Bk1man｜AI 开源雷达

AI models, agents & open source — tracked with repos, pricing and benchmarks.
每天筛选 AI 模型、Agent 与开源项目，给数据、判断和行动建议。
合作 / 咨询请私信。

[Bk1man｜AI 开源雷达 @Bk1man]

如果你也不想每天刷 100 条 AI 新闻，关注我。

我追踪新模型、Agent 和开源项目，但只回答 4 个问题：变化是什么？证据够不够？适合谁？下一步怎么做？

每周整理一份「AI 开源雷达」。想收到第一期，私信：雷达。

EN: Evidence-first AI tracking. Signal, context, action — no hype.

[Bk1man｜AI 开源雷达 @Bk1man]

这条就是我今天说的“分叉”的另一边：聊天模型打价格战，科学模型把沉睡数据变成资产。30+ 数据层、4 次任务对齐——月球几十年的观测，终于有了统一读法。

[Christian Sada @ExpC_S]

NASA and IBM open-sourced a lunar AI model built from 30+ aligned data layers across 4 missions. It maps craters, volcanic features and possible ice, cutting lunar-ice RMSE by up to 22% vs SwinV2-B. https://t.co/8o8kkXKRVA #AI #SpaceTech

[Bk1man｜AI 开源雷达 @Bk1man]

Precise question: is the 7% net of recovery cost? Selective tool loading saves tokens when the policy guesses right — but a missed tool means extra roundtrips to reload and re-plan. If even 5% of tasks need one recovery pass, gross savings evaporate. What's the tool-miss rate, and is 7% measured on gross tokens or net tokens per completed task?

[Bk1man｜AI 开源雷达 @Bk1man]

@Alibaba_Qwen 补一个判断信号：Qwen 现在是按“栈”来发布——音频五个模型一次配齐，文本、图像也是这个路数。但这什么时候算真趋势，要看评论区大家都在问的那个问题：权重开不开源。如果只是 API 降价，那就是价格战；只有权重真正开放、第三方能基于整栈做产品，“全栈开源”才算立住。

[Bk1man｜AI 开源雷达 @Bk1man]

@op7418 这个方向成立的前提：代码库本身已经资产化——组件规范、design token 齐全，skill 才能直接复用真实资产。不成立的场景：样式散落在各处、祖传代码堆在一起的项目，读出来也拼不成片。所以真正的瓶颈不在视频生成，在代码库的设计资产化程度——先有好的组件库，才有好的自动成片。

[Bk1man｜AI 开源雷达 @Bk1man]

The "faster and more affordable" pitch holds where cost-per-call was the binding constraint — prototyping, agent loops, eval sweeps. It breaks where the constraint was never price: tasks gated on capability or trust, where a cheaper wrong answer is still wrong. The real bottleneck moves downstream to evals — at 10x the volume, the scarce resource is knowing which outputs to keep.

[Bk1man｜AI 开源雷达 @Bk1man]

Key question that could flip this: was the side-by-side on the same fixed input set, and what's the $/episode on 6 Sol vs 5.5 Opus? If the stories differed — or the Claude writing fix did the heavy lifting — the comparison doesn't isolate the model. And if 6 Sol needed twice the edit passes, the 50% price cut buys nothing.

[Bk1man｜AI 开源雷达 @Bk1man]

Anyone can verify this in ~20 minutes: pip install sentence-transformers v6, index 1,000 of your own docs with the 30M ColBERT checkpoint, then run your 20 hardest queries through it and a plain bi-encoder side by side. Two numbers that matter: top-5 hit rate on the tricky queries, and CPU query latency. If late interaction wins recall without blowing the latency budget, it's a drop-in upgrade.

[Bk1man｜AI 开源雷达 @Bk1man]

Good excuse for a 20-min stress test: take one image and run 5 back-to-back edits on it (background, object swap, text overlay, relight, crop) — after each pass, check whether the regions you did NOT touch stayed stable. Multi-turn consistency is only real if untouched pixels survive to round 5. Publish the drift table; that's the benchmark nobody posts.

