VOL.2026-10-21 · 33 STORIES · SKILLHUB DAILY
SkillHub.news Daily AI
Oct 21, 2026 · #39 · updated 08:00 daily · sources 4
01MODEL RELEASES4 stories
MiniMax M3.1 Flash: nearly 3x faster and cost reduced by 70%, over 200T Token processed during the anonymous period
MiniMaxInfoQ（网易转载）+ 36氪Confidence MidProgress · Day 2
During the anonymous period, it processed over 200T Token under the Space Bunny identity; the company says Agentic tasks/programming/Computer Use saw generational improvement, with the Pro version teased as underway; this is today's update to the 10-20 reported item (cost reduction and anonymous traffic scale are given for the first time).
Take · Running real traffic under an anonymous identity before claiming it is like using the online distribution for a free holdout validation, cheaper than a benchmark; but the generational improvement claim has no public baseline and is unquantified, so it should be discounted.
Shanghai Memory Tensor open-sources MemT-1.0: a memory-native large model, with 288B Flash and 36B Nano tiers
上海记忆张量新民晚报（腾讯新闻）Confidence MidOpen-source and testable.
Historical information and task experience enter model computation directly without retransmitting the original text; the company says the large model proves the upper bound and the small model proves deployability.
Take · Replacing 'retransmit history every turn' with 'write once into memory' is a second cost-reduction path beyond long context—but there is no third-party reproduction yet; first establish interfaces and acceptance criteria before discussing replacement.
Kimi K3 now on official website: 2.8T-parameter MoE, 104B activated, 1M context, open weights
Kimikimi.com 官网抓取 + arXiv 2607.24653Confidence MidPending verification
Activation ratio about 3.7%, native vision, focused on Agentic programming / Swarm clusters; official site self-evaluation still lags Claude Fable 5 and GPT-5.6 Sol, but beats other open-source and closed-source models; no first-release date is marked, logged as 'available today' rather than 'released today'.
Take · In a market where everyone self-rates as SOTA, explicitly stating it lags uses falsifiable self-deprecation to buy credibility—it is not seizing the performance slot, but the 'most honest open source' position.
Xiaomi MiMo-V2.6-Pro goes live: 1M context, focused on long-horizon workflows
小米快科技（网易转载）Confidence Mid
Supports 1M context, now live on WorkBuddy, aimed at long-horizon workflow scenarios.
Take · 1M context and memory roadmap debut on the same day; which has lower per-task cost will determine the next round of Agent selection divergence.
02PRODUCT5 stories
OpenAIIT之家 + TechCrunch + 官方博客Confidence High
Team and technology go into Rosalind/Rosalind Workbench; the latter has already landed clients including Moderna and Novo Nordisk.
Take · The first batch of real examples of model companies vertically integrating upstream into industry software: what's sold is no longer API calls but industry workbenches.
Original ↗Google's in-car Gemini adds "Personal Intelligence": can access Gmail, Calendar, Maps and search history
谷歌IT之家Confidence Mid
Android Auto first, requiring active driver authorization and can be turned off at any time.
Take · The permissions narrative shifts from "can it do it" to "dare you turn it on"—in-car is the least error-tolerant authorization scenario; getting it working first sets the rules for all platforms.
DatabricksDatabricks 官方博客 + 官方视频台账Confidence HighProgress · Day 2
CustomerLake is an enterprise-grade Agentic CDP (already live at leading enterprises); Agent Memory and MemCurrent are production-grade persistent memory; this item is the same-day increment beyond the 10-20 Agent Memory item (CustomerLake Beta and MemCurrent).
Take · Memory breaks away from 'prompt engineering' into a standalone module: define the interface first, then choose the implementation; whoever defines the interface collects the rent.
Original ↗ElevenLabs官方·ElevenLabs（YouTube）Confidence Mid
Posted on the official channel on 10-21, positioned as a simpler way to build natural voice Agents.
Take · Real-time voice integration costs keep falling; competition shifts from 'does the voice sound alike' to 'can it plug into business workflows'.
Original ↗Musk officially announces Hagglebot negotiation agent: to take over enterprise procurement negotiations
马斯克 / xAI猎云网（网易转载）Narrative-level · not evidence
Autonomous price comparison, multi-round negotiations, and term finalization; claims efficiency gains of over 5% and up to $6 trillion in annual global savings.
