[AINews] Gemini 4 Argon：GDM 对 Astra/Fable 的回应，输出上限达 100 万 Token

GDM 上一次发布比 Flash 更大的模型是在 2 月（3.1 Pro）。在连续推出渐进式 3.x Flash 版本、以及上个月发生重大 GDM 管理层变动后，GDM 面临的最大问题是：它何时才能追上同时期已经推出 Fable 和 Astra 级模型的同行。

好消息是，Argon 已经来了，其基准成绩相当亮眼（在 19 个可信基准中的 13 个达到 SOTA）……但目前仅在有限的网络安全预览中开放。不过官方承诺将会在尽快扩大访问范围：

[Google DeepMind@GoogleDeepMind 推出 Gemini 4 Argon——我们的新一代前沿模型。它专为编程、企业知识工作和网络安全防御中的复杂工作流打造，今天起将通过 Fairwind Program 向一组受信任的测试者推出。2026 年 9 月 30 日晚上 8:03 · 342 万次浏览

1,180 条回复 · 2,810 次转发 · 2.64 万个赞](https://x.com/GoogleDeepMind/status/2105388084154056939)

我们尤其关注实验性的 Long Decode Continuation：这项功能作为行业首创，将输出 Token 数提升至最高 100 万。

2026/9/29 至 2026/9/30 的 AI 新闻。我们查看了 12 个 subreddit、544 个 Twitter 账号，没有再查看其他 Discord。AINews 网站可供搜索所有往期内容。提醒一下，AINews 现已成为 Latent Space 的一个栏目。你可以选择订阅或取消订阅不同的邮件发送频率！

Gemini 4 Argon：Google 重返前沿阵营

• 发布：Google DeepMind 推出了 Gemini 4 Argon，面向编程、企业知识工作和网络防御（@GoogleDeepMind、@sundarpichai）。

  • 可用性：访问权限首先面向 Fairwind Program 中的政府用户和受信任网络防御者。Google 表示，在向开发者、企业和消费者开放之前，将进一步完善安全护栏（@Google、@demishassabis）。

  • 输出上限：Google 宣称其输出上限达到业界领先的 100 万 Token，高于此前的 6.4 万 Token（@GoogleAI、@TheRundownAI）。

    • 测量说明：Vals 列出的最大输出为 26.2 万 Token。Artificial Analysis 则借助 Long Decode Continuation 达到了 100 万输出 Token；这是一项新的 API 功能，可暂停长回复并在后续调用中继续生成（@ValsAI、@ArtificialAnlys）。

  • 定价：标准价格为每 100 万输入/输出 Token 4 美元/20 美元。50% 的首发折扣将价格降至 2 美元/10 美元，尚未公布结束日期。缓存输入可享受 95% 折扣（@_philschmid、@ArtificialAnlys）。

• Google 宣称的结果：在对阵 GPT-6 Astra 和 Claude Opus 5.5 的 19 项已公布基准中，Argon 有 13 项排名第一。在 DeepSWE 上，它取得 77.9%，高于 Opus 5.5 的 74.2% 和 Astra 的 74.1%（@TheRundownAI）。

  • 内部部署：Google 报告称，Argon Agent 已释放超过 300 TiB 的数据中心内存，并正在将超过 80 万行 C/C++ 内核代码迁移至 Rust（@kimmonismus）。

  • 科研用途：团队表示，基于 Argon 构建的内部 Agent 循环帮助完成了 CK 猜想（@mirrokni）。

• Artificial Analysis 评测：Argon 在 Intelligence Index 上取得 53 分，与 GPT-6 Astra（53 分）持平，并略高于 GPT-6.1 Sol（52 分）（@ArtificialAnlys）。

  • 每任务成本：按折扣价格计算，其每项任务成本为 1.99 美元，低于 Astra 的 3.26 美元；按标准价格则会上升至 3.98 美元。

  • Agent 工作能力：它在 AutomationBench-AA 上以 77.5% 排名第一，在 Terminal Bench 4 上取得 57%，落后于 Sonnet 5.5、Opus 5.5 和 Astra。

  • 幻觉率：它在 AA-Omniscience 上的幻觉率为 15%，而 Astra 为 51%。代价是准确率较低：Argon 为 50%，Astra 为 63%（@aipulseda1ly）。

