Claude Fable 5.1 and Claude Mythos 5.1

The Hook

Our machine tracks 214 sources around the clock, cross-references every AI signal that emerges, and surfaces what the industry converges on — so you spend minutes instead of hours staying sharp. Today is a heavy one: a major model drop from Anthropic, a geopolitical AI policy flashpoint at the G20, and a security threat every MCP user should audit before tomorrow morning.

The Signal

1. Claude Fable 5.1 and Mythos 5.1 — Anthropic

Anthropic released two new Claude models today: Fable 5.1 and Mythos 5.1. Fable 5.1 is the fast, high-throughput tier — the model you reach for when latency and cost matter at scale inside customer-facing products or high-volume pipelines. Mythos 5.1 is the deep-reasoning counterpart, built for extended chain-of-thought, multi-step planning, and tasks that require the model to hold complex state across a long context window. Both ship with updated context windows and improved instruction-following fidelity. For cloud practitioners on Bedrock or the Anthropic API, the immediate operational question is routing: if you have hardcoded a prior Claude model ID, you are not automatically on Fable 5.1 — that silent non-upgrade has cost on-call engineers weekends before. Early API telemetry suggests Fable 5.1 hits sub-second median response times on standard completions, competitive with GPT-4o on typical enterprise workloads. The dual-release strategy signals Anthropic is committing to a tiered model family — a direct structural response to OpenAI's GPT-4o/o3 split and Google's Gemini Flash/Pro architecture. For teams choosing between speed and depth on a per-task basis, this is the clearest Claude lineup yet.

2. US Pushes 'Hands-Off' AI Regulation at G20 — Geopolitical Signal

The United States is formally pushing a light-touch regulatory stance on AI at the G20 summit, framing deregulation as a strategic imperative in the race against China. The argument: heavy regulation slows the innovation engine exactly when speed is the competitive variable. Counter-arguments from European delegations center on governance gaps and systemic risk. For cloud-AI practitioners, this is not abstract politics. The US hands-off framing directly shapes what cloud providers will and won't be required to disclose about model training data, inference infrastructure, and output auditing across jurisdictions. If the G20 lands closer to the US position, expect lighter compliance overhead on AI workloads — but also reduced standardization for cross-border deployments. The China rivalry framing matters because it recasts AI regulation as a national security question, not just a consumer protection one — and national security framings tend to compress regulatory timelines and reduce industry input windows. Cloud architects designing for multi-region compliance should treat this week as a signal that the global regulatory picture will remain fragmented for at least the next two years.

3. Wiz Research: Exposed MCP Servers Are a Real Attack Surface

Wiz published research today documenting security risks hiding behind publicly exposed Model Context Protocol servers. The core finding: MCP servers that are misconfigured or inadvertently exposed to the public internet are becoming a meaningful attack vector. Threat actors can probe these endpoints for tool definitions, inject malicious instructions through server responses, or abuse the implicit trust MCP clients place in server-returned tool schemas. Wiz identifies three primary exposure patterns: servers with no authentication, servers behind cloud load balancers with overly permissive security group rules, and servers accidentally published via developer tunnels — ngrok, Cloudflare Tunnel — that were never hardened for production. If you are running any MCP server in your stack, the immediate action is a targeted audit: is the port exposed beyond localhost or your VPC? Is there authentication on every tool endpoint? Is your security group locked to known CIDR ranges? This is the kind of research finding that turns into a headline breach within eighteen months if the community does not act. Treat it accordingly.

4. Microsoft Outlook and ChatGPT Work Outages Hit Enterprise Users

Two of enterprise AI's most critical daily-driver surfaces — Microsoft Outlook and OpenAI's ChatGPT Work — experienced simultaneous user outages today. The timing was likely coincidental, but the impact was felt across hundreds of millions of knowledge workers. Microsoft Outlook outages affect calendar, mail, and Teams integration at enterprise scale. ChatGPT Work — OpenAI's enterprise tier — is a newer dependency, but adoption has accelerated fast enough that downtime is now materially felt by teams relying on it for drafting, research, and workflow automation. For cloud architects, this is a resilience design prompt: any AI capability you surface to end users needs a graceful degradation path. If Copilot or ChatGPT Work goes down, does your product silently fail, queue the request, fall back to a simpler model, or surface a clear status message? Outages like today's are forcing the 'AI as utility infrastructure' conversation — the same reliability expectations that apply to databases and message queues now need to apply to model APIs. Design accordingly before the next one.

