Skip to content

#部署/工程

11 today

Apr 6Monday

X · @dotey

Xiaomi MiMo lead Luo Fuli on token costs in the Agent era

Luo Fuli said Agent workloads can resend 100k+ tokens across repeated tool calls, and global compute cannot keep up with that burn. She said OpenClaw makes several times more requests than Claude Code and can push real API cost to tens of times the subscription price; the post does not disclose a pricing formula.

Why it matters: Xiaomi MiMo's lead says agents that drag 100k-token contexts into every tool call can cost dozens of times a subscription, forcing third-party frameworks to fix context management.

Apr 4Saturday

X · @dotey

DeepSeek's next-generation V4 model will run on Huawei chips

DeepSeek delayed V4 for months and rewrote some low-level modules with Huawei and Cambricon so it runs on Huawei's Ascend 950PR, with launch expected in weeks, per The Information. The post cites 112GB memory, 1.4TB/s bandwidth, 600W power, and FP4 inference support; it does not disclose V4 size, pricing, or measured performance.

Why it matters: DeepSeek rewrote low-level model modules to fit domestic hardware such as Huawei Ascend 950PR, pushing V4 back by months; it shows a concrete path for domestic chips to run large-model inference.

Mar 26Thursday

TheValley101 (硅谷101)

E230 | Behind the $1 trillion revenue forecast: NVIDIA's peak and weak spots

Jensen Huang said at GTC that NVIDIA expects at least $1 trillion in cumulative orders for Blackwell and Vera Rubin by the end of 2027, above the roughly $600B global semiconductor market in 2024 cited in the episode. The discussion adds that Vera Rubin launched 7 chips at once, NVL72 delivers 10x inference efficiency over Blackwell, cuts cost per token to one-tenth, and improves token per watt by 35x; the real constraint discussed is CoWoS, HBM4, and power capacity, not demand alone.

Why it matters: Nvidia's Jensen Huang put a $1 trillion order figure on Blackwell and Vera Rubin through 2027, and the show tests whether that holds up.

Mar 24Tuesday

Lex Fridman (YouTube RSS)

Jensen Huang: NVIDIA - The $4 Trillion Company & the AI Revolution | Lex Fridman Podcast #494

Jensen Huang said on the Lex Fridman podcast that NVIDIA uses “extreme co-design” for AI clusters, aiming to beat linear scaling across 10,000 computers. The interview cites Amdahl’s Law, model and data sharding, networking, power, and cooling as hard constraints; Huang also said he has 60+ direct reports. The key shift is that NVIDIA now competes at rack and data-center level, not only at single-GPU level.

Why it matters: Jensen Huang explains on a podcast why NVIDIA moved to rack-scale and data-center co-design, and why single-chip gains hit Amdahl's law limits without algorithm sharding plus network, power and cooling work.

Mar 17Tuesday

NVIDIA Blog

GTC spotlights NVIDIA RTX PCs and DGX Spark running latest open models and AI agents locally

NVIDIA used GTC to showcase RTX PCs and DGX Spark for running local AI agents, and announced Nemotron 3 Nano 4B, Nemotron 3 Super 120B, and the open source NemoClaw stack. The post says DGX Spark has 128GB unified memory for models above 120B parameters; Nemotron 3 Super scored 85.6% on PinchBench, and Qwen 3.5 supports a 262,000-token context window. The key signal is local inference for privacy and zero token cost, while the full “latest open models” lineup and pricing are not disclosed in the post.

Why it matters: At GTC, NVIDIA showed AI assistants running locally on RTX PCs and DGX Spark, whose 128GB unified memory holds models above 120B parameters, with privacy and token cost as the two payoffs.

Mar 13Friday

MIT Technology Review · AI

Future AI chips could be built on glass

Absolics plans to start commercial glass-substrate production in 2026 for AI data-center chip packaging. The post gives three concrete metrics: up to 10x more connections per millimeter, 50% more silicon in the same package area, and 12,000 square meters of annual panel capacity. The real issue is packaging limits, not material hype; Intel has shown a glass-core device that booted Windows, while large-scale yield and cost are not disclosed.

Why it matters: South Korea's Absolics plans to mass-produce glass substrates for AI data center chip packaging at its US plant this year, holding 50% more silicon per area with lower power draw.