Take · $6 trillion ≈ 7% of global GDP; no single-point agent can account for this magnitude—treat as narrative, not evidence; the named scenario of “procurement negotiation” is what warrants attention.
03INDUSTRY6 stories
OpenAI plans to raise $30 billion: valuation may exceed $1.4 trillion, aiming to lock in commitments before Anthropic’s IPO
OpenAI界面新闻（36氪 8点1氪引）Confidence Mid
SoftBank, Abu Dhabi’s MGX, and existing shareholders are in talks to participate; the company says post-GPT-6 Astra annualized revenue has reached $50 billion and is eyeing $70 billion by year-end.
Take · The real move is not the valuation level but locking in commitments before Anthropic’s expected mid-November IPO—once the reference point is anchored by IPO pricing, subsequent financing will have to reprice around it.
Nscale in talks with Anthropic on $45 billion data center: compute costs pinned into multi-year rigidity
Nscale × Anthropic界面新闻（新浪转载）Confidence Mid
Scale is about $45 billion, roughly ≈90% of a leading model company's revenue over the past year
Take · Revenue is elastic, long-term contracts are rigid; a mismatch between them is tail risk—AI infrastructure credit risk is shifting from the equity side to the contract side
Meituan DeepSkill catalogs 372,000 skills that have 'actually run successfully'
美团搜狐网Confidence MidSingle-source pending verification
Officials say it has been validated through one million calls, and it is dubbed the 'Dianping' of the agent era; the selling point is not the number of skills but proof of having 'run successfully'
Take · Without reuse rates, the 372,000 are just a massive plugin repository; with reuse rates, they are priceable, depreciable skill assets
DeepSeek search billed per request: 5 yuan/1000 successful requests; server-side calls free before 11-01
DeepSeek36氪Confidence Mid
Search and Fetch are open and priced at ≈0.005 yuan/request; server-side calls within Harness/Claude Code/Codex are free until 11-01 00:00.
Take · Search now has a public price tag for the first time: agentic loops multiply retrieval counts; if cost models omit this line item, call volume will eat through gross margins.
US DHS orders removal of DeepSeek from civilian government systems: order announced 10-19, effective since end-July
美国 DHSBBC 中文Confidence Mid
Covers civilian agencies including Treasury, State Department, and Justice Department; agencies under the Department of Defense are not within the restriction.
Take · Delayed disclosure is itself a signal: compliance risk has moved from "anticipated" to "established fact"; overseas-expansion model selection and procurement contracts need repricing.
GTC Berlin keynote tonight 17:00—19:00: the main theme shifts from selling GPUs to selling 'AI factories'
英伟达NVIDIA 官网Confidence HighProgress · Day 2
Third GTC of 2026; this is the same-day update to the 10-20 item 'GTC Berlin opens today' (agenda timing and main-theme framing).
Take · Only check whether there are hard numbers that can be converted into orders/backlog; if there is only narrative and no numbers, treat it as a valuation story.
04RESEARCH2 stories
Both K3 and MemT-1.0 lack third-party reproduction: none of today's leaderboard-type materials provide recomputable figures
评测口径深度版·榜单维度Definitional caveat
Today's two most closely watched technical releases (Kimi K3, MemT-1.0) are sourced only from official websites or a single media outlet, with no independent evaluation or third-party reproduction.
Take · A continuation of benchmark re-basing: cited scores must carry version and source; vendor self-reported claims can only indicate direction, not conclusions.
Material freshness ledger: h10=1, verdict=WEAK; only 4–5 of 19 items are same-day entries
素材自检11 篇元素情报文件头 + 总协调情报Confidence HighMethodology self-check
Still no improvement after 11 intelligence streams + 4 supplementary items; the root cause is that the 05:30 collection time is earlier than the Chinese-language media publishing peak, not “insufficient item count.”
Take · Recommend moving the time to 08:30 and keeping the 05:30 snapshot for difference-set comparison—otherwise “queryable that day” will be mistakenly recorded as “published that day.”