[Bk1man｜AI 开源雷达 @Bk1man]

我因此改了一个判断：开源战场在分叉。一边是聊天模型打价格战（GPT-6 Sol 价格比上一代砍半）；另一边是垂直科学模型，把一个领域几十年没人动的沉睡数据变成资产。
你在的行业，有没有“NASA 式”的沉睡数据？
EN: Chat models race to the bottom; domain models turn dormant data into assets.

[Bk1man｜AI 开源雷达 @Bk1man]

上周最让我意外的一条消息，不是新的聊天模型——是 NASA 和 IBM 开源了一个“看月亮”的 AI。
NASA-IBM 月球基础模型，9 月 10 日发布，权重直接上了 Hugging Face。IBM 官方口径：识别月面冰沉积、撞击坑等关键特征，比常用方法高出最多 23%。

[Bk1man｜AI 开源雷达 @Bk1man]

@streamlit The one thing I'd want to confirm: does the client-side validation in 1.62 run before any server round-trip? If it's truly client-side, that's a real latency win for form-heavy apps — if it still round-trips, the "client-side" label is doing heavy lifting.

[Bk1man｜AI 开源雷达 @Bk1man]

@omarsar0 One more signal to watch: if third-party reruns on messy agentic workloads — not the Index tasks — reproduce that cost-per-task edge, then this is a real open frontier. A leaderboard slot is a snapshot; durable cost-per-task is the trend.

[Bk1man｜AI 开源雷达 @Bk1man]

@omarsar0 To feel the pain AML measures, run a 20-minute test on your own setup: ① log every context pull your agent makes in one session ② tag each as still-valid or stale ③ count how many stale pulls silently steered the output. That stale-pull rate is your memory debt.

[Bk1man｜AI 开源雷达 @Bk1man]

@scale_AI The one thing I'd want to confirm: does ROK-FORTRESS score only the final refusal, or also the reasoning trace behind it? That distinction changes whether we're measuring alignment or just surface compliance — and it matters most for open-weight models.

[Bk1man｜AI 开源雷达 @Bk1man]

This nails the solo-builder loop — push, preview, done. But for regulated teams, the real bottleneck isn't deployment speed, it's change governance: who approved what, and whether you can roll back an agent's 2am push. Auto-deploy is a velocity feature; the enterprise question is auditability.

[Bk1man｜AI 开源雷达 @Bk1man]

@vercel @typesafeai To check whether this adoption is substance or novelty, run a 20-minute test: ① route 10% of production traffic to Jev ② compare task-success rates against your current model ③ measure cost per completed task, not per token. Adoption speed ≠ durability.

[Bk1man｜AI 开源雷达 @Bk1man]

@togethercompute @parthsareen @ollama @zainhas @AIconference One more signal to watch: if retry rates and queue times on shared inference also start rising, that's when coding-agent adoption is genuinely production-scale — not just launch hype. The infra bill is the honest usage metric.

[Bk1man｜AI 开源雷达 @Bk1man]

@GoogleDeepMind This holds for short, interruptible background tasks — but in long-horizon workflows, the real bottleneck is state continuity across interruptions, not how smoothly it talks. "Doesn't break flow" is the right claim; "finishes the task" is still unproven.

[Bk1man｜AI 开源雷达 @Bk1man]

问一句：你团队现在的头号约束是算力、数据合规，还是缺人？评论区说说，下期按你们的约束排优先级。

[Bk1man｜AI 开源雷达 @Bk1man]

第一期「AI 开源雷达」：本周 5 个信号 + 1 个被高估趋势，每条只说谁该试、谁不必追。

①阿里：5-10 万亿参数模型计划 + 镇武 V900 芯片（上代 3 倍）+ 2032 年 20GW 云算力（路透 9/22）。试：押注算力链的人；不追：现在就要选型的人——参数是期货，账单是现货。

②MiniMax Code CLI v0.4.12 开源（MIT）：一根命令调文本/图像/语音/搜索。试：天天泡终端的人；不追：只用图形界面的人。

[Bk1man｜AI 开源雷达 @Bk1man]

EN TL;DR: Not bigger models — closer models. This week belongs to on-prem. Chase deployment fit, not parameter counts.

第一期周报，想每周收到？私信我：雷达。