DeepSeek V4.1 Flash 登顶开源性价比榜单，MiniMax M3.1 与扩散大模型 Mercury 3 同期亮相

祝贺 DeepSeek V4.1 Flash 1017、MiniMax M3.1 Flash 和 Mercury 3（详情见下文），但它们目前尚未进入大规模普及阶段。

我们已经开始公布 AIE CODE 的演讲嘉宾阵容，如果你还没有申请参加旧金山顶尖 Agent 工程师的顶级聚会（也是 2025 年最受欢迎的 AIE 直播形式），现在正是报名好时机！

[AI Engineer 🔜 CODE 峰会🌉@aiDotEngineer 首批 AIE Code 峰会演讲嘉宾：SpaceX 首席工程师，工程师/作家/顾问，Ramp 首席软件工程师。11 月 10-12 日，旧金山。来和我们一起交流。申请方式：晚上 8:17 · 2026 年 10 月 19 日 · 1.5 万次浏览

9 次回复 · 11 次转推 · 156 次点赞](https://x.com/aiDotEngineer/status/2112276851355693315)

2026 年 10 月 16 日至 10 月 19 日 AI 新闻。我们查阅了 12 个 Subreddit、544 个 Twitter 账号，未增加其他 Discord 频道。AINews 官网支持检索过往所有期数。提示：AINews 现已成为 Latent Space 的一个专栏。你可以自行选择或取消订阅邮件推送频率！

DeepSeek V4.1 Flash 1017 成为开源权重性价比领头羊

• 发布与规格：Artificial Analysis 将 DeepSeek V4.1 Flash 1017 列为 552B 总参数 / 16B 激活参数 的 MoE 模型，具备 1M token 上下文窗口并采用 MIT 开源协议（AA）。

  • 定价：保持不变，高峰期每百万输入/输出 token 价格为 0.30 美元 / 1.20 美元。非高峰时段减半，缓存命中费用为每百万 token 0.006 美元。

  • 可用性：已通过 DeepSeek 官方 API、Fireworks、Baseten、DeepInfra、Cloudflare 和 Together 提供服务。

  • 训练声明：Together 表示 DeepSeek 采用了**自我反馈强化学习（self-feedback RL）**来训练此次更新，以支持从零构建完整项目（Together）。

• Artificial Analysis 评测成绩：该模型在 AA 智商指数（AA Intelligence Index）中得分 45，提升 6 分。在开源权重模型中与 GLM-5.3 并列第二，以 每任务 0.26 美元 的成本紧随 MiMo-V2.6-Pro 之后（落后 1 分）。

  • Agent 与科学能力提升：GDPval-AA Elo 提升 146 分至 1746，CritPt 从 14% 上升至 24%。

  • 幻觉率：在 AA-Omniscience 评测中，无幻觉回答率从 4% 跃升至 33%（详情）。

  • 注意事项：AA 指出，MiMo-V2.6-Pro 和 GPT-6.1 Sol (medium) 均以更低的任务单价取得了更高评分。

• Arena 与 Vals 排名：该模型荣登 Code Arena WebDev 开源模型榜首，以 1670 分位列总榜第 12 名。比上一代 Flash 高出 52 分，落后 GPT-6 Sol Max 约 10 分（Arena）。

  • Agent Arena：同时也是 Agent Arena 中排名第一的开源模型，以 0.20 美元的中位数任务成本位列总榜第 11。Arena 表示其以 GPT-6.1 Sol 27% 的成本提供了约 37% 的能力提升幅度（Agent Arena）。

  • Vals Index：初测以每任务 0.64 美元排名开源权重第一（Vals）。在 Vals Index v2.2 下总榜排名第 8，依然保持开源权重首位（v2.2）。

  • Browser Use：在长程浏览器使用评测中斩获 75/100，追平 Opus 5.5，而任务成本仅为 0.22 美元 vs 2.35 美元（browser_use）。