05INSIGHTS6 stories
Pricing-unit reset day: money locked into long-term contracts, intelligence per use, credentials per call
前瞻深度版研判Confidence Mid
OpenAI anchors its valuation at 1.4 trillion with USD 30 billion in financing, Nscale pins compute costs with a USD 45 billion long-term contract, MiniMax cuts inference prices by 70%, DeepSeek puts price tags on retrieval, and Meituan prices skills with “run-through records”—three price lines were recalibrated the same day.
Take · This is not a discount war but a phase shift in pricing language; those who change the quote structure first will secure the next round of contracts first.
28x or 20x: two readings of a 1.4 trillion valuation
复算界面新闻（自算）Confidence MidRecalculation
1.4 trillion ÷ 50 billion annualized = 28x PS; ÷ 70 billion at year-end = 20x.
Take · 20x is not expensive in software, provided that '70 billion can be realized and gross margin does not collapse'—revenue is on the company's investor-facing basis and unaudited.
Cost rigidity vs. price-decline slope: depreciation period × annual unit-price decline, to see what share of contracts turn into floating losses
压力测试深度版研判Confidence Mid
In the same week, M3.1 Flash unit price was cut 70%, while compute costs are locked in by annual amortization.
Take · If the price-decline slope is faster than the amortization curve, long-term contracts get trapped at high levels—this is the first question the compute chain should calculate for 2027–2028.
372,000 lacks a denominator: any 'verifiable' leaderboard header must list two numbers side by side.
商机搜狐网 + 惊蛰／坎流／总协调评论Confidence Mid
The platform gives 'successful runs + millions of calls' but not 'how many times the same skill is reused by two or more different tasks.'
Take · Successful-run counts can be gamed (like view counts); leaderboards should look at 'monthly active skills / total skills' and 'share of calls by the top 1% of skills'; a leaderboard that shows only the former will be useless in three months.
Agent cost ledger and pricing overhaul: before 11-01 is a free dependency observation window.
动作坎流、巽风情报Confidence Mid
Starting today, instrument two metrics—retrievals per task and token count per task—and write 'retrievals per single task' into the cost model and contract terms.
Take · Frame the price tag in user language as 'about 3 fen per research session'; it will be too late to backfill data after billing takes effect on 11-01.
Skill QA + delisting mechanism: the opportunity is not in building another skill marketplace, but in occupying the verification node.
商机炎离、厚坤情报Confidence Mid
Deliverables: three tables—reuse rate / task failure rate / marginal contribution per skill; eliminated if not fully run quarterly.
Take · Occupying the validation node means occupying the only path every skill must pass; waiting for the platform to set its own standard closes the window (about 3 months).
Ecosystem Pulseself-updating
News600Skills72OpenHub44Reddit252026-09-02 → 2026-09-10 Star growth：obra/superpowers +90,630★ · mattpocock/skills +82,178★ · multica-ai/andrej-karpathy-skills +67,613★ · anthropics/skills +56,067★ · Shubhamsaboo/awesome-llm-apps +43,732★
✦Action Items3
1. Agent per-task cost ledgerAdd three columns to running Agent products: retrievals per task, tokens per task, and gross margin per task; translate the price tag into user language with “about 3 cents per research run.” Path: free cost checkup form for lead gen → subscription monthly ledger → pricing system overhaul (including retrieval-count clauses in contracts).
window Starting now, a 10-day free window to observe dependency before DeepSeek billing takes effect on 11-01 2. Skill reuse rate audit reportConduct third-party spot checks on DeepSkill-type skill libraries: three tables for reuse rate, task failure rate, and per-skill marginal contribution; publish a two-column leaderboard of 'number of successful runs + secondary use rate'. Path: free publication of first-round spot checks to set the benchmark → paid vendor certification and quarterly re-checks → delisting mechanism consulting.
window Within 3 months (platforms have not yet published secondary use rates; a standards vacuum) 3. Memory asset migration solutionProvide private-domain clients with 'exportable memory + cross-model redundancy': a three-part package of memory checkups, contamination rollback, and model-switch rehearsals, delivering an auditable memory inventory. Path: free memory checkup → memory repository hosting (billed per entry) → model-switch drill report.
window 6–12 months (MemT-1.0 has no third-party reproduction yet; first define interfaces and acceptance standards, and do not weld customer memory to a single vendor's format)