AI Wiki

Explore AI Wiki

Citation and evidence

Magic (AI software engineer)

22 min full readUpdated 36 references

This article's verification

Report a problem with this article

More

Use this article

Raw MarkdownExplore connections

Improve this page

Suggest editRevision history

Topics

AI Code GenerationAI Companies

Cite this article

Magic (legally Magic AI, Inc., often referenced by its primary domain magic.dev) is an American artificial intelligence company, founded in 2022 by Eric Steinberger and Sebastian De Ro, that trains proprietary frontier large language models for code with the stated aim of automating software engineering and AI research.[1][2] Magic is best known for an ultra-long-context bet: in August 2024 it announced LTM-2-mini, which it described as its first model with a 100 million-token context window. It raised $465 million through August 2024 according to TechCrunch, including a $320 million round backed by Eric Schmidt, and announced a partnership with Google Cloud to build AI training supercomputers.[3][4] In September 2026, after roughly two years without a technical update, Magic published a research post claiming that its pretraining recipe had become more than 10 times more compute-efficient than those behind leading open-weight base models; as of October 2026 its website said the company had raised $768 million.[2][21][22]

Magic logo
Magic logo, from the official website magic.dev. Trademark of its owner.

The company positions itself as building an "AI colleague" or "coworker" for code, rather than the autocomplete-style "copilot" products that dominated the early generative-AI coding market.[1][7] Magic distinguishes itself from rivals such as Cognition AI (the maker of Devin), Cursor, Claude Code, and GitHub Copilot through two technical bets: an emphasis on ultra-long-context language models, culminating in the LTM-2-mini research preview, and a custom long-context evaluation called HashHop designed to replace "needle in a haystack" tests.[3][4] The company has also built much of its identity around a small, secretive research team, a stated mission of building "aligned superintelligence," and a strategy of training proprietary frontier models rather than fine-tuning third-party systems.[5][28]

Through its August 2024 round, Magic's disclosed backers included Nat Friedman and Daniel Gross (whose fund is known as NFDG), Alphabet's CapitalG, Eric Schmidt, Atlassian, Sequoia, Jane Street, and Elad Gil, and it announced a partnership with Alphabet's Google Cloud to construct two AI training supercomputers called Magic-G4 and Magic-G5.[4][6] As of October 2026 Magic had not released a model or a generally available product.[21][22]

Key facts

FieldValue
Founded2022[1]
FoundersEric Steinberger (CEO); Sebastian De Ro (CTO)[1][7]
HeadquartersSan Francisco, California; open roles listed in San Francisco[4][12]
IndustryArtificial intelligence; developer tools; foundation models[4]
Flagship model (announced)LTM-2-mini, 100 million-token context window (Aug 2024), not publicly released[3]
Headcount23 people (company figure, August 2024)[3]
Total funding$465 million through August 2024 (TechCrunch); the company said $515 million at the time and $768 million as of September 2026[3][4][26]
StatusPrivate; valuation not disclosed. Reuters reported in July 2024 that Magic was in talks to raise at a valuation of about $1.5 billion[24]
Compute"Thousands of GB200s" (company website, 2026)[2]

Expanded article table

What is Magic?

Magic is a San Francisco AI startup that is trying to build an autonomous software engineer rather than a coding assistant. Founded in 2022 by Eric Steinberger and Sebastian De Ro, the company trains its own frontier language models specialized for code, and frames its product as a "coworker, not copilot."[1][7] Magic's core wager is that a genuinely useful autonomous engineer requires models that can read and reason over an entire codebase at once, which has led it to pursue context windows orders of magnitude larger than those of mainstream chat models.[3] Its website describes the company as "building frontier code models to automate software engineering and research," and says that the most promising path to safe AGI is "to automate AI research and code generation."[2]

Sequoia Capital's description of its podcast interview with Steinberger summarizes the founding rationale in terms of timelines to artificial general intelligence: "In 2022, Eric realized that AGI was closer than he had previously thought and started Magic to automate the software engineering necessary to get there."[5] This artificial general intelligence-oriented mission, rather than a near-term focus on incremental developer-productivity tools, is a defining feature of how Magic presents itself.[5][20] In its February 2023 funding announcement, Magic described itself as "a public benefit corporation dedicated to building and safely deploying aligned, superhuman AGI."[10]

