Mistral Large 4 Open-Sources a Trillion Parameters, Microsoft Offers Voluntary Severance to 60,000 Employees

Trillion-Parameter Open Source: Mistral Large 4 Arrives

French company Mistral AI has officially released the weights of Mistral Large 4, a natively multimodal model with one trillion total parameters. Training was completed in a European sovereign data center using 3,800 NVIDIA Grace Blackwell GPUs. According to the Artificial Analysis Coding Agent Index, the model scores 48, placing it in the top tier for coding agents.

On pricing, Mistral Large 4 costs $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens for online use. The license is a modified MIT, permitting free use for personal, academic, and conventional commercial scenarios, but cloud providers offering model-as-a-service (MaaS) resale must obtain a separate commercial license. This "open source but restricted resale" approach is becoming a common strategy for leading model companies to balance ecosystem and commercial interests.

A trillion-parameter open-weight model is a landmark event for the entire open-source ecosystem. It not only means developers can deploy near-frontier models locally or in private environments, but also that the barrier to application innovation around multimodal capabilities will be further lowered.

Why MoE Architecture Has Become the Mainstream Choice for Trillion-Parameter Models

A trillion total parameters does not mean all parameters are used in every inference. Current large open-weight models generally adopt a Mixture of Experts (MoE) architecture: the model consists of multiple "expert" subnetworks, and each token is routed to only a few of them, so the number of activated parameters per forward pass is far smaller than the total parameter count. This sparse activation design lets models stack extremely large parameter capacity to improve knowledge breadth and expressiveness while keeping inference compute within an acceptable range. The trade-off is that memory footprint still scales with total parameters, requiring a higher memory budget for deployment, and training routing strategies and load balancing becomes significantly harder. Understanding this explains why "trillion parameters" and "locally deployable" are not contradictory.

SanDisk Earnings: AI Inference Is Rewriting Storage Demand

On its fiscal Q1 2027 earnings call, SanDisk raised a notable observation: AI inference is reshaping the supply-demand landscape of the flash memory market, and customers are underestimating the NAND capacity required to run AI infrastructure. Management revealed that the order pipeline for a new business model extends to 2030, and roughly 50 percentage points of bit shipments in fiscal 2027 have already been locked in under NBM agreements.

SanDisk reported quarterly revenue of $11.03 billion and non-GAAP earnings per share of $49.51, both above market expectations, and announced an eight-for-one stock split. The market is exploring replacing some high-cost solutions with SLC NAND to help systems retain broader context and reduce redundant computation.

Behind these numbers is a structural shift: as AI workloads move from training to inference, they place different demands on storage capacity, bandwidth, and endurance. Inference scenarios require frequent reading and writing of context, and traditional storage designs oriented solely around capacity may not be optimal. The SLC NAND substitution approach SanDisk mentioned is precisely a technical adjustment targeting this pain point.

SLC NAND and Storage Trade-offs in Inference Scenarios

NAND flash is divided into SLC, MLC, TLC, and QLC types by the number of bits each cell can store. SLC stores only 1 bit per cell, with the lowest capacity density and highest cost per unit, but with fast read/write speeds, long lifespan, and high reliability; QLC stores 4 bits per cell, with high density and low cost, but weaker write performance and endurance. AI inference requires frequent reading and writing of context caches (KV Cache), repeated loading of model weights, and handling large numbers of concurrent requests, making it especially sensitive to random read/write latency and write endurance. Replacing some capacity-oriented high-cost solutions with SLC NAND is essentially re-balancing between "cost per GB" and "cost per IOPS," trading less capacity for lower latency and higher endurance—a clear departure from the training phase's capacity-centric approach.

Google Gemini Launches In-App File Editing

Google announced that all Gemini users can now generate, preview, and edit PDFs, Google Docs, Google Sheets, Slides, Markdown files, Text files, and code files directly within the app. PDF preview and editing are supported on both web and mobile, while other formats are currently limited to web. Users can open documents in full screen to edit, and can select text within a file to ask Gemini questions or have it modify content.

Anthropic launched a similar feature earlier this year, allowing users to generate and edit documents and presentations within the Claude chat window. Two leading AI companies independently making "conversational file editing" a product direction shows that AI assistants are migrating from pure Q&A tools to productivity workbenches.

The essence of this trend: AI is no longer satisfied with "telling you how to write," but directly "editing for you." For knowledge workers who handle large volumes of documents daily, this means one fewer round trip of copy-pasting and format adjustment in their workflow.

OpenAI Raises Codex Usage Quotas

OpenAI announced an immediate increase in usage quotas for Work and Codex across ChatGPT, Pro, and Business Premium subscriptions, with quotas for Pro100, Pro200, and Pro500 directly doubled. Previously, OpenAI had planned to cut Pro200 subscription quotas by about 50 percentage points, triggering strong backlash from subscribers who accused it of undermining paid commitments. OpenAI has now made clear it will not reduce usage limits for the relevant subscriptions, and expanded usage for always-on AI assistance will remain independently calculated.