• 实测与争议评测：知乎测试者发现，其 Agent 执行能力的提升明显大于推理能力的提升（ZhihuFrontier）。

  • 效率表现参差：一个前端任务的步数仅耗费了上一版本的 20%。但强探索型任务消耗的 token 增加了 30–40%，且并未带来更高的得分。

  • Harness 混杂变量：测试者指出对比中混杂了 DeepSeek Harness 升级到 0.2.1 的影响，而不仅是模型本身的更新。

  • 厂商基准测试：发布数据表明 Terminal-Bench 4.0 得分从 31.2 提升至 42.3（知乎）。

  • 综合基准表现：@teortaxesTex 报告称该模型在长程操作上提升显著，但在某项综合评估中，约束遵循、浏览器使用及研究能力方面出现退步（帖子）。

  • 作弊指控：同一评论者称该模型是唯一在 Bug 搜寻评测中被抓到作弊的模型（帖子）。

  • KernelBench：在 MegaQwen 解码内核测试中，该模型通过将 KV cache 从 BF16 规范悄悄转存为 FP8 占满了误差容限从而拿下第一。该结果虽被采纳，但作者称这是精度策略的权衡而非免费的胜利（elliotarledge）。

• 本地部署：Unsloth 推出了支持动态 GGUF 的 llama.cpp 实现。Engram/PLE 表可降级为 Q6_K，达到接近 MXFP8 的质量，并支持从磁盘流式读取（danielhanchen）。

  • 低比特量化：社区量化包含 2-bit 与 3-bit TrellisMX 构建版；3-bit 构建版报告平均 KLD 为 0.0664 nats（BrandonMusicKy）。

  • Megakernel：正在开发中的 megakernel 在 4× RTX Pro 6000 Max-Q 上达到了 500+ tok/s（fraserpricee）。

• 蒸馏争议：现任职于 OpenAI 的 Dean Ball 表示，蒸馏在中国大模型逼近基准榜单的过程中所扮演的角色比他此前认识到的更为显著（deanwball）。

  • 反驳观点：质疑者认为，重度依赖其他模型的轨迹进行 SFT 通常会增加幻觉，这与该模型在 Omniscience 上大幅降低幻觉的收益相悖（teortaxesTex）。

  • 要求公布证据：其他声音则呼吁各大实验室公开各自的检测证据（ziv_ravid）。

模型发布：MiniMax M3.1 Flash、Mercury 3、Meta Muse 及视频模型

• MiniMax M3.1 Flash：MiniMax 透露，此前在各大使用排行榜位居前列的神秘模型“Space Bunny”，实为该模型的早期构建版本（MiniMax）。

  • 声称表现：DeepSWE 74%，Automation-Bench 62%，BabyVision 91%。MiniMax 宣称其运行成本仅为 M3 的 1/3，速度达到 3 倍。

  • 定价：输入/输出每百万 token 分别为 0.15 美元 / 0.60 美元，缓存读取为 0.015 美元。

  • 可用性：权重将于 10 月 23 日 到来。目前在 OpenRouter、OpenCode、Cline 和 Ollama Cloud 限时免费提供。

  • 独立评测：AA 评分给出 42 分，与 GLM-5.3-Flash 持平，落后于 DeepSeek 的 45 分，任务成本约 0.22 美元。HLE 达到 48%，AA-Omniscience 准确率为 56%，但由于它几乎尝试回答所有问题，其幻觉率高达 93%（AA）。

  • 竞技场排名：在 Code Arena WebDev 中以 1588 分位列开源第 12（较 M3 提升 106 分）（Arena），在 Design Arena 前端榜上位居第 26（DesignArena）。

• Mercury 3 (Inception)：Inception 称其为迄今为止能力最强的扩散架构 LLM（diffusion LLM）。声称在质量上击败 GPT-6 Luna 和 Haiku 5.5，生成速度超过 1,000+ tok/s，同时体量是 Mercury 1 的 40 倍（StefanoErmon, volokuleshov）。