Founding and early years

Eric Steinberger, who was born in Vienna, became interested in artificial intelligence in his early teens and, while still in secondary school, was mentored by Johannes Heinrich, a former PhD student of David Silver and a researcher at DeepMind, who agreed to review his work regularly.[5] He studied computer science at the University of Cambridge, leaving after about a year, and worked as a researcher at Meta's Fundamental AI Research (FAIR) lab, where he collaborated with Noam Brown on game-theoretic reinforcement learning.[4][5][10] His published work from this period includes "Single Deep Counterfactual Regret Minimization" (2019), a simplified variant of Deep CFR that the paper reports outperforms Deep CFR in poker, and "DREAM" (2020), a model-free deep reinforcement learning algorithm for imperfect-information games co-authored with Adam Lerer and Noam Brown.[29][30] In 2019 Steinberger co-founded ClimateScience, a non-profit climate-education organisation.[5][10]

Sebastian De Ro, Magic's co-founder and chief technology officer, comes from an industry rather than academic research background. He studied higher informatics at HTBLVA Spengergasse in Vienna between 2013 and 2018 and held engineering and leadership roles at the Austrian and German software companies Automic Software, twinformatics GmbH, and the business-process-management firm FireStart, where he rose to development lead, VP of technology, and finally chief technology officer before leaving to co-found Magic.[9] TechCrunch reported that the two met through ClimateScience, where De Ro volunteered.[4][9] The Austrian outlet Trending Topics gives a different account, reporting that the pair already knew each other from a programme for gifted students at HTL Spengergasse, where they once gathered the school's computers over a summer holiday to train AI models, and that De Ro was also involved in ClimateScience.[23]

The company was originally pitched as a "coworker, not copilot" alternative to GitHub Copilot. Magic says it raised a $5 million seed round in summer 2022.[20] In February 2023 it announced a $23 million Series A round led by Alphabet's growth-equity arm CapitalG, with participation from Nat Friedman, Elad Gil, and Amplify Partners, bringing total funding to $28 million.[1][10][16] Magic's own announcement also listed individual backers including Noam Brown.[20] CapitalG partner Jill Chase led the investment and joined Magic's board.[10] At the time, TechCrunch reported that Magic was pre-revenue with "a distributed workforce of six people," and the company's press release said its product operated "like an AI pair programmer."[1][10] Magic's announcement described Friedman as the former CEO of GitHub and co-creator of Copilot.[20]

How does Magic's technology work?

Magic's technology is built around long-context language models, a custom benchmark for measuring long-context performance, and a large training-infrastructure partnership with Google Cloud. The company presents model architecture and training infrastructure as joint prerequisites for an autonomous coding agent.[3] In 2026 it described its research program as resting on three parts: pretraining, agentic reinforcement learning, and long context.[21]

What are Magic's long-context LTM models?

Magic's central technical bet is that producing a useful autonomous software engineer requires language models capable of reading and reasoning over far larger code contexts than typical chat-tuned LLMs. The company has pursued this through a model family it calls LTM (for Long-Term Memory); it calls the underlying architecture the Long-term Memory Network, or "LTM Net."[3][19]

The first prototype, LTM-1, was announced on 6 June 2023. Magic described it as an LLM "with 5,000,000 prompt tokens," equal to "~500k lines of code or ~5k files, enough to fully cover most repositories," and as "a prototype of a neural network architecture we designed for giant context windows."[17][18] Magic's blog post at the time said LTM-1 "enables 50x larger context windows than transformers," and that its coding assistant built on the architecture was in closed alpha.[34] The company said it had tried to scale standard GPT context windows, "quickly got stuck," and built a custom ML stack to train and serve LTM Nets.[19] It also cautioned that "LTM-1 has fewer parameters than today's frontier models, making it less smart."[19]

What is Magic's 100 million token context window?

On 29 August 2024 Magic announced LTM-1's successor, LTM-2-mini, which it described as "our first 100M token context model."[3] The company stated that "100M tokens equals ~10 million lines of code or ~750 novels."[3] TechCrunch noted at the time that this was by far the largest context window claimed for any commercial model, the next-largest being Google's Gemini models at 2 million tokens.[4] Magic emphasised efficiency rather than raw scale alone: it stated that "for each decoded token, LTM-2-mini's sequence-dimension algorithm is roughly 1000x cheaper than the attention mechanism in Llama 3.1 405B for a 100M token context window."[3] It added that "running Llama 3.1 405B with a 100M token context requires 638 H100s per user just to store a single 100M token KV cache," whereas LTM "requires a small fraction of a single H100's HBM per user for the same context."[3] Magic said that to train and serve these models it had written "an entire training and inference stack from scratch (no torch autograd, lots of custom CUDA)."[3]