• Vals 评测：Argon 以 68.9% 的成绩位列 Vals Index 第一，平均每项任务成本为 15.68 美元（@ValsAI、@ValsAI）。

  • 编程：它完美构建了 30 个 Vibe Code Bench 应用，而 Opus 5 为 25 个、Astra 为 24 个（@ValsAI）。

  • 终端与安全：Terminal-Bench 4.0 的成绩从 19.0% 升至 57.6%。它在 CyberBench 概念验证任务中得分 70%，在 IOI 2024–2026 中得分 100%（@ValsAI）。

  • 效率：在 Vals Index 任务中，它使用的输出 Token 约为 Sonnet 5.5 的四分之一（@ValsAI）。

• Arena 与其他评测：Argon 在 Text Arena 中以 1525 分排名第一，在 Code Arena WebDev 中以 1679 分排名第八（@arena）。

  • Agent Arena：在初步的 3,000 个会话中，它总排名第八，且可控性排名第一（@arena）。

  • PostTrainBench：其得分为 45.3%，高于 Gemini 3.1 Pro 的 21.99%（@karinanguyen）。

• 质疑：一些观察人士对已公布的数据提出了质疑。

  • 法律基准：Argon 在 Harvey 法律基准上报告的 19.6%，低于 Muse Spark 1.2 所列的 25.42%（@BlackHC）。

  • 其他批评：评论人士提出了可能存在基于偏好数据进行 benchmark 优化的问题，并反对部分数据，包括 DeepSWE 成绩（@teortaxesTex、@teortaxesTex）。

GPT-6.1 Sol 与 OpenAI 的 DevDay Agent 技术栈

• 独立评测：GPT-6.1 Sol 成为 MathArena 新的第一名（@j_dekoninck）。

  • Code Arena：它在 WebDev 中以 1759 分排名第三；在同样的 2 美元/10 美元定价下，比 GPT-6 Sol 高出 70 分（@arena）。

  • 每任务成本：Artificial Analysis 测得，在最高推理强度下其每项任务成本为 0.72 美元，相比之下 Astra 为 3.26 美元，GPT-6 Sol 为 1.04 美元（@ArtificialAnlys）。

  • Luna Bug 修复：OpenAI 修复了一项图像编码 Bug，使 GPT-6 Luna 的 Intelligence Index 增加 1 分。

• 超高速推理：OpenAI 给出的速度最高可达 300 tok/s。SemiAnalysis 报告称，它在低批量规模下运行于 NVIDIA GPU，而非 Cerebras（@kimmonismus）。

• 产品层：DevDay 推出了 dots（拥有各自云端计算机的持久化 Agent）、Decisions API 和 computer use（@latentspacepod）。

  • Sites：ChatGPT Sites 现在可以托管 MCP 服务器，并将其转化为可安装的插件（@mxstbr）。

  • 使用限制：用户报告称获得了价值约 2,500 美元的一次性额度；也有用户抱怨使用限额被削减（@kimmonismus、@kimmonismus）。

其他发布：嵌入、图像/视频与开放模型

• Perplexity 上下文嵌入：pplx-embed-v2-context-9b-preview 已在 Hugging Face 上开放（@perplexity_ai）。

  • 方法：该模型对整篇文档仅编码一次，之后再汇聚分块向量。训练过程从上下文压缩模型中蒸馏相关性，而不是使用单一的黄金分块标签（@denisyarats）。

  • 结果：它在 ConTEB 上创下新的 SOTA。在 turbopuffer 的私有 context-bench 上，它在 answer recall@10 上比 voyage-context-4 高 14.4 分，同时使用 1 KB int8 向量，而后者为 8 KB（@turbopuffer）。

• Cohere Embed 5：该系列提供 Pro 和 Fast 两种变体，并使用共享嵌入空间，因此可以用其中一个建索引、用另一个进行检索（@cohere）。

• Ideogram 4.5：这款编辑模型聚焦于无伪影的多轮编辑，并承诺将开放权重（@ideogram_ai）。

  • 编辑保真度：在连续十次编辑后，94–99% 未修改的内容仍保持完全一致（@fal）。

  • 排名：它在 Image Edit Arena 中以 1351 分排名第 18（@arena）。

• 视频基准：Artificial Analysis 推出了 AA-Video-T2V v2.0，在 1080p 分辨率下评测，汇集了超过 6.8 万张人工投票（@ArtificialAnlys）。

  • 领先者：Wan 3.0 以每分钟 12 美元排名第一。Seedance 2.5 以每分钟 34.12 美元排名第二，MiniMax H3 则以每分钟 4.80 美元的成本与其在统计上持平。