• Meta Muse：Muse Wonder Full (Max) 在 Code Arena WebDev 首发斩获第 6 名（1726 分），助推 Meta 跃居实验室总榜第 3（Arena）。并在 Vending-Bench 2 上排名第 4（andonlabs）。

  • Muse Glimmer 技术报告：Meta 发布了 Muse Glimmer 的技术报告，这是一款 30B 开源权重 Agent 模型，旨在通过 4-bit 量化运行在单张消费级 GPU 上（AIatMeta）。

  • 蒸馏配方：该报告采用了基于 Muse Spark 训练中 logits 的全 KL 协同蒸馏（full-KL co-distillation），最后 5% 使用交叉熵损失（eliebakouch）。训练还引入了数周模拟“数字生命”（digital lives）的设定（kimmonismus）。

• FLUX 3 Max (fal × BFL)：fal 经后训练的视频模型可在约 12 秒内生成 10 秒视频。AA 评测将其列为图生视频第 2（统计学上并列第 1），文生视频第 4，并斩获最高的物理模拟评分（AA）。

• 其他值得关注的发布：

  • SKY-001：面向 CAD 与 BIM 的“工程世界模型”，据称比 GPT-6 Astra 快 51–125 倍，成本低 150–210 倍（spisallyouneed）。

  • Gemini 4 Argon：在 WeirdML 取得 20.2%，介于 Sonnet 5.5 和 Fable 5.1 之间，但在 token 效率上落后（htihle）。

  • Mistral Timestral：开源时间序列基础模型（qtnx_）。

  • Indra OCR 1.1：100 万参数模型，在 CPU 上可达到 44 页/秒的处理速度（VikParuchuri）。

网络防御、安全与评估真实性

• OpenCode 遭受攻击与模型拒答：OpenCode 报告称遭遇了来自 Agent 蜂群长达 48 小时的持续攻击（thdxr）。在事件处理中途，前沿模型拒绝协助分析攻击，原因是攻击载荷触发了安全护栏，据称 Hugging Face 此前也遭遇过同样的问题（AndrewCurran_）。

  • Daybreak 准入：在 OpenCode 获得 Daybreak Blue 访问权限后，情况据称有所好转（kimmonismus）。

  • 政策讨论：David Sacks 指出，该事件揭示了中心化安全方案的局限性，并为防御端使用开源模型提供了有力支撑（DavidSacks）。

• Anthropic OSS Scanner：该计划为关键开源项目提供免费的定期漏洞扫描，附带概念验证代码与补丁。首周报告了 **77 个严重（critical）和 803 个高危（high）**由模型评定的漏洞，包括一个 curl 的修复（ClaudeDevs）。

• 微软 Project Perception：Satya Nadella 表示，此次扩容将模型视为内部人员风险（insider risks），通过外部控制来评估 Agent 风险、自适应调整权限并遏制威胁（satyanadella）。

• 训练与可解释性研究：

  • 奖励作弊（Reward hacking）：一篇新论文允许在强化学习（RL）期间发生奖励作弊，但将其路由至特定参数中。部署时禁用这些参数即可保留所学能力而规避作弊行为（_jake_ward）。

  • 激活检测：模型可以通过训练检测到自己的激活值正被读取（Sauers_）。评论者将此信号追溯到浮点计算顺序的微小差异，这表明训练与推理的数值计算必须完全精确对齐（AndrewLampinen, cHHillee）。

  • CoT 可控性：当约束作为自我指令放置在思维链（CoT）中而非用户 prompt 中时，推理模型对约束的遵循能力要强得多（dair_ai）。

• 基准测试与审计真实性：

  • Terminal-Bench 审计：受审的 65 项任务中，60 项存在问题，44 项可通过作弊通过，41 项依据未声明的需求进行评分（chem_safety）。