Magic was candid about the limits of the prototype. It said it had trained a prototype model on text-to-diff data with its long-context mechanism that was "several orders of magnitude smaller than frontier models," and that its code synthesis abilities "were not good enough yet." It showed two demonstrations: the model built a calculator using a custom GUI framework supplied only in context, and implemented a password strength meter for the open-source Documenso repository without human intervention.[3] The post said Magic was "now training a large LTM-2 model" on its new supercomputer, and a footnote estimated that context for its largest LTM-2 model would cost about twice as much as for LTM-2-mini, "still 500x cheaper than Llama 405B."[3] The company also set out a goal of spending "$100 and 10 minutes on an issue" and reliably getting "a great pull request for an entire feature," and said it saw inference-time compute as "the next frontier in AI."[3] As of October 2026 no LTM-2 model had been released for outside use.[21][22]

What is the HashHop benchmark?

Alongside LTM-2-mini, Magic introduced and open-sourced a new long-context evaluation called HashHop. The company argued that prior long-context tests, such as the popular "Needle in a Haystack" evaluation, let models exploit semantic cues: a needle that stands out from its haystack can be found without storing the rest of the context, which can let even retrieval-augmented systems or small-state recurrent models appear successful.[3] HashHop replaces those needles with random hash strings; according to Magic, "hashes are random and thus incompressible, requiring the model to be able to store and retrieve from the maximum possible information content for a given context size at all times."[3] The model is asked to complete chains of hash-to-hash mappings (an approach Magic credited to the earlier RULER benchmark), with the order of pairs shuffled to test order and position invariance and the chain length controlling how many "hops" are required. A harder variant asks the model to skip directly from the first hash to the last, without writing out the intermediate steps.[3] Magic published result charts for its LTM architecture trained on hashes with chain of thought, and reported that, for the hyperparameters of that model, it saw strong results when skipping two hops at once without chain of thought but worsening performance at three or more.[3] The HashHop code is published on GitHub under the MIT license.[31]

In a January 2026 Hugging Face community post, the developer Asankhaya Sharma described HashHop as "a benchmark left behind by an AI startup that raised half a billion dollars and then went silent," and reported building a solver that reached perfect HashHop accuracy by treating each hash string as a single token, an approach that relies on retrieval rather than a long context window.[32]

What infrastructure does Magic use to train its models?

Magic disclosed in August 2024 that it had agreed a partnership with Google Cloud to build two purpose-built AI training systems:

  • Magic-G4, powered by "NVIDIA H100 Tensor Core GPUs."[3][4][6]
  • Magic-G5, powered by "NVIDIA GB200 NVL72, with the ability to scale to tens of thousands of Blackwell GPUs over time."[3][4][6][11]

TechCrunch reported that Magic said the two clusters together would be able to reach 160 exaflops, and SiliconANGLE attributed the same figure to a Google blog post.[4][11] At the time Magic described itself as "23 people (+ 8000 H100s)" and said it was hiring supercomputing and systems engineers to work with Ben Chess, a former supercomputing lead at OpenAI.[3][4] Internet Archive captures of Magic's website from late 2024 through December 2025 say the company had "8,000 H100s"; by March 2026 the site said it had "thousands of GB200s."[2][33][36]

What products does Magic offer?

The core product Magic has been working toward is an autonomous coding agent that the founders have repeatedly described as a "coworker" or "AI colleague" capable of writing, reviewing, debugging, and planning code changes across an entire codebase, rather than offering inline autocompletion.[1][7] This positions Magic in the broader field of AI code generation.

At its 2023 Series A, TechCrunch reported that Magic's tool was "not yet generally available" and that the company planned to launch it without giving a date.[1] Magic's website in 2023 invited developers to join a waitlist, and its June 2023 LTM-1 post said the assistant was in closed alpha.[34] At the time of its $320 million round in August 2024, TechCrunch reported that Magic's tools were still "not yet for sale," and that the company employed roughly two dozen people.[4] In its September 2026 research post, Magic said it looked forward to "releasing the thing," and Trending Topics noted that the company had still not shipped a model.[21][22]

