-35B-A3B solved coding problems Qwen3.5-27B couldn’t · Reddit ML / LocalLLaMA
  144. 2026-04-18 · 第 2 期
    Qwen 3.6 35B crushes Gemma 4 26B on my t
    1. Qwen 3.6 35B crushes Gemma 4 26B on my tests · Reddit ML / LocalLLaMA
    2. Qwen3.6. This is it. · Reddit ML / LocalLLaMA
    3. Qwen 3.6 is the first local model that actually feels worth the effort for me · Reddit ML / LocalLLaMA
  145. 2026-04-17 · 第 1 期
    Qwen3.6-35B-A3B: Agentic coding power, n
    1. Qwen3.6-35B-A3B: Agentic coding power, now open to all · Hacker News
    2. PSA: Qwen3.6 ships with preserve_thinking. Make sure you have it on. · Reddit ML / LocalLLaMA