Mar 10Tuesday

NVIDIA Blog

NVIDIA and Thinking Machines Lab Announce Long-Term Gigawatt-Scale Strategic Partnership

NVIDIA and Thinking Machines Lab formed a multiyear deal to deploy at least 1 gigawatt of NVIDIA Vera Rubin systems, targeted for early next year, for frontier model training and customizable AI platforms. The partnership also covers training and serving system design for NVIDIA architectures and broader access to frontier and open models for enterprises and researchers; the post does not disclose the investment size. The key signal is the explicit 1-gigawatt compute commitment, not a routine cloud purchase.

Why it matters: NVIDIA and Thinking Machines Lab signed a multi-year deal to deploy at least 1 gigawatt of Vera Rubin systems for training frontier models.

Mar 7Saturday

Bloomberg Technology

Oracle and OpenAI End Plans to Expand Flagship Data Center

Oracle and OpenAI ended talks to expand a flagship AI data center in Abilene, Texas, after financing delays and OpenAI's changing needs. Meta is considering leasing the site from Crusoe, and Nvidia helped facilitate talks; the post only says such projects cost tens of billions of dollars.

Why it matters: Oracle and OpenAI halted expansion of the Abilene, Texas flagship data center over financing and shifting demand, with Meta weighing a takeover of the site.

Bloomberg Technology

US Considers Permits for Global Nvidia, AMD AI Chip Sales | Bloomberg Tech 3/6/2026

The US Commerce Department has reportedly drafted rules that would require American approval before Nvidia and AMD AI chips ship anywhere globally. The RSS snippet also says Oracle plans thousands of job cuts amid cash strain from AI data center expansion, and the Pentagon told lawmakers Anthropic poses a US supply-chain risk. The post does not disclose permit thresholds, layoff details, or the basis for the Anthropic finding.

Why it matters: A draft Commerce rule would require licenses for Nvidia and AMD AI chip sales to any country, widening export controls from specific regions to the whole map.

Bloomberg Technology

OpenAI, Oracle Won't Expand Flagship AI Data Center in Texas

OpenAI and Oracle have scrapped plans to expand a flagship AI data center in Texas after financing talks dragged and OpenAI's needs changed. The RSS snippet confirms only the Texas site; the post does not disclose the facility name, target capacity, capex, or revised timeline. The signal to watch is shifting compute demand, not just a stalled real estate project.

Why it matters: The stalled Texas build-out shows how financing talks and OpenAI's own compute needs can slow the pace at which leading AI labs construct their own capacity.

Bloomberg Technology

Oracle and OpenAI End Plans to Expand Flagship Data Center

Oracle and OpenAI ended plans to expand a flagship AI data center in Texas. The RSS snippet says talks dragged over financing and OpenAI’s changing needs; the post does not disclose the site’s size, budget, or timeline. The real signal is financing friction plus a demand reassessment.

Why it matters: Oracle and OpenAI dropped expansion of their flagship Texas AI data center, with talks stuck on financing and OpenAI reassessing its compute needs.

Feb 28Saturday

36Kr (direct RSS)

Qwen plans AI glasses, earbuds, and rings as tech giants race for a new AI entry point

A report says Alibaba's Qwen plans AI glasses, earbuds, and rings for a global launch in 2026; the glasses are slated for MWC 2026, with reservations opening on March 2. The post adds that Qwen app functions like food delivery and ride hailing will move to these devices, and cites Qwen3.5-Plus with 60% lower memory use, up to 19x inference throughput, and RMB 0.8 per million tokens. The real point is distribution: if the hardware connects Alipay, Amap, and Taobao, Alibaba is chasing the consumer AI entry layer, not just device sales.

Why it matters: Alibaba plans Qwen AI glasses, earbuds and a ring for 2026, moving food delivery and ride hailing from its app onto wearables, as model apps push past the phone.

36Kr (direct RSS)

36Kr 9AM Briefing: Lynk apologizes after voice command headlight crash; OpenAI raises $110B; miHoYo reports employee death

OpenAI said it raised $110B, with $30B each from SoftBank and NVIDIA and $50B from Amazon, at a $730B pre-money valuation. The post adds a strategic partnership with Amazon and a next-gen inference compute deal with NVIDIA.

Why it matters: OpenAI closed a $110 billion round at a $730 billion pre-money valuation, with SoftBank and Nvidia at $30 billion each and Amazon at $50 billion, alongside compute and cloud partnerships.