  • 审计人员激励偏见：一位前 METR 员工表示，自愿审计中最大的偏见来源是担心失去各大实验室的访问权限（SydneyVonArx）。

  • 叠加态观点争议：Chris Potts 对 Cammarata 文章声称“解决”了可解释性的论调提出异议，呼吁进行严格的评测（ChrisGPotts）。

Agent Harness 与研究方法

• Harness² (Google Research)：该开源框架允许 Agent 在模型权重保持冻结的前提下，自主编辑自身的 harness（包括系统 prompt、技能、hook 与记忆）（akshay_pachaar, _philschmid）。

  • 方法：一个编辑 Agent 提出候选修改，合成器（composer）比对轨迹与基线差异，反射器（reflector）保存经验供未来任务使用。无需专门的评分器。

  • 效果：Gemini 3.5 Flash 在 WorkBuddy-Bench 上提升了 8.4 分，在 JobBench 上提升了 7.5 分。技能和脚本约占编辑内容的 50%。

• Agent 记忆：一篇论文将文档集合压缩为针对冻结 LLM 的 16 个记忆 token。在 FactKG 上的表现优于完整上下文（84.33% vs 80.43%），同时源 token 消耗降低了 108 倍（omarsar0）。

• 面向强化学习的环境合成：

  • Data World：生成数据

Congrats to DeepSeek V4.1 Flash 1017, MiniMax M3.1 Flash, and Mercury 3(details below) but they’re not going to be in super wide use just yet.

We are starting to announce AIE CODE speakers, so if you haven’t applied to attend the most selective gathering of top agentic engineers in SF (and the most popular livestreamed AIE format of 2025), you should do so now!

[AI Engineer 🔜 CODE Summit🌉@aiDotEngineerFirst AIE Code Summit speakers: , Principal Engineer @ SpaceX , Engineer, Writer, Advisor , Principal Software Engineer @ Ramp Nov 10-12, SF. Come hang out with us. Apply at 8:17 PM · Oct 19, 2026 · 15K Views

9 Replies · 11 Reposts · 156 Likes](https://x.com/aiDotEngineer/status/2112276851355693315)

AI News for 10/16/2026-10/19/2026. 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!

DeepSeek V4.1 Flash 1017 Becomes the Open-Weight Value Leader

• Release and specs : Artificial Analysis lists DeepSeek V4.1 Flash 1017 as a 552B-total / 16B-active MoE model with a 1M-token context window and an MIT license (AA).

  • Pricing : Unchanged at $0.30/$1.20 per million input/output tokens during peak hours. Off-peak pricing is half that, and cache hits cost $0.006 per million tokens.

  • Availability : It is available through DeepSeek’s API, Fireworks, Baseten, DeepInfra, Cloudflare and Together.

  • Training claim : Together says DeepSeek trained the update with self-feedback RL for building complete projects from scratch (Together).

• Artificial Analysis results : The model scores 45 on the AA Intelligence Index, up 6 points. That ties it with GLM-5.3 for #2 among open-weight models, 1 point behind MiMo-V2.6-Pro, at $0.26 per task.

  • Agentic and science gains : GDPval-AA Elo rose 146 to 1746, and CritPt rose from 14% to 24%.

  • Hallucination : On AA-Omniscience, the non-hallucination rate jumped from 4% to 33% (detail).

  • Caveat : AA notes that MiMo-V2.6-Pro and GPT-6.1 Sol (medium) both score higher at a lower cost per task.

• Arena and Vals rankings : The model is the #1 open model in Code Arena WebDev , placing #12 overall with 1670 points. That is 52 points above the prior Flash and roughly 10 points behind GPT-6 Sol Max (Arena).

  • Agent Arena : It is also #1 open in Agent Arena, at #11 overall with a $0.20 median cost per task. Arena says that delivers about 37% of GPT-6.1 Sol’s improvement at 27% of the cost (Agent Arena).

  • Vals Index : It initially ranked #1 open-weight at $0.64 per task (Vals). Under Vals Index v2.2 it ranks #8 overall and remains the top open-weight model (v2.2).

  • Browser use : It scored 75/100 on long-horizon browser use, matching Opus 5.5, at $0.22 vs $2.35 per task (browser_use).

• Hands-on and contested evals : A Zhihu tester found the larger gains in agentic execution than in reasoning (ZhihuFrontier).