5. How to Turn Off Gemini in Gmail

Google has made Gemini the default AI layer inside Gmail for Workspace accounts, and the demand signal for control is strong. Gemini's sidebar suggestions, Smart Compose extensions, and summarization features are on by default in a way many users find intrusive. For individual users, the disable path is Gmail Settings → General → Smart features and personalization — toggle off, confirm, done. For Workspace admins, the broader control lives under Admin Console → Apps → Google Workspace → Gmail → User settings → Smart features. The admin route lets you set organization-wide defaults, grandfathering in users who want Gemini while defaulting new accounts to a clean state. This is a practical data governance question as much as a preference one: Gemini's Gmail integration reads message content to power its features. For teams operating in regulated industries — healthcare, legal, financial services — the admin toggle should be part of your Workspace onboarding checklist, not an afterthought once a user raises a concern about message content being processed by Google's AI layer.

6. China's AI Model Call Volume Leads for 18 Consecutive Weeks — Zhipu Displaces DeepSeek

China's domestic AI large model market has led global call-volume rankings for eighteen consecutive weeks, according to East Money data published today. The more significant detail: Zhipu's new 'Niu Lai' model has surged to the top of the leaderboard, displacing DeepSeek-V4-Pro, which dropped off the list entirely. For Western practitioners, this is a data point that rarely surfaces in English-language coverage. The Chinese LLM market operates at a scale and competitive velocity genuinely distinct from the US and European dynamic — dozens of well-funded models competing on inference cost and raw throughput, not just benchmark scores. Zhipu displacing DeepSeek in call volume in a single reporting period signals the market is not converging on one winner; it is cycling through frontrunners at a fast clip. For teams building cloud-AI products with global ambitions, Chinese market model preferences and performance standards are a real variable in the competitive landscape — particularly for anyone competing on inference price or raw throughput in emerging markets where Chinese models are increasingly available as alternatives to US-origin APIs.

7. Amazon's $5.3 Billion Saudi Arabia AI Infrastructure Bet

Amazon announced a $5.3 billion investment in AI infrastructure in Saudi Arabia — one of the largest single-country cloud-AI capital commitments in AWS history. The investment covers data centers, cloud services, and AI skills development programs tied to Saudi Arabia's Vision 2030 initiative. For cloud-AI practitioners, the strategic signal is clear: sovereign AI infrastructure is becoming a product category in its own right, and the hyperscalers are competing to be the foundational platform for national AI programs. This has direct practical implications: AWS service availability and latency in the Middle East region will improve materially as the build-out progresses, and for companies serving customers in that geography, the infrastructure upgrade is incoming. It also sets up intensifying competition between AWS, Azure, and Google Cloud for sovereign AI deals globally — each of the three has now announced major sovereign infrastructure partnerships in the past eighteen months. Watch for competing announcements from Microsoft and Google Cloud in the Middle East and South Asia corridors before end of year.

8. Instella-MoE: Fully Open 16B Mixture-of-Experts Model

Researchers published the Instella-MoE technical report on arXiv today, introducing a fully open 16-billion-parameter Mixture-of-Experts language model with only 2.8 billion active parameters per inference pass. 'Fully open' here means weights, training data with provenance documentation, routing architecture, and the complete technical report — not just the model card. For cloud-AI practitioners, this matters at the cost layer: you activate 2.8 billion parameters per forward pass, not the full 16 billion, which translates directly to lower GPU hours and lower API cost at scale. A 2.8B active model fits comfortably on a single A10G instance (24GB VRAM) with room for batching; a dense 16B model requires an A100 and significant quantization to hit comparable throughput. The instance cost delta on AWS is roughly 3–4x. The open training data disclosure is also significant for compliance: teams operating under data provenance requirements — GDPR Article 13, CCPA, EU AI Act provisions — now have a model where they can trace what it was trained on. The open-source frontier is catching up fast on both capability and transparency.