Feb 20Friday

Hugging Face Blog

GGML and llama.cpp join Hugging Face to support the long-term progress of Local AI

Hugging Face said the GGML and llama.cpp team is joining the company, while Georgi Gerganov’s team will still spend 100% of its time maintaining llama.cpp. The post says the project remains 100% open source and community driven, with full technical and community autonomy. The key angle is tighter delivery from transformers model definitions into llama.cpp, aiming for near “single-click” shipping; the post does not disclose timeline, team size, or deal terms.

Why it matters: The GGML and llama.cpp team joined Hugging Face while staying full-time on llama.cpp, keeping technical direction and community management independent, so the core local-inference engineering effort gains long-term backing.

Feb 6Friday

Dwarkesh Patel

Elon Musk: “In 36 months, the cheapest place to put AI will be space”

Elon Musk predicts that in 30–36 months, space will become the cheapest place to deploy AI compute. He cites flat power growth, permitting bottlenecks, and roughly 5x better solar output in space without batteries; the interview does not disclose a cost model or validation data.

Why it matters: Musk gives a concrete timeline and cost argument for putting AI compute in orbit rather than on the ground.

Jan 30Friday

Bloomberg Technology

Perplexity Inks Microsoft AI Cloud Deal Amid Dispute With Amazon

Perplexity signed a $750 million Azure cloud deal with Microsoft while facing a legal dispute with its longtime cloud partner Amazon. The RSS snippet discloses the deal size, cloud provider, and dispute context, but not the contract term, compute scale, or lawsuit details. The key signal is a cloud supply rebalance that can affect training and inference costs.

Why it matters: Perplexity signed a $750 million Microsoft Azure cloud contract while a long-time AWS user, against the backdrop of a legal dispute with AWS, shifting the cloud supply picture.

Bloomberg Technology

Amazon in Talks to Invest Up to $50 Billion in OpenAI and Expand Ties

Amazon is in talks to invest up to $50 billion in OpenAI and expand their existing relationship. The RSS snippet says the tie-up includes Amazon selling compute to OpenAI; the post does not disclose deal structure, timing, or whether talks will close. The key signal is compute linkage, not just capital.

Why it matters: Amazon is negotiating an investment of up to $50 billion in OpenAI while also selling it compute, tying the money to cloud services.

Jan 12Monday

36Kr (direct RSS)

He Xiaopeng: The best AI companies in the future will build their own chips

He Xiaopeng said XPeng's four 2026 vehicle models will use its Turing AI chip, and Ultra SE and Ultra trims will run a second-gen VLA model for entry-level L4-assisted driving. The post says MAX uses one 750 TOPS chip, Ultra SE uses two, and Ultra uses three; XPeng has entered 60 countries and regions, and VLA 2.0 is already being road-tested in Europe. The real signal is that automakers are pulling chips, models, and deployment in-house as a ceiling-on-performance play, not just a cost move.

Why it matters: He Xiaopeng says top AI companies will all build their own chips, and put Turing silicon in four new cars, tying chip design to how far product performance can go.

Jan 6Tuesday

NVIDIA Blog

NVIDIA RTX Accelerates 4K AI Video Generation on PC With LTX-2 and ComfyUI Upgrades

NVIDIA said GeForce RTX and related devices can run LTX-2 and updated ComfyUI for local AI video generation up to 3x faster with up to 60% lower VRAM use. The post attributes this to PyTorch-CUDA optimizations, native NVFP4/FP8 support in ComfyUI, and an RTX Video 4K upscaling node due next month; LTX-2 open weights are available now and the workflow ships next month. The real signal for AI builders is that local 4K video is shifting from VRAM-bound demos to usable RTX workflows.

Why it matters: NVIDIA's CES push for local AI video generation shows how much speed and VRAM headroom RTX cards gain from low-precision work.

NVIDIA Blog

NVIDIA presents Rubin platform, open models and autonomous driving roadmap at CES

At CES 2026, NVIDIA said its six-chip Rubin AI platform is now in full production and cuts token generation cost to about one-tenth of the prior platform. The post cites 50 petaflops NVFP4 inference for Rubin GPUs, 5x gains from its KV-cache storage tier, and the new open autonomous-driving model family Alpamayo; the key signal is production status and cost curve, not the “AI everywhere” framing.

