DeepSeek gave away its serving stack the week OpenAI shipped its most expensive model ever

WEEK OF FEB 22 – 28, 2025

DeepSeek gave away its serving stack the week OpenAI shipped its most expensive model ever

DeepSeek closed an open-source week by publishing the kernels and file system behind V3 and R1. OpenAI answered with a model its own materials declined to call frontier — at roughly 30x GPT-4o's input price.

DeepSeek open-sourced one production component per day: attention and MoE communication kernels, an FP8 GEMM library, an expert-parallel load balancer, and finally 3FS, the parallel file system behind V3 and R1, reporting 6.6 TiB/s aggregate read throughput on a 180-node cluster.

OpenAI released GPT-4.5 as a research preview, priced at $75/$150 per million tokens and openly described as not a frontier model. Nvidia reported a record $39.3B quarter three weeks after DeepSeek's efficiency shock.

FRI · Feb 28, 2025open sourcedeepseekinferencemoeinfrastructurecost

DeepSeek ends Open Source Week with 3FS, handing over the storage and inference stack behind V3/R1

DeepSeek open-sourced one production component per day from 24 to 28 February: FlashMLA, an MLA decoding kernel for Hopper reporting 3,000 GB/s memory-bound and 580 TFLOPS BF16 compute-bound on H800; DeepEP, the first open expert-parallel all-to-all communication library for MoE with native FP8 dispatch; DeepGEMM, an FP8 GEMM library with runtime JIT and no install-time CUDA compilation; DualPipe plus the EPLB expert-parallel load balancer; and the finale 3FS.

3FS is an MIT-licensed parallel file system for AI training and inference reporting 6.6 TiB/s aggregate read throughput on a 180-node cluster, 3.66 TiB/min on GraySort with 25 nodes, and 40+ GiB/s per client node for KV-cache lookups, with Smallpond as its data-processing layer. These are the pieces DeepSeek actually runs in production, published with the numbers rather than benchmark slides.

DeepSeek ends Open Source Week with 3FS, handing over the storage and inference stack behind V3/R1 — opengraph.githubassets.com

WHY IT MATTERS

The strongest release of the week is not weights but serving economics. Anyone building agents or a self-hosted inference stack got DeepSeek's licensed, production-tested kernels and file system for free, which sets a hard ceiling on what closed labs can charge for the same capability. The cost of self-hosting a frontier-adjacent open MoE is no longer gated by unpublished systems work.

DeepSeek — open-infra-index (day-by-day releases)deepseek-ai/3FS on GitHub (MIT)deepseek-ai/DeepGEMM

THU · Feb 27, 2025openaigpt-4.5model releasepretrainingpricing

OpenAI releases GPT-4.5 as a giant, expensive research preview — and says it is not a frontier model

OpenAI shipped GPT-4.5, its largest chat model, as a research preview for Pro users and API developers, built by scaling unsupervised pre-training rather than chain-of-thought. OpenAI calls it the last of its pre-reasoning models.

On OpenAI's own page: GPQA 71.4%, AIME 2024 36.7%, MMMLU 85.1%, SWE-bench Verified 38.0% — behind o3-mini's 87.3% on AIME and 61.0% on SWE-bench Verified. A much bigger, much costlier model still loses to a small reasoning model on hard math and code. The Verge reported OpenAI's own materials initially called GPT-4.5 not a frontier model with performance below o1 and o3-mini on most preparedness evaluations before that line was removed, and Sam Altman publicly called it a giant, expensive model that won't crush benchmarks. API pricing was $75 per million input tokens and $150 per million output, roughly 30 times GPT-4o's input rate.

OpenAI releases GPT-4.5 as a giant, expensive research preview — and says it is not a frontier model — platform.theverge.com

WHY IT MATTERS

Direct, same-week contrast with the open story: DeepSeek published the kernels that push serving cost down while OpenAI charged frontier-premium prices for a model its own documentation declined to call frontier. For anyone deciding between closed frontier APIs and self-hosting open weights, this week is the argument in miniature — and it is evidence that test-time compute, not raw pre-training scale, is where capability is coming from.

OpenAI — Introducing GPT-4.5 GPT-4.5 System Card (PDF)The Verge — OpenAI announces GPT-4.5, warns it's not a frontier AI model

WED · Feb 26, 2025nvidiacomputeearningsdatacentercapex

Nvidia posts a record $39.3B quarter and $130.5B year, three weeks after the efficiency shock

Nvidia reported fiscal Q4 2025 revenue of $39.3 billion, up 78% year over year, with record data center revenue of $35.6 billion — up 93% and 91% of total sales. Full-year revenue was $130.5 billion, up 114% from $60.9 billion.