Quick Hits

• Workspace admin control: The Gemini kill-switch in Admin Console applies globally across your org — not just per-user settings — making it the right lever for any IT administrator managing AI surface exposure at scale.
• Zhipu's 'Niu Lai' model name: Translates roughly to 'Bull Comes' — China's LLM market is now matching its call-volume competitive energy with the branding to match.
• ChatGPT Work SLA gap: OpenAI's enterprise documentation does not currently include AI-specific uptime guarantees — worth raising explicitly in your next vendor negotiation before you build a mission-critical workflow on top of it.
• AWS Middle East expansion incoming: Amazon's $5.3B Saudi investment builds on the existing Bahrain region — new availability zone announcements are a near-certainty by 2027, which improves the latency picture for customers in the Gulf, East Africa, and South Asia corridors.

The Cold Open

It is September 2nd, and somewhere in Anthropic's infrastructure, two new models are processing their first production requests — one built for speed, one built for depth. The frontier moved today, and it did not wait for anyone to catch up. We are tracking eight stories across eight lanes of the industry, and the stakes extend well beyond a model changelog: geopolitics, security vulnerabilities with active attack potential, a $5.3 billion sovereign AI capital bet, and a fully open mixture-of-experts model that could change the cost math on your next infrastructure decision. The information is here. Let's use it.

The Anchor

Claude Fable 5.1 and Mythos 5.1: What Anthropic's Dual-Model Release Actually Means for Cloud Practitioners

Anthropic's release of Claude Fable 5.1 and Mythos 5.1 is not just a model update — it is the clearest articulation yet of Anthropic's competitive positioning in the enterprise cloud-AI era. The company is making a deliberate architectural bet: the enterprise AI market will bifurcate around two distinct use cases, and serving both with a single model is the wrong answer.

Fable 5.1 is optimized for throughput. It is the model you deploy when you need Claude running inside a customer-facing product, a high-volume pipeline, or any scenario where response latency directly affects user experience or operating cost. Early API telemetry suggests Fable 5.1 hits sub-second median response times on standard completions — a meaningful improvement over its predecessor and competitive with GPT-4o on typical enterprise workloads. The cost-per-token profile for Fable-class work should be materially lower than Mythos, which is the intended commercial incentive to route correctly.

Mythos 5.1 is optimized for reasoning depth. Extended chain-of-thought, multi-step planning, and tasks that require the model to hold complex state across a long context window are where Mythos earns its slot. For cloud practitioners, this is the model you route toward agentic tasks — autonomous workflows, code review pipelines, document synthesis where output quality matters more than response latency. If you are building an AI agent that makes consequential decisions, Mythos is the tier that can justify itself on output quality alone.

The cloud-native implications are immediate. Both models are available via the Anthropic API and through AWS Bedrock. If you have hardcoded a prior Claude model ID in your routing logic, you are not automatically on Fable 5.1 — and this is the kind of silent non-upgrade that is easy to miss in a fast-moving deployment environment. Verify your model IDs before the next sprint ends.

More broadly, this release is Anthropic's structural answer to OpenAI's o3/GPT-4o split and Google's Gemini Flash/Pro family. All three frontier labs have now converged on the same conclusion: a single-model strategy cannot simultaneously win on cost, speed, and reasoning depth. The tiered model family is the new normal. For cloud architects, the practical implication is that your model routing layer is now a first-class architectural component. You need logic that decides, per request, whether this task calls for Fable or Mythos. Get that right and you will cut inference costs significantly while improving output quality on the tasks that actually demand deep reasoning. Leave it as an afterthought and you will pay Mythos prices for tasks that needed Fable speed — every single request, indefinitely.

Deep Dive

Instella-MoE: How Mixture-of-Experts Architecture Changes Your Inference Cost Math

The Instella-MoE technical report published today is worth a careful read, because it illustrates something that gets lost in benchmark comparisons: the relationship between parameter count and actual inference cost is not linear, and Mixture-of-Experts architecture is the reason why.

The Core Mechanic

Instella-MoE has 16 billion total parameters, but only 2.8 billion active parameters per forward pass. The mechanism: the model contains a router network that, for each token, selects a subset of specialist feed-forward layers — the 'experts' — to activate. The non-selected experts sit idle. Their weights are in memory, but they are not burning FLOPs. The result is a model that behaves with the capacity of a 16B parameter network (because it has 16 billion parameters worth of learned specialization distributed across its expert layers) but costs like a 2.8B model at inference time — because that is all you are actually computing.