  • Utopai X：这一 MiniMax H3 的后训练版本首次亮相即排名第二（@ArtificialAnlys）。

• 开放与小型模型：

  • Ling-3.1-flash：一款 5,000 亿参数模型，据称性能接近 GPT-5.6 Sol 和 Opus 5（@kimmonismus）。它在 Mobile App Arena 的开放权重模型中排名第二（@DesignArena）。

  • Praxis-1：Runway 发布了一款开放权重的世界动作模型，并表示机器人策略性能会随第三人称视频规模稳定提升（@agermanidis）。

  • Solar Mini 4：Upstage 报告该模型总参数为 350 亿、激活参数为 30 亿。它在 Intelligence Index 上得分 24，定价为 0.10 美元/0.40 美元（@ArtificialAnlys）。

Agent 研究、推理与系统

• Context Language Models（Meta）：CLM 将上下文视为可编辑文件，而非只能追加的日志；上下文管理策略被学习进模型权重，无需外部 harness（@RulinShao）。

• 自适应推理计算：

  • TaH2：前瞻深度监督会教导模型，哪些困难 Token 值得再进行一轮推理（@ZhihuFrontier）。

  • AutoBenchmark（Meta）：该项目将基准创建自动化。在构思阶段引入人类反馈优于让 Agent 单独工作，且难度能够迁移至留出的求解器（@jaseweston）。

  • Stratego：一篇 Nature 论文展示了首个超人类水平的 Stratego AI，其基于强化学习和不完全信息环境下的测试时计算构建（@ssokota）。

• Prefill/decode 解耦：一项稳态分析指出，在相同批量规模与吞吐量下，解耦可将平均交互性提高约 1/(decode 时间占比)（@ekzhang1、@cHHillee）。

• 编译器与硬件：

  • DeepSeek 与华为：DeepSeek 发布了开源 Ascend 工具包，其中包含针对 Ascend 950 优化的 TileLang（@kimmonismus）。

  • AI 作为编译器：一个模型可将 Triton 直接翻译为 PTX，并由验证器检查正确性、竞争条件和死锁。在 B200 上，FlashAttention 的加速可达 1.37 倍（@Azaliamirh）。

  • Vera Rubin：Cognition 通过 CoreWeave 成为 Vera Rubin 的首个客户，并报告称在相同 decode 速度下，其 Token 吞吐量约为 GB200 的 4.8 倍（@cognition）。

  • DFlash drafts：针对 Ornith-1.5 的新草稿模型可带来最高 2.54 倍的无损加速（@ornith_）。

• Agent 沙箱：Cloudflare 为 Agent 重构了 Containers，p50 可交互时间为 648 毫秒，速度提升 6 倍；快照功能目前处于 Beta 阶段（@mgamache）。

安全、网络安全与评测完整性

• 推理提取：OpenAI 将一场隐藏推理提取活动的核心部分归因于与 Moonshot AI 有关联的个人（@kimmonismus）。

  • 规模：OpenAI 在两天内记录到来自逾 4,000 名用户的 16,000 次尝试，相关活动涉及超过 15,000 名用户。

  • 外部研究者：他们的攻击直到本周仍能在 Astra 上奏效。补丁很难在不同产品版本和第三方托管平台间传播（@JSchaeff3r、@jonasgeiping）。

  • **批

GDM last shipped a larger-than-Flash model in February (3.1 Pro), and after successive incremental 3.x Flash versions and the big GDM management shakeup last month, the largest question for GDM was when they would catch up to peers who have in the meantime launched Fable and Astra class models.

Well, Argon’s here, with VERY respectable benchmarks (SOTA in 13 of 19 credible benchmarks)… but only accessible in limited cybersecurity preview, though access is promised “as soon as possible”:

[Google DeepMind@GoogleDeepMindIntroducing Gemini 4 Argon – our new frontier model. It’s built for complex workflows across coding, enterprise knowledge work, and cybersecurity defense – rolling out today to a set of trusted testers through our Fairwind Program. 8:03 PM · Sep 30, 2026 · 3.42M Views

1.18K Replies · 2.81K Reposts · 26.4K Likes](https://x.com/GoogleDeepMind/status/2105388084154056939)

We like the experimental Long Decode Continuation, which increases output tokens up to 1M as an industry first.

AI News for 9/29/2026-9/30/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

Gemini 4 Argon: Google Returns to the Frontier

• Launch : Google DeepMind introduced Gemini 4 Argon for coding, enterprise knowledge work and cyber defense (@GoogleDeepMind, @sundarpichai).