Why it matters: Rubin's claimed tenfold cut in token cost and 50 petaflops at NVFP4 set the bar for next-gen inference hardware.

NVIDIA Blog

NVIDIA DGX SuperPOD Sets the Stage for Rubin-Based Systems

NVIDIA introduced Rubin-based DGX SuperPOD systems, with DGX Vera Rubin NVL72 and DGX Rubin NVL8 slated for the second half of this year. One DGX SuperPOD can combine eight NVL72 systems for 576 Rubin GPUs, 28.8 exaflops FP4, and 600TB memory; NVIDIA says inference token cost drops by up to 10x versus the prior generation. The key detail is rack-scale design: 260TB/s NVLink per rack, which the post says removes model partitioning.

Why it matters: The DGX SuperPOD roadmap gives concrete numbers, 260TB/s per rack and a claimed 10x lower inference cost, for Rubin-based systems.

Jan 5Monday

Import AI (Jack Clark)

Import AI 439: AI kernels; decentralized training; and universal representations

Meta says KernelEvolve cut kernel development from weeks to hours and delivered up to 17x over PyTorch baselines in production tests. The system uses Llama, GPT, and Claude to generate kernels, validates them, and feeds results into a knowledge base across NVIDIA, AMD, and MTIA; the post also says decentralized training is growing 20x per year but still uses about 1000x less compute than frontier runs. The real signal is continuous self-optimizing infra in production, while decentralized training matters if that 1000x gap keeps shrinking.

Why it matters: KernelEvolve shows LLMs writing and tuning low-level kernels, cutting weeks of work to hours and beating PyTorch baselines by up to 17x.

Dec 17, 2025Wednesday

Mistral AI

Mistral releases OCR 3 with better forms and handwriting, plus Document AI Playground

Mistral released Mistral OCR 3, which wins 74% of head-to-head comparisons against Mistral OCR 2 on forms, scanned documents, complex tables and handwriting. Mistral says its accuracy beats enterprise document-processing tools and AI-native OCR options.

Why it matters: Mistral OCR 3's upgrades on forms, handwriting and complex tables, plus its $2 per 1,000 pages pricing, are useful when evaluating document parsing options.

Oct 13, 2025Monday

OpenAI News

OpenAI and Broadcom announce collaboration to deploy 10 gigawatts of OpenAI-designed AI accelerators

OpenAI and Broadcom announced a multi-year deal to deploy 10 gigawatts of OpenAI-designed AI accelerators, with rack deployments starting in H2 2026 and completing by the end of 2029. OpenAI will design the accelerators and systems, while Broadcom provides accelerator deployment plus Ethernet, PCIe, and optical networking for OpenAI sites and partner data centers. The key signal is OpenAI's custom-chip plus Ethernet cluster path, but the post does not disclose process node, chip specs, or capex.

Why it matters: It lays out OpenAI's compute route: custom accelerators plus an all-Ethernet cluster, with Broadcom handling networking and full racks.

Oct 1, 2025Wednesday

OpenAI News

Samsung and SK join OpenAI’s Stargate initiative to expand global AI infrastructure

OpenAI said on Oct. 1, 2025 that Samsung and SK joined Stargate, with the partnership centered on Korea’s AI chip supply and data center expansion. The post gives one hard target: Samsung Electronics and SK hynix plan to scale advanced memory output to 900,000 DRAM wafer starts per month, while OpenAI also signed Korean data center exploration agreements with MSIT, SK Telecom, and Samsung affiliates. The key gap is execution detail: the post does not disclose investment size, timeline, or facility scale.

Why it matters: Samsung and SK Hynix join Stargate, plan to expand advanced memory output to 900,000 DRAM wafers a month, and sign a deal for an AI data center outside Seoul.

Sep 23, 2025Tuesday

OpenAI News

OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites

OpenAI, Oracle, and SoftBank announced five new U.S. Stargate AI data center sites, bringing planned capacity to nearly 7 GW and investment to over $400 billion in three years. The post says this keeps Stargate on track to reach its full $500 billion, 10 GW commitment by the end of 2025; Oracle-linked sites account for over 5.5 GW, while two SoftBank-OpenAI sites can scale to 1.5 GW in 18 months. The key signal is supply progress: Abilene is already running early training and inference workloads with first NVIDIA GB200 racks delivered in June.

Why it matters: Five new US sites, close to 7 GW of planned capacity and over $40 billion committed, give AI buildout a concrete scale.