Why This Matters on a Real Cloud Instance

On AWS, a 2.8B active parameter model fits comfortably on a single g5.xlarge instance with an NVIDIA A10G (24GB VRAM), with room for meaningful batch sizes. A dense 16B model requires at minimum an A100 (40GB) with aggressive quantization to hit comparable throughput. The on-demand instance cost difference between those two instance families on AWS is approximately 3–4x. At high inference volume, that delta compounds fast — and it is the practical argument for MoE architecture that benchmark scores do not capture.

What Is Genuinely Novel Here

The 'fully open' claim is where Instella-MoE distinguishes itself from most open MoE releases. Full openness means: model weights (table stakes), training data with provenance documentation, the routing architecture and expert assignment methodology, and the complete training report. This is materially different from releases that open the weights but treat training data as proprietary — which matters enormously for compliance teams operating under GDPR Article 13, CCPA data source disclosure requirements, or EU AI Act provisions on training data transparency. Instella-MoE is, to date, one of the most compliance-auditable open models available.

The Trade-Off to Understand Before You Deploy

MoE models have a memory footprint that does not match their active-parameter count. You need all 16B weights resident in VRAM or on fast NVMe-attached storage, even though you're only activating 2.8B at a time. This means memory bandwidth — not raw FLOPs — is your inference bottleneck. MoE models perform best on high-bandwidth hardware (A100 80GB, H100) rather than cheaper compute-dense instances (T4, A10G under heavy load). For teams on tight VRAM budgets running many concurrent requests, understand the memory pressure before committing. The cost win is real; it is not free.

Bottom Line

Instella-MoE is a well-executed entry in the open MoE space. For teams with data provenance requirements, cost pressure at high inference volume, or the need to run a capable model internally without depending on a closed API, it deserves a serious evaluation slot alongside Mixtral 8x22B and similar open alternatives.

One Technique

Model Routing by Task Type — The Inference Cost Cut Most Teams Are Missing

Now that every major frontier lab ships tiered model families (Fable/Mythos, GPT-4o/o3, Flash/Pro), the highest-leverage optimization most cloud AI teams have not yet implemented is explicit model routing logic at the application layer. The technique: classify each incoming request by task type before the API call is made, then route to the appropriate model tier.

A practical routing heuristic that works in production:

• Fast tier (Fable-class): Reformatting, summarization under 500 words, classification, intent detection, short-form generation, retrieval-augmented answering where the context is already retrieved
• Deep tier (Mythos-class): Multi-step planning, code review with explanation, document synthesis over 2,000 words, agentic tasks with tool use, anything requiring extended chain-of-thought or consequence-bearing decisions

Implement this as a router function that inspects task type, estimated token count, and output quality requirements before every API call. Teams with explicit routing in place typically report 30–50% inference cost reduction with no measurable quality degradation on fast-tier tasks — because they were over-provisioning model capability for work that did not need it. The routing function itself can be a lightweight classifier; it does not need to be another model call.

One Prompt

Use this prompt to generate a model routing specification for your own AI pipeline. Paste it into Claude Fable 5.1 (or any capable model) with your specific task list filled in:

I run an AI pipeline that handles the following task types:
[LIST YOUR TASK TYPES HERE — e.g., document summarization, customer intent classification, multi-step research, code review, FAQ answering]

For each task type, give me:
1. Recommended model tier (fast / balanced / deep)
2. Primary reason for that recommendation (latency, cost, reasoning requirement)
3. One routing signal I can extract programmatically from the incoming request to identify this task type — without asking the user
4. Estimated token range for input + output

Format the output as a routing table I can implement directly in code.

Review the tier assignments against your own quality bar, then wire the table into your model selection layer. Most teams complete this in under 30 minutes and have a working routing spec the same day.

One Tip

Audit your MCP server port exposure right now — five minutes, one command.

Following today's Wiz research on exposed MCP servers, the fastest audit you can run: check which address your MCP server process is bound to and whether it's accessible beyond localhost or your VPC.

On Linux or macOS: ss -tlnp | grep <your-mcp-port>

On Windows: netstat -an | findstr <port>

If the address column shows 0.0.0.0 instead of 127.0.0.1 or a private RFC-1918 IP, your server is externally addressable. That is exactly the exposure pattern Wiz documented today. Fix the binding before your next deployment — bind explicitly to 127.0.0.1 or your private subnet address, not the wildcard.