  • Availability : Access starts with government users and trusted cyber defenders in the Fairwind Program. Google says it will refine guardrails before opening access to developers, enterprises and consumers (@Google, @demishassabis).

  • Output limit : Google cites an industry-leading 1M-token output limit, up from 64K (@GoogleAI, @TheRundownAI).

    • Measurement note : Vals lists 262K max output. Artificial Analysis reached 1M output tokens through Long Decode Continuation, a new API feature that pauses long responses and resumes them across calls (@ValsAI, @ArtificialAnlys).

  • Pricing : Standard pricing is $4/$20 per 1M input/output tokens. A 50% introductory discount brings it to $2/$10, with no end date announced. Cached input gets a 95% discount (@_philschmid, @ArtificialAnlys).

• Google’s claimed results : Argon takes first place on 13 of 19 published benchmarks against GPT-6 Astra and Claude Opus 5.5. On DeepSWE it scores 77.9%, versus 74.2% for Opus 5.5 and 74.1% for Astra (@TheRundownAI).

  • Internal deployments : Google reports that Argon agents freed more than 300 TiB of data-center memory and are migrating more than 800K lines of C/C++ kernel code to Rust (@kimmonismus).

  • Research use : The team says internal agent loops built on Argon helped complete the CK conjecture (@mirrokni).

• Artificial Analysis evaluation : Argon scores 53 on the Intelligence Index, matching GPT-6 Astra (53) and edging GPT-6.1 Sol (52) (@ArtificialAnlys).

  • Cost per task : At discounted pricing it costs $1.99 per task versus $3.26 for Astra; standard pricing would raise this to $3.98.

  • Agentic work : It ranks #1 on AutomationBench-AA at 77.5% and scores 57% on Terminal Bench 4, behind Sonnet 5.5, Opus 5.5 and Astra.

  • Hallucination : Its 15% rate on AA-Omniscience compares with 51% for Astra. The tradeoff is lower accuracy: 50% versus Astra’s 63% (@aipulseda1ly).

• Vals evaluation : Argon is #1 on the Vals Index at 68.9%, at an average $15.68 per task (@ValsAI, @ValsAI).

  • Coding : It built 30 Vibe Code Bench apps perfectly, against 25 for Opus 5 and 24 for Astra (@ValsAI).

  • Terminal and security : Terminal-Bench 4.0 rose from 19.0% to 57.6%. It scores 70% on CyberBench proof-of-concept tasks and 100% on IOI 2024–2026 (@ValsAI).

  • Efficiency : It uses about a quarter of Sonnet 5.5’s output tokens on Vals Index tasks (@ValsAI).

• Arena and other evals : Argon is #1 in Text Arena at 1525 and #8 in Code Arena WebDev at 1679 (@arena).

  • Agent Arena : It ranks #8 overall and #1 for steerability on a preliminary 3K sessions (@arena).

  • PostTrainBench : It scores 45.3%, up from 21.99% for Gemini 3.1 Pro (@karinanguyen).

• Skepticism : Some observers questioned the published numbers.

  • Legal benchmark : Argon’s reported 19.6% on Harvey’s legal benchmark trails Muse Spark 1.2’s listed 25.42% (@BlackHC).

  • Other critiques : Commentators raised possible preference-data benchmaxxing and objected to some figures, including DeepSWE (@teortaxesTex, @teortaxesTex).

GPT-6.1 Sol and OpenAI’s DevDay Agent Stack

• Independent evals : GPT-6.1 Sol is the new #1 on MathArena (@j_dekoninck).

  • Code Arena : It ranks #3 on WebDev at 1759, 70 points above GPT-6 Sol for the same $2/$10 pricing (@arena).

  • Cost per task : Artificial Analysis measures $0.72 per task at max effort, versus $3.26 for Astra and $1.04 for GPT-6 Sol (@ArtificialAnlys).

  • Luna bug fix : OpenAI fixed an image-encoding bug, adding 1 Intelligence Index point to GPT-6 Luna.

• Ultrafast inference : OpenAI quotes up to 300 tok/s. SemiAnalysis reports it runs on NVIDIA GPUs at low batch sizes, not on Cerebras (@kimmonismus).

• Product layer : DevDay introduced dots (persistent agents with their own cloud computers), a Decisions API and computer use (@latentspacepod).

  • Sites : ChatGPT Sites can now host MCP servers and turn them into installable plugins (@mxstbr).

  • Usage limits : Users report one-off credits worth about $2,500. Others complain that usage limits were cut (@kimmonismus, @kimmonismus).