Guidance for the current quarter was about $43 billion, above the roughly $41.8 billion analysts expected, and the CFO pointed to a significant ramp of Blackwell sales. Nvidia also flagged its role as a key technology partner in the $500B Stargate Project.

WHY IT MATTERS

This is the compute half of the picture and the answer to the question DeepSeek raised in January: cheap inference has not reduced demand for accelerators. If you are planning a self-host, open weights or not, the hardware floor is still set by Nvidia's supply and pricing — and 114% annual growth is the clearest signal that 2025 capex is going into capacity regardless of whose model wins.

NVIDIA Q4 FY2025 press release (SEC EDGAR)CNBC — Nvidia sales grow 78% on AI demand

WED · Feb 26, 2025open weightsmicrosoftphi-4small modelsfunction callingedge

Microsoft open-weights Phi-4-mini and Phi-4-multimodal under MIT

Microsoft added two small open models to the Phi family: Phi-4-mini, a 3.8B dense decoder-only transformer with grouped-query attention, a 200,000-token vocabulary, 128K context and function calling, and Phi-4-multimodal, a 5.6B model with a mixture-of-LoRAs architecture that takes text, audio and image input in one representation space.

Both shipped under the MIT license on Hugging Face, Azure AI Foundry, GitHub Models, Ollama and the NVIDIA API catalog, making them commercially usable in a way the earlier 14B Phi-4 research-only release was not. Microsoft says Phi-4-multimodal tops the Hugging Face OpenASR leaderboard at 6.14% word error rate.

WHY IT MATTERS

Not near-frontier scale, so this is the other half of the open-weights story: the cheap, permissively licensed, tool-calling tier you can run on a laptop or edge box. 128K context plus function calling on a 3.8B MIT model is exactly the profile for local agent sub-tasks — and it shows the closed labs are willing to give away the small end while the flagship stays behind an API at $75 per million tokens.

Microsoft Azure blog — the next generation of the Phi family Hugging Face — microsoft/Phi-4-mini-instruct (MIT, 128K context)

TUE · Feb 25, 2025googlegeminicoding agentsdeveloper toolspricing

Google makes Gemini Code Assist free for individuals — 180,000 completions a month

Google launched Gemini Code Assist for individuals in public preview at no cost, powered by Gemini 2.0, available in VS Code and JetBrains with a personal Gmail account, supporting all public-domain programming languages and up to 128,000 input tokens in chat.

Google offers up to 180,000 code completions per month, against roughly 2,000 typical of free competitors. In the same announcement Google put Gemini Code Assist for GitHub into preview, giving free AI code review on pull requests in both public and private repositories, with custom style guidelines, bug detection and suggested fixes.

WHY IT MATTERS

Agent tooling economics from the other direction: the completion and review layer of the coding workflow just went free at a scale that makes paid tiers hard to justify for individuals. Combined with Claude Code's gated preview the day before, the week makes clear that the harness — tool calling, code review, repository context — is where the pricing fight is moving.

Google blog — Get coding help from Gemini Code Assist, now for free SiliconANGLE — Google launches free Gemini Code Assist tier

MON · Feb 24, 2025anthropicclaude 3.7reasoningcoding agentsclaude codeagent harness

Anthropic ships Claude 3.7 Sonnet and opens the Claude Code agent preview

Anthropic released Claude 3.7 Sonnet, which it calls the first hybrid reasoning model on the market: one set of weights that answers instantly in standard mode or runs extended, user-visible step-by-step thinking, with API developers controlling the thinking budget and thinking tokens billed as output tokens. It is a 200K-context model available on all Claude plans, with extended thinking everywhere except the free tier.

Alongside it, Anthropic launched Claude Code as a limited research preview — a terminal agent that searches and reads code, edits files, writes and runs tests, uses command-line tools and commits and pushes to GitHub, keeping the human in the loop at each step. Anthropic says 3.7 Sonnet sets state of the art on SWE-bench Verified and TAU-bench.

WHY IT MATTERS

The harness story of the week for anyone building agents: an agentic coding CLI plus a model with an explicit, priced reasoning budget, from a US lab shipping closed weights on the same day DeepSeek was giving its serving stack away. Claude Code made the permission-and-approval layer — what the agent may touch without asking — a first-class product surface, which is the design question every agent builder now has to answer.

Anthropic — Claude 3.7 Sonnet and Claude Code