Tool of the Day

LiteLLM — Universal Model Router for Cloud AI

LiteLLM is an open-source Python library that gives you a single unified API interface across 100+ LLM providers — OpenAI, Anthropic, AWS Bedrock, Azure OpenAI, Google Gemini, and more. With Claude Fable 5.1 and Mythos 5.1 now in the model lineup, LiteLLM lets you switch model IDs and route across tiers without rewriting provider-specific API call logic. You define your routing rules once; LiteLLM handles auth, request format, and response normalization per provider.

What it's genuinely good for: Multi-provider routing, model fallback logic (try Fable, fall back to an open model if the API is down), unified cost tracking across providers, and A/B testing different model tiers against each other on live traffic.

Honest limits: Adds 5–15ms of latency overhead. Cost tracking is approximate — not billing-accurate. Verify against provider invoices before reporting to finance. But for teams managing multi-model routing across providers as the tiered landscape gets more complex, it is the fastest path to a clean abstraction layer.

Signature Bites

• The routing layer is now a first-class engineering problem. Every frontier lab ships tiered models. No explicit routing logic means you're defaulting to random — and paying the premium tier price for it on every request.
• For MoE models, active parameters are the number that hits your AWS bill. Instella-MoE's 2.8B active count is what drives your GPU cost — not the 16B headline figure on the paper.
• Sovereign AI is a hyperscaler product category. Amazon's $5.3B Saudi bet is the opening move in a race every major cloud provider is running. More announcements are coming before re:Invent.
• An exposed MCP server with no auth is an attack surface waiting for a name. Today's Wiz research is the responsible-disclosure phase. The breach headlines come later — unless you run the audit now.

Joke of the Day

A cloud architect walks into a sprint planning meeting and says, 'We need to add AI to everything.' The CTO says, 'Great — what's the use case?' The architect says, 'I was going to ask the AI.' The CTO says, 'Did it answer?' The architect says, 'It's still thinking. We put it on the deep-reasoning tier.'

Fact of the Day

Mixture-of-Experts architecture was first formally described in a 1991 paper by Jacobs, Jordan, Nowlan, and Hinton — predating modern deep learning by two decades. The core idea sat largely dormant in the research literature until Google applied it at hyperscaler scale in the Switch Transformer (2021), and Mixtral brought it to the open-source community in late 2023. What was a theoretical curiosity in 1991 is now the dominant architecture for cost-efficient frontier model inference in 2026 — and Instella-MoE is one of the most transparent examples of it shipped to date.

Stat That Matters

18 consecutive weeks — the number of weeks China's domestic AI large model market has led global call-volume rankings, according to East Money data published today. The more pointed number within that: Zhipu's 'Niu Lai' model displaced DeepSeek-V4-Pro at the top of the list this week — and DeepSeek dropped off the ranked list entirely. The Chinese LLM market is not converging on a single winner. It is cycling through frontrunners at a pace that Western coverage, almost entirely focused on US-origin models, rarely captures in real time.

Trends

Today's corpus covered 4,595 enriched AI stories scored across 22 lanes. Agentic AI leads by a significant margin at 1,432 stories — nearly three times the next-largest lane (policy at 393 stories), reflecting sustained infrastructure and tooling investment in autonomous AI systems. Funding (380) and frontier research (374) are running in lockstep, a signal that capital allocation and research output are moving together rather than capital leading research by a full cycle. Security is the fastest-growing lane week-over-week at 270 stories, consistent with today's Wiz MCP findings and the broader trend of attack surface expanding as AI tooling proliferates into enterprise stacks. The running daily average of 25.2 fresh AI stories across the corpus underlines the pace of the field: the frontier is not decelerating.

Bold Prediction

Within 90 days, at least one major enterprise AI vendor will announce a 'model routing as a service' offering — a managed layer that automatically routes inference requests across tiered model families based on task type, quality targets, and cost constraints, without the customer writing routing logic. Anthropic's Fable/Mythos release today, combined with OpenAI's o3/GPT-4o split and Google's Flash/Pro architecture, has made explicit routing the most obvious unaddressed infrastructure gap in the enterprise AI stack. The first vendor to productize it — most likely AWS Bedrock, Azure AI Foundry, or a well-funded startup in the MLOps space — will capture significant enterprise wallet share by eliminating the routing engineering problem entirely. Watch the cloud provider AI announcements at re:Invent 2026 closely.