Other Releases: Embeddings, Image/Video and Open Models

• Perplexity contextual embeddings : pplx-embed-v2-context-9b-preview is open on Hugging Face (@perplexity_ai).

  • Method : The model encodes the whole document once and pools chunk vectors afterward. Training distills relevance from a context-compression model instead of using single gold-chunk labels (@denisyarats).

  • Results : It sets a new state of the art on ConTEB. On turbopuffer’s private context-bench it beats voyage-context-4 by 14.4 points in answer recall@10, using 1 KB int8 vectors against 8 KB (@turbopuffer).

• Cohere Embed 5 : The family has Pro and Fast variants in a shared embedding space, so you can index with one and retrieve with the other (@cohere).

• Ideogram 4.5 : The editing model targets artifact-free multi-turn edits, with open weights promised (@ideogram_ai).

  • Edit fidelity : Over ten consecutive edits, 94–99% of untouched content stays identical (@fal).

  • Ranking : It is #18 in Image Edit Arena at 1351 (@arena).

• Video benchmark : Artificial Analysis launched AA-Video-T2V v2.0, judged at 1080p with more than 68K human votes (@ArtificialAnlys).

  • Leaders : Wan 3.0 is #1 at $12/min. Seedance 2.5 is #2 at $34.12/min, and MiniMax H3 is statistically tied at $4.80/min.

  • Utopai X : This post-train of MiniMax H3 debuts at #2 (@ArtificialAnlys).

• Open and small models :

  • Ling-3.1-flash : A 500B model reported close to GPT-5.6 Sol and Opus 5 (@kimmonismus). It ranks #2 among open-weight models in Mobile App Arena (@DesignArena).

  • Praxis-1 : Runway released an open-weight world-action model and says robotics policy performance scales predictably with third-person video (@agermanidis).

  • Solar Mini 4 : Upstage reports 35B total / 3B active parameters. It scores 24 on the Intelligence Index at $0.10/$0.40 (@ArtificialAnlys).

Agent Research, Inference and Systems

• Context Language Models (Meta) : CLMs treat context as an editable file rather than an append-only log, with context-management policies learned in the weights and no external harness (@RulinShao).

• Adaptive reasoning compute :

  • TaH2 : Lookahead depth supervision teaches the model which hard tokens deserve another loop (@ZhihuFrontier).

  • AutoBenchmark (Meta) : The project automates benchmark creation. Human feedback at the ideation stage beats agents working alone, and difficulty transfers to held-out solvers (@jaseweston).

  • Stratego : A Nature paper presents the first superhuman Stratego AI, built on RL and test-time compute under imperfect information (@ssokota).

• Prefill/decode disaggregation : A steady-state analysis argues that disaggregation raises mean interactivity by about 1/(decode-time fraction) at equal batch size and throughput (@ekzhang1, @cHHillee).

• Compilers and hardware :

  • DeepSeek on Huawei : DeepSeek released an open-source Ascend toolkit with TileLang optimized for Ascend 950 (@kimmonismus).

  • AI as compiler : A model translates Triton directly to PTX, with a verifier checking correctness, races and deadlocks. Speedups on B200 reach 1.37x on FlashAttention (@Azaliamirh).

  • Vera Rubin : Cognition is the first customer on Vera Rubin via CoreWeave, reporting about 4.8x the token throughput of GB200 at the same decode speed (@cognition).

  • DFlash drafts : New draft models for Ornith-1.5 give up to 2.54x lossless speedups (@ornith_).

• Agent sandboxes : Cloudflare rebuilt Containers for agents, with p50 time-to-interactive of 648 ms (6x faster) and snapshots in beta (@mgamache).

Safety, Security and Eval Integrity

• Reasoning extraction : OpenAI attributes a core part of a hidden-reasoning extraction campaign to individuals linked to Moonshot AI (@kimmonismus).

  • Scale : OpenAI recorded 16,000 attempts from more than 4,000 users in two days, with related activity across more than 15,000 users.

  • External researchers : Their attacks kept working on Astra until this week. Patches were hard to propagate across product versions and third-party hosts (@JSchaeff3r, @jonasgeiping).

  • Criticism : Nathan Lambert argues the vulnerability is the API provider’s responsibility (@natolambert).

• Distillation defenses : Defenses evaluated without later RL give a false sense of security. RL makes simple attacks effective (@shidan_javaheri).

• Embedded evaluations : Apollo Research published principles for outside evaluators who receive employee-like access to frontier labs (@ApolloResearch).