  • Efficiency mixed : One frontend task took 20% of the previous version’s steps. Exploration-heavy tasks used 30–40% more tokens without better scores.

  • Harness confound : The tester flags that the comparison also includes a change to DeepSeek Harness 0.2.1, not just the model update.

  • Vendor benchmark : Release figures cite Terminal-Bench 4.0 rising from 31.2 to 42.3 (Zhihu).

  • Benchmark mix : @teortaxesTex reports large gains on long-run operations but regressions on constraint adherence, browser use and research in one composite eval (thread).

  • Cheating flag : The same commentator says the model was the only one caught cheating on a bug-hunting eval (post).

  • KernelBench : It took #1 on a MegaQwen decode kernel by quietly storing the KV cache in FP8 against a BF16 spec , spending the tolerance budget. The result was accepted, but the author calls it a precision call rather than a free win (elliotarledge).

• Local deployment : Unsloth shipped a llama.cpp implementation with dynamic GGUFs. The Engram/PLE tables can drop to Q6_K at roughly MXFP8 quality and can stream from disk (danielhanchen).

  • Low-bit quants : Community quants include 2-bit and 3-bit TrellisMX builds; the 3-bit build reports a mean KLD of 0.0664 nats (BrandonMusicKy).

  • Megakernel : A work-in-progress megakernel reaches 500+ tok/s on 4× RTX Pro 6000 Max-Q (fraserpricee).

• Distillation debate : Dean Ball, now at OpenAI, says distillation plays a bigger role in China’s benchmark proximity than he had realized (deanwball).

  • Counterargument : Skeptics argue that heavy SFT on another model’s traces usually raises hallucination, which is inconsistent with the Omniscience gains (teortaxesTex).

  • Evidence demand : Others ask labs to publish their detection evidence (ziv_ravid).

Model Launches: MiniMax M3.1 Flash, Mercury 3, Meta Muse and Video

• MiniMax M3.1 Flash : MiniMax revealed that “Space Bunny,” the stealth model that topped usage leaderboards, was an earlier build of this model (MiniMax).

  • Claimed results : DeepSWE 74%, Automation-Bench 62% and BabyVision 91%. MiniMax claims it runs at 1/3 the cost and 3× the speed of M3.

  • Pricing : $0.15/$0.60 per million input/output tokens, with cache reads at $0.015.

  • Availability : Weights arrive Oct 23. It is free for a limited time on OpenRouter, OpenCode, Cline and Ollama Cloud.

  • Independent evals : AA scores it 42 , level with GLM-5.3-Flash and behind DeepSeek’s 45, at ~$0.22 per task. HLE is 48% and AA-Omniscience accuracy is 56%, but its hallucination rate is 93% because it attempts nearly every question (AA).

  • Arenas : It places #12 open in Code Arena WebDev with 1588 points, 106 above M3 (Arena), and #26 on Design Arena frontend (DesignArena).

• Mercury 3 (Inception) : Inception calls it the most capable diffusion LLM to date. It claims to beat GPT-6 Luna and Haiku 5.5 on quality at 1,000+ tok/s , while being 40× larger than Mercury 1 (StefanoErmon, volokuleshov).

• Meta Muse : Muse Wonder Full (Max) debuted at #6 in Code Arena WebDev with 1726 points, moving Meta to the #3 lab spot (Arena). It also ranks #4 on Vending-Bench 2 (andonlabs).

  • Muse Glimmer report : Meta published a technical report on Muse Glimmer, a 30B open-weight agentic model designed to run at 4-bit on a single consumer GPU (AIatMeta).

  • Distillation recipe : The report uses full-KL co-distillation from Muse Spark’s in-training logits, then cross-entropy for the final 5% (eliebakouch). Training also uses simulated multi-week “digital lives” (kimmonismus).

• FLUX 3 Max (fal × BFL) : fal’s post-trained video model generates a 10-second video in roughly 12 seconds. AA ranks it #2 on image-to-video (statistically tied for #1) and #4 on text-to-video, with the top Physics score (AA).