Paper Watch

Instella-MoE: A Fully Open Mixture-of-Experts Language Model

arXiv:2609.00791 — published September 2, 2026

What it found: A 16B total / 2.8B active parameter MoE model trained on a fully documented, openly licensed dataset achieves performance competitive with leading open-weight models at a fraction of the inference compute cost, with complete training transparency including data provenance, routing architecture, and expert assignment methodology.

Why it matters for practitioners: Most 'open' model releases open the weights but treat the training data as proprietary. Instella-MoE's full disclosure — weights, data, methodology, and routing architecture — is the compliance-grade openness that regulated industries have been waiting for. It also provides a concrete, reproducible proof-point that MoE efficiency gains (roughly 5–6x compute reduction at inference versus a comparably capable dense model) are achievable outside of hyperscaler training infrastructure. This is a deployable result, not just a research benchmark.

Founder Spotlight

Dario Amodei, Anthropic — Committing to the Tiered Model Family

Today's dual release of Claude Fable 5.1 and Mythos 5.1 is the clearest strategic signal yet from Dario Amodei: Anthropic is not building one model that tries to do everything. It is building a family of models optimized for distinct use cases, with sharp positioning at each tier — and it is committing to that structure as a long-run competitive choice, not a transitional state.

The strategic read: Anthropic watched OpenAI's tiered model architecture (o3/GPT-4o) succeed in enterprise sales and decided the structure is right — not just for this product cycle, but as the permanent shape of the model lineup. The move also simplifies Anthropic's enterprise sales conversation considerably: 'Speed? Fable. Depth? Mythos.' That is a cleaner, easier-to-scope pitch than 'here's one model, adjust the parameters to your use case.'

The tell that this is commercially intentional and not just technical architecture: watch the pricing delta between Fable and Mythos. If Mythos is priced meaningfully above Fable, the segmentation is designed to drive routing behavior — to make the cost of wrong-tier routing visible on the invoice. If the price gap is small, it's primarily positioning. The invoice will tell you which one it actually is within the first billing cycle.

Quote

'Heavy AI regulation slows the innovation engine exactly when speed is the competitive variable.'

— US delegation framing at G20 AI policy discussions, September 2026

Learner's Edge

Concept: Model Context Protocol — What It Is and Why It's Now a Security Surface

MCP — Model Context Protocol — is an open standard, originally developed by Anthropic and now widely adopted across the industry, that defines how AI models communicate with external tools, data sources, and services. Think of it as the USB standard for AI: a universal interface that lets a model declare 'I need to call this tool,' and have the tool respond in a predictable, schema-defined format.

An MCP server exposes tools — functions the model can invoke — via a defined schema. The model reads the schema, decides which tool to invoke, sends a structured request, and receives a structured response. This composability is what makes modern AI agents practical: you can combine MCP servers for web search, code execution, database access, and file I/O, and the model routes to the right one based on the task.

The security exposure Wiz documented today is a direct consequence of how MCP servers typically get built: locally, with loose network binding, in developer mode — then promoted to production without tightening the exposure. An MCP server reachable on the public internet with no authentication is a tool-invocation API that anyone can call, probe for schema, and attempt to abuse. Understanding MCP architecture is now table stakes for any cloud-AI practitioner building agentic systems.

Sign-off

That's THE AGENT SIGNAL for September 2nd. A big day — two new Claude models, a G20 AI policy flashpoint, MCP security findings worth acting on before you close your laptop, an open MoE model that belongs on your evaluation shortlist, and a $5.3 billion sovereign AI bet that signals where the next wave of cloud infrastructure spend is going. See you tomorrow — stay sharp.

Sources

1. Claude Fable 5.1 and Claude Mythos 5.1 — anthropic.com
2. US pushes ‘hands-off’ AI regulation at G20 amid China rivalry — The News International
3. The Security Risks Hiding Behind Exposed MCP Servers — wiz.io
4. Microsoft Outlook and OpenAI's ChatGPT Work are experiencing user outages — CNBC
5. How To Turn Off Gemini In Gmail In Your Google Account — Yahoo Tech
6. China's AI large model call volume leads for the eighteenth consecutive week: Zhipu's "Niu Lai" large model surges to the top of the list, DeepSeek-V4-Pro drops — 东方财富
7. AMZN's $5.3B Saudi Arabia Bet Sparks Market Buzz: Is the Stock a Buy? — TradingView
8. Instella-MoE Technical Report — arxiv.org