Sep 22, 2025Monday

OpenAI News

OpenAI and NVIDIA announce strategic partnership to deploy 10 gigawatts of NVIDIA systems

OpenAI and NVIDIA signed a letter of intent to deploy at least 10 gigawatts of NVIDIA systems for OpenAI’s next-generation AI infrastructure. NVIDIA plans to invest up to $100 billion into OpenAI as each gigawatt is deployed, and the first 1 GW phase is targeted for H2 2026 on the Vera Rubin platform. The key detail is execution: this is still an LOI, and final terms are not yet closed.

Why it matters: A letter of intent for at least 10 GW of NVIDIA systems and up to $100 billion, with the first gigawatt due in late 2026.

Aug 7, 2025Thursday

OpenAI News

GPT-5 System Card

OpenAI published the GPT-5 System Card on Aug. 7, 2025, stating GPT-5 combines gpt-5-main, gpt-5-thinking, and a real-time router, with mini models used after limits are hit. The API exposes gpt-5-thinking, gpt-5-thinking-mini, and gpt-5-thinking-nano, while ChatGPT adds gpt-5-thinking-pro; the post does not disclose pricing, context window, or benchmark scores. The key signal is safety: OpenAI classifies gpt-5-thinking as High capability in biological and chemical domains and applies the related safeguards.

Why it matters: OpenAI's GPT-5 system card describes a fast model, a deep reasoning model and a real-time router that assigns work by conversation type and complexity.

Aug 5, 2025Tuesday

OpenAI News

Introducing gpt-oss

OpenAI released gpt-oss-120b and gpt-oss-20b under Apache 2.0, with the 120B model running on one 80GB GPU and the 20B model on devices with 16GB memory. Both are MoE Transformers with 117B and 21B total parameters, 5.1B and 3.6B active params per token, 128k context, and support for the Responses API and Structured Outputs. The part that matters is the lower deployment bar plus open weights; the post excerpt claims strong reasoning, but full benchmark scores are not disclosed here.

Why it matters: OpenAI open-sourced gpt-oss-120b and gpt-oss-20b under Apache 2.0, its first language model weights since GPT-2, and the 120B runs on a single 80GB GPU.

Jul 31, 2025Thursday

OpenAI News

Introducing Stargate Norway

OpenAI said Stargate Norway, its first European AI data center project, is planned for 230MW with a further 290MW expansion target. Nscale and Aker will build it in Narvik, aiming for 100,000 NVIDIA GPUs by end-2026, using renewable power and closed-loop direct-to-chip liquid cooling. The key detail is allocation: OpenAI is an initial offtaker, while surplus capacity is intended for users in Norway, the UK, the Nordics, and Northern Europe; the post does not disclose capex or exact GPU models.

Why it matters: OpenAI's first European data center, in Narvik, Norway, starts at 230 MW and targets 100,000 Nvidia GPUs by the end of 2026, all on renewable power.

Jul 30, 2025Wednesday

Mistral AI

Mistral ships Codestral 25.08 and an enterprise coding stack

Mistral AI released Codestral 25.08 along with a full enterprise coding stack: Codestral, Codestral Embed, Devstral and a Mistral Code IDE plugin.

Why it matters: Codestral 25.08 arrives with an enterprise coding stack spanning completion, retrieval and IDE plugins, deployable in cloud, VPC or on-prem, with completion acceptance up 30%.

Jul 22, 2025Tuesday

OpenAI News

Stargate advances with 4.5 GW partnership with Oracle

OpenAI and Oracle agreed to add 4.5 GW of Stargate data center capacity in the U.S., bringing capacity under development to over 5 GW and more than 2 million chips. OpenAI says this advances its January pledge to build 10 GW of U.S. AI infrastructure with $500 billion over four years, and it now expects to exceed that target. The concrete signal is deployment: Stargate I in Abilene has started receiving Nvidia GB200 racks and is already running early training and inference workloads.

Why it matters: Shows how fast OpenAI is locking up data center capacity, with over 5 GW under construction and room for more than 2 million chips.

Jun 4, 2025Wednesday

Mistral AI

Mistral AI launches Mistral Code enterprise coding assistant

Mistral AI released Mistral Code, an enterprise AI coding assistant that combines four models: Codestral, Codestral Embed, Devstral and Mistral Medium. It runs in the cloud, on dedicated capacity or on air-gapped local GPUs, so code stays inside the company's boundary.