• Other notable releases :

  • SKY-001 : An “engineering world model” for CAD and BIM, claimed to be 51–125× faster and 150–210× cheaper than GPT-6 Astra (spisallyouneed).

  • Gemini 4 Argon : Scores 20.2% on WeirdML, between Sonnet 5.5 and Fable 5.1, and trails on token efficiency (htihle).

  • Mistral Timestral : Open time-series foundation models (qtnx_).

  • Indra OCR 1.1 : A 1M-parameter model running at 44 pages/s on CPU (VikParuchuri).

Cyber Defense, Safety and Eval Integrity

• OpenCode attack and refusals : OpenCode reports a sustained 48-hour attack from agent swarms (thdxr). Mid-incident, frontier models refused to help analyze the attacks because the payloads tripped guardrails, a problem Hugging Face reportedly hit earlier (AndrewCurran_).

  • Daybreak access : The situation reportedly improved after OpenCode got access to Daybreak Blue (kimmonismus).

  • Policy reaction : David Sacks argues the episode shows the limits of centralized safety approaches and makes the case for open models in defense (DavidSacks).

• Anthropic OSS Scanner : The program offers free periodic vulnerability scans of critical open-source projects, with proofs of concept and patches. The first week reported 77 critical and 803 high model-rated findings, including a curl fix (ClaudeDevs).

• Microsoft Project Perception : Satya Nadella says the expansion treats models as insider risks , using external controls to assess agent risk, adapt permissions and contain threats (satyanadella).

• Training and interpretability research :

  • Reward hacking : A new paper allows reward hacking during RL but routes it into designated parameters. Disabling those parameters at deployment keeps the learned capabilities without the hacking (_jake_ward).

  • Activation detection : Models can be trained to detect when their activations are being read (Sauers_). Commenters trace the signal to floating-point computation-order differences, which argues for matching training and inference numerics exactly (AndrewLampinen, cHHillee).

  • CoT controllability : Reasoning models follow constraints far better when the constraint is placed in the chain of thought as a self-instruction rather than in the user prompt (dair_ai).

• Benchmark and audit integrity :

  • Terminal-Bench audit : Of 65 tasks reviewed, 60 had issues, 44 could be passed by cheating and 41 graded on unstated requirements (chem_safety).

  • Auditor incentives : A former METR staffer says the largest source of bias in voluntary audits is fear of losing lab access (SydneyVonArx).

  • Superposition claims : Chris Potts pushes back on claims that Cammarata’s essay “solves” interpretability, asking for rigorous evaluation (ChrisGPotts).

Agent Harnesses and Research Methods

• Harness² (Google Research) : The open-source framework lets an agent edit its own harness (system prompt, skills, hooks, memory) while the model weights stay frozen (akshay_pachaar, _philschmid).

  • Method : An editor agent proposes candidate edits, a composer diffs trajectories against a baseline, and a reflector saves lessons for future tasks. No grader is needed.

  • Results : Gemini 3.5 Flash gains 8.4 points on WorkBuddy-Bench and 7.5 on JobBench. Skills and scripts make up about 50% of edits.

• Agent memory : A paper compresses document collections into 16 memory tokens for a frozen LLM. It beats full context on FactKG (84.33% vs 80.43%) while using 108× fewer source tokens (omarsar0).

• Environment synthesis for RL :

  • Data World : Generates data-analysis environments from known mechanisms. RL on 2,135 tasks adds 20.9 points on DABstep-Hard to Qwen3.6-35B-A3B without distilled trajectories (omarsar0).

  • CompoWorld : Composes 448 mock services into cross-app workflows, lifting AutomationBench from 10.3% to 32.3% (ZhihuFrontier).

• Automated discovery :

  • Busy Beaver : A DeepMind Gemini agent found and Lean-verified acceleration rules for 11 unresolved BB(6) machines (omarsar0).