Note the consumption differences between modes: Fast mode consumes quota about 2.5 times faster than standard mode, and UltraFast up to 8 times, with actual output still depending on the model and task complexity. For developers who use Codex heavily, doubled quotas are a substantive benefit, but users who frequently use Fast mode still need to watch their consumption pace.

Diffusion Language Model Clover 35BA3B: Twice the Speed

Diffusion Intelligence released a new diffusion language model, Clover 35BA3B Preview, using a MoE architecture with 35 billion total parameters and about 3 billion activated parameters per dimension. In the 49 tasks of the Wind Tunnel web agent benchmark, Clover's median execution time was 1.52 seconds, compared with 3.21 seconds for the Jeff and Mercury2.5 combination—more than half the time.

Through fast diffusion and a reconstructed gated Delta network, it lets historical memory and current generation each play their roles. This breakthrough pushes diffusion language models further into code and agent tasks, providing a new way to handle dependencies between content. Progress of diffusion models on language tasks has long been constrained by the balance between generation quality and speed; Clover compressing execution time to about half in agent scenarios shows that this technical route has practical value in scenarios requiring low-latency interaction.

The Difference Between Diffusion Language Models and Traditional Autoregressive Generation

Mainstream large language models use autoregressive generation, producing tokens one by one from left to right serially; the nth token must wait for the previous n-1 to be generated, so latency grows linearly with output length. Diffusion language models borrow the diffusion approach from image generation: starting from a noise-perturbed initial state, they iteratively denoise over multiple rounds to gradually recover the complete text, theoretically allowing multiple positions to be updated in parallel and thus compressing generation time. Early diffusion language models were constrained by the balance between generation quality and the number of iteration steps, with text coherence inferior to autoregressive models. Clover, through fast diffusion and a gated Delta network that divides labor between historical memory and current generation, is alleviating this bottleneck, demonstrating competitiveness in latency-sensitive tasks such as code and agents.

Microsoft Offers Voluntary Buyouts to 60,000 Employees

Microsoft is offering voluntary buyout packages to about 60,000 employees in the United States, United Kingdom, Canada, and India to reduce operating costs and headcount, thereby avoiding cuts to investment in AI infrastructure. Microsoft currently employs about 223,000 people, down 5,000 from 228,000 in 2025, marking the first year-over-year decline in total headcount in nearly a decade.

The terms offered by this buyout program are essentially the same as Microsoft's standard severance, including cash compensation based on tenure, accelerated stock vesting, and several months of career services. By continuously adjusting team size, Microsoft is trying to achieve more efficient resource allocation and higher employee compensation levels during its AI transition.

Viewed alongside the previous stories, this news sketches a clear thread: capital expenditure on AI infrastructure is squeezing the traditional labor cost structure. Microsoft chose to control headcount through buyouts while directing resources toward compute—pointing in the same direction as SanDisk's observed surge in inference storage demand and OpenAI's expanded compute quotas.

The Structural Background of Capex Squeezing Labor Costs

The main body of capital expenditure at large tech companies has shifted from office buildings and server procurement to AI data centers and accelerator chips. Single investments in GPU clusters, power supply, and cooling facilities often run into billions of dollars and require continuous additional spending over several years, directly compressing the labor budget within operating expenses. Voluntary severance, compared with layoffs, reduces legal and reputational risk, and through conditions such as accelerated stock vesting, induces employees to leave voluntarily while reallocating the saved costs to compute and high-paying AI roles. This is why several companies show a combination of "declining headcount, rising per-capita compensation": the workforce structure is tilting toward AI-related roles rather than simply contracting.

One-line Summary

Mistral open-sources a trillion parameters, SanDisk is bullish on inference storage, and AI capex is reshaping the industry landscape.

Paragraph Summary

This issue covers four threads: models, storage, products, and organization. Mistral AI releases the weights of its trillion-parameter multimodal model Mistral Large 4, entering the top tier of coding agent evaluations under a modified MIT license restricting MaaS resale; SanDisk's earnings report indicates AI inference is reshaping NAND demand, with an order pipeline extending to 2030 and exploration of replacing some high-cost solutions with SLC NAND; Google Gemini launches in-app file generation and editing, aligning with Anthropic's product direction; OpenAI raises Codex usage quotas, doubling Pro tiers; the diffusion language model Clover 35BA3B achieves about half the execution time of comparison solutions in agent evaluations; Microsoft offers voluntary buyouts to about 60,000 employees, with total headcount declining year-over-year for the first time in nearly a decade.