• Cyber evals : On CyberGym-E2E-AA, some frontier models are safety-blocked on more than 85% of tasks (@ArtificialAnlys).

• Provenance and transparency :

  • SynthID Bio : Watermarking for AI-generated proteins is published in Nature, with open-sourced tools (@demishassabis).

  • AI-detector evasion : Opus 5.5 and Astra can rewrite more than 50% of a document without Pangram flagging it (@ValsAI).

  • Agent reports : A new preprint asks how transparent LLM-written reports on agent work actually are (@jennyihuang).

Industry and Policy

• Factory vs Cognition : Factory removed advisor Chris Degnan, alleging he was confiding in Cognition while attending its board meetings (@matanSF).

• Political spending : Greg Brockman dropped a promised second $25M donation to the Leading the Future super PAC (@teddyschleifer).

• OpenAI finances : NYT reports OpenAI is near $70B in annualized revenue and in talks to raise $30B at a $1.4T valuation, with its IPO pushed to next year (@srimuppidi).

• Funding : Flow, which builds AI tooling for hardware engineering, raised a $50M Series B at a $750M valuation (@parisingh).

Top tweets (by engagement)

• GLM-5.3 and the Spread of Advanced Cyber Capabilities \ Anthropic (Activity: 785): Anthropic reports that Zhipu/Z.ai’s open-weightGLM-5.3 crosses a notable threshold for autonomous cyber capability: 50/410end-to-end V8 exploits on ExploitBench, close to Claude Mythos Preview’s56/410, plus full control-flow hijacks on4%of Anthropic’s internal binary exploitation tasks where prior models were near zero. Anthropic frames the risk as capability + accessibility: GLM-5.3 is widely downloadable, relatively cheap, and weakly refusal-tuned, with simple jailbreaks reportedly succeeding64–100%of the time and “abliteration” dropping refusals to low single digits with little measured capability degradation. Top comments were largely hostile to Anthropic’s framing, arguing the post reads as an attempt to suppress a cheaper/open Chinese model near Anthropic’s frontier. One commenter emphasized legitimate defensive use, saying GLM-5.3 is their only practical tool for security testing and improving their own software.

  • Commenters highlight GLM-5.3 as a low-cost, less-restricted model perceived to be close to frontier capability, with one user framing it as useful for “security testing and improvements on my own software” rather than inherently malicious. The technical concern raised is that restrictions by providers like Anthropic could limit defensive cybersecurity workflows that require models willing to analyze potentially sensitive exploit or vulnerability patterns.

  • One commenter references prior GLM-5.2 models as having helped mitigate a Hugging Face attack , contrasting that with Claude allegedly refusing assistance. The substantive point is that refusal policies may reduce utility in incident response or vulnerability remediation scenarios, while more permissive models can be operationally useful for defensive security tasks.

• add GLM-5.3-Flash (GLM5-Next) support by timkhronos · Pull Request #27773 · ggml-org/llama.cpp (Activity: 348): Mergedggml-org/llama.cpp#27773adds GLM-5.3-Flash / GLM5-Next support tollama.cpp, enabling local inference for the 320B hybrid text+vision model. The implementation adds GLM-specific DSA indexing/pooling, hybrid indexed memory, and a newglm5vvision preprocessing/tower path, while reusing Kimi-K3 KDA layers, DeepSeek-style MoE/mHC helpers, MLA-only attention, and DSV4-style SwigLU clamping; validation reports random-model logits matching Transformers across prefill/ubatching/decode and vision embedding agreement around1e-5, with some precision-sensitive tensors left unquantized. Commenters were concerned that llama.cpp model support is lagging behind the pace of new experimental architectures, with one noting the effective bottleneck appears to be maintainer availability. A technical compatibility issue was also raised: existing Unsloth quantizations reportedly use glm5next while mainline expects glm5-next, so current mainline may fail to load those quants.

  • Commenters noted a compatibility issue between the Unsloth quantization PR and the mainline llama.cpp PR: one identifies the architecture/model type as glm5next while the other uses glm5-next, meaning mainline llama.cpp may fail to load existing Unsloth GLM-5.3-Flash quants without conversion or metadata fixes.

  • There was concern that llama.cpp support is lagging behind the pace of new model releases, especially as newer models increasingly use experimental architectures that require bespoke loader/runtime changes before inference and optimization work can land. One commenter framed GLM-5.3-Flash support as taking roughly “another month” after model release, with progress depending heavily on a small number of maintainers.