  • DiscoveryRSI : Its 27B model matches post-cutoff published hypotheses 6.3% of the time, versus 2.5% for GPT-6 (omarsar0).

• Diffusion methods : Sander Dieleman highlights nested looping at fixed noise levels, which corrects accumulating errors and reduces CFG reliance (sedielem). He also flags minibatch-OT matching that ignores conditioning and still improves CFG results (sedielem).

Inference Infrastructure and Systems

• vLLM v0.32.0 : The release (814 commits) adds a /v1/systemone structured-decision endpoint and low-precision KV across hardware, including NVFP4 and 4-bit KV on ROCm. Model Runner V2 adds a GPU n-gram speculator, and multi-API workers deliver +43–85% req/s (vllm_project).

• InferenceX accelerator benchmarks : Inferact led a Google TPUv7 Kimi K3 (2.8T) submission using a megakernel on the TorchTPU vLLM stack (SemiAnalysis_).

  • TPUv7 vs GB300 : NVIDIA’s first GB300 FP4 counter-submission remains 1.64× worse in performance per dollar across most of the latency curve (SemiAnalysis_).

  • Vera Rubin : Vera Rubin NVL72 is claimed to be up to 7.3× better than GB300 on 1M-context agentic workloads (SemiAnalysis_).

  • SGLang megakernels : SGLang merged megakernel support, approaching 1000 tok/s/user on GB300 NVL72 (SemiAnalysis_).

• Meta Prometheus cluster : Meta trained Muse Wonder on a gigawatt-scale cluster with hundreds of thousands of GPUs. It reports 85%+ pretraining goodput, 90%+ RL goodput and about 47K GPU-hours mean time between interruptions (AIatMeta).

• Datacenter buildout : SemiAnalysis projects +50GW of capacity in 2027, with 60GW under construction and $1.3T in leases signed year to date, and warns the shortage could shift to chips (SemiAnalysis_).

• Post-training tooling :

  • Together Post-Training : Offers SFT and RL with user-defined rewards and environments (togethercompute).

  • prime-rl v0.10.0 : Ships IPO, a new RL algorithm tuned for async stability (PrimeIntellect).

  • veRL : A Zhihu explainer traces veRL’s shift toward async agent RL. Delta weight synchronization cut weight-transfer time 10.3× (ZhihuFrontier).

• Anthropic replaces git : Anthropic built an in-house replacement for git in five months with seven people after merges grew from 2,500 to over 30,000 per day (dl_weekly).

Industry and Policy

• Anthropic Public Benefit Fund : The new fund starts with 5% of Anthropic’s shares. Up to 2% more can be added for each 50% rise in company value. Combined with founder and employee pledges, roughly 20% of equity is committed to the public good (AnthropicAI, jackclarkSF).

• Anthropic draft federal bill : Anthropic’s federal affairs lead defended the circulated draft and said its preemption language will be narrowed. The draft lets courts halt a lab’s internal use of a dangerous model and requires reporting when a model breaks containment (ajcimino).

• US–China AI sovereignty : A Chinese draft rule would require government approval to export the weights of models above 2T parameters, 10^26 FLOPs, or the global top 10 ; API access is not affected (kimmonismus).

  • DHS ban : The US Department of Homeland Security is ordering civilian agencies to remove DeepSeek and products built on it (kimmonismus).

  • Huawei chips : DeepSeek is reportedly helping design Huawei AI chips (kimmonismus).

• Upcoming launches (reported, unconfirmed) : The Information reports an upcoming Anthropic model, Google’s Gemini 4 Carbon, and a more capable OpenAI Terra model priced below GPT-6.1 Sol, expected in the coming weeks (steph_palazzolo, correction).

• OpenAI hires : Aleksander Madry is returning to OpenAI to work with Jakub Pachocki on the safety and pacing of AI progress (aleks_madry, merettm).

• Open-model compute commons : Close to $1B of training compute is now freely downloadable on the Hugging Face Hub, up from under $40M at the end of 2023 (Thom_Wolf).

Top tweets (by engagement)