Why it matters: Mistral AI bundles four coding models with on-prem deployment and an admin console into an enterprise assistant, already running at Spain's Abanca and France's SNCF.

May 22, 2025Thursday

OpenAI News

Introducing Stargate UAE

OpenAI, with G42, Oracle, NVIDIA, Cisco, and SoftBank, will deploy a 1GW Stargate UAE cluster in Abu Dhabi, with 200MW expected online in 2026. The project is the first OpenAI for Countries deal; OpenAI says the UAE will be the first country with nationwide ChatGPT access, and the site can serve a 2,000-mile radius. What matters is sovereign compute tied to U.S. coordination; the post does not disclose capex split, GPU counts, or how nationwide ChatGPT access will work.

Why it matters: Stargate's first site outside the US, planned at 1GW with 200MW due in 2026, shows how OpenAI is structuring compute deals abroad.

May 8, 2025Thursday

OpenAI News

OpenAI’s response to the Department of Energy on AI infrastructure

On May 7, 2025, OpenAI submitted AI infrastructure proposals to the US Department of Energy, urging federal land use, faster permitting, and financial incentives for AI supercomputer hubs. The post says the first Stargate campus is underway in Abilene, Texas, and more sites are being evaluated in Texas and other states; it does not disclose specific tax, power-pricing, or lease terms. The real signal is policy positioning: OpenAI is framing data centers, energy, and permitting as a national industrial agenda.

Why it matters: OpenAI asks the Department of Energy for federal land, faster permitting and power tax credits, with the Abilene Stargate campus already under construction.

May 7, 2025Wednesday

Mistral AI

Mistral AI launches Le Chat Enterprise, powered by Mistral Medium 3

Mistral AI released Le Chat Enterprise, an enterprise AI assistant powered by the new Mistral Medium 3 model. It includes enterprise search, an agent builder, connectors for custom data and tools, a document library, custom models and hybrid deployment. All features will roll out over the next two weeks.

Why it matters: Mistral AI launches Le Chat Enterprise, an enterprise assistant powered by Mistral Medium 3 with enterprise search, agent building and self-hosted deployment, all due within two weeks.

Mistral AI

Mistral AI releases Mistral Medium 3, targeting low cost and enterprise deployment

Mistral AI released Mistral Medium 3, which it says reaches or exceeds 90% of Claude Sonnet 3.7 across benchmarks. Pricing is $0.4 per million input tokens and $2 per million output tokens.

Why it matters: Mistral Medium 3 benchmarks against Claude Sonnet 3.7 at lower cost, and the post gives a path to private enterprise deployment and customization.

Mar 31, 2025Monday

OpenAI News

OpenAI raises $40 billion at a $300 billion post-money valuation

OpenAI said it raised $40 billion at a $300 billion post-money valuation. The post names SoftBank Group as a partner and says the funds will expand compute infrastructure and support tools for ChatGPT's 500 million weekly users. The AGI framing is broad; the post does not disclose deal structure, funding timing, or product roadmap details.

Why it matters: OpenAI raised $40 billion at a $300 billion post-money valuation, with SoftBank as the key partner, to fund compute infrastructure and ChatGPT development.

Feb 7, 2025Friday

Mistral AI

Mistral AI launches new le Chat with mobile apps and Pro, Team plans

Mistral AI released a new version of le Chat with iOS and Android apps, Pro and Team tiers, and an Enterprise private preview. It adds a code interpreter, document analysis and image generation, and Flash Answers is in open preview for all users at a claimed ~1000 words per second. Most features are free; Pro starts at $14.99 per month.

Why it matters: Mistral AI is giving le Chat both a mobile entry point for daily use and enterprise deployment options for work inside a company's own environment.

Jan 30, 2025Thursday

Mistral AI

Mistral AI releases Mistral Small 3, a 24B model under Apache 2.0

Mistral AI released Mistral Small 3, a 24B model built for low-latency tasks, with pretrained and instruction-tuned checkpoints under Apache 2.0. The company says it matches Llama 3.3 70B instruct, runs more than 3x faster on the same hardware, and scores above 81% on MMLU. The model was not trained with RL or synthetic data.

Why it matters: Mistral AI benchmarks Small 3 against Llama 3.3 70B and reports more than 3x the speed on the same hardware.