The "coworker, not copilot" framing has been a persistent feature of how Magic positions its product. In its 2023 interview with TechCrunch, Steinberger said that "early releases will need human supervision, but our ultimate aim is for AI to complete large tasks reliably for you, end-to-end, without babysitting."[1] By 2024 the company framed the goal as a system that, given an issue, reliably produces a full pull request for an entire feature.[3]

2026 pretraining research update

On 8 September 2026 Magic published ">10x More Efficient Pretraining," its first technical post since August 2024.[21][35] All figures in it are the company's own and had not been independently verified when Trending Topics covered the post.[22]

ClaimCompany figure
Compute efficiencyPretraining recipe ">10x more compute-efficient" than that of leading open-weight base models[21]
Matching DeepSeek V4 Pro BaseAbout 50x fewer FLOPs, "around half of GPT3's pretraining compute, or ~$0.5M on GB200"[21]
Larger runScaled a further 10x (about $4 million), which it said "meaningfully outperformed all publicly available open base models on perplexity evals"[21]
Progress across generationsInternal models V2 (late 2024), V3 (early 2026), V4 (July 2026) and V5 (September 2026); V5 charted at about 505x the compute efficiency of V2[21]
Math reinforcement learning check72% pass@1 for a V5 model (the e24 run) on a private set of 400 competition math problems after a short RL run with a 16k-token chain-of-thought budget, started directly from the base model without SFT or distillation[21]

Expanded article table

Magic measured bits-per-byte loss on held-out data rather than standard benchmarks, arguing that base models are too sensitive to prompting for sampling-based evaluations. Its held-out sets were its own codebase and private codebases it had acquired from other startups, recent low-citation research papers, and private math problems, and it said it used a different parser than its pretraining pipeline to build them.[21] The comparison models were DeepSeek V4 Pro and Flash, Kimi K2 and Nemotron 3 Ultra; Magic said base models for Claude, Gemini and GPT were not openly available, and that it worked with Fireworks AI to check baseline log-probabilities on Fireworks' own inference engine.[21]

The post said Magic had trained "a small dense model with an architecture designed for very long context windows" in late 2024, and that its first pretraining scale-ups "kept blowing up in a wide variety of ways" before it rebuilt a stable training foundation.[21] It named long-horizon reinforcement learning, "training agents to keep learning after deployment through long-context," and alignment training as next priorities, and described Magic as "likely the smallest team in the world training trillion parameter models."[21] Trending Topics summarized the post under the headline "AI Startup Claims Frontier-Level Pretraining for a Few Million Dollars" and stressed that "every figure comes from the company, none has been independently verified, and Magic has still not shipped a model."[22]

Safety policy

On 2 July 2024 Magic published version 1.0 of an "AGI Readiness Policy," which it said was created with assistance from METR. The policy committed Magic to implementing dangerous-capability evaluations and mitigations before publicly deploying models that exceed the frontier of coding performance, using a 50% score on LiveCodeBench as one trigger threshold, and named four covered threat models: cyberoffense, AI R&D, autonomous replication and adaptation, and potentially biological weapons assistance.[27] By October 2026 the policy page carried a note saying it was outdated and that an update was in progress, and the September 2026 research post said an updated policy would cover "deployment gates, safety requirements during RL training, and how we approach alignment research."[21][27] Magic's safety page states that its mission is "to build aligned superintelligence" and argues that voluntary safety frameworks "would be most impactful if written into law, internationally."[28]

How much funding has Magic raised?

Magic has announced four funding rounds. Publicly reported totals rose from $28 million in early 2023 to $465 million after its August 2024 round, according to TechCrunch; Magic itself put the total at $515 million at that time.[1][3][4]

RoundDateAmountLead / notable investors
Seedsummer 2022$5 millionnot disclosed[20]
Series A6 February 2023$23 millionCapitalG (lead); Nat Friedman, Elad Gil, Amplify Partners, Noam Brown and others[1][10][20]
Series BFebruary 2024$117 millionNFDG (Nat Friedman and Daniel Gross, lead); CapitalG, Elad Gil[7][13]
Unlabeled round29 August 2024$320 millionNew investors Eric Schmidt, Jane Street, Sequoia, Atlassian; existing investors Nat Friedman and Daniel Gross, Elad Gil, CapitalG[3][4][6]

Expanded article table

Reuters reported in early July 2024, citing three sources, that Magic was in talks to raise more than $200 million at a valuation of $1.5 billion, which would triple the $500 million valuation PitchBook data attached to its February 2024 round.[8][24] The same report said Magic was "earning no revenue and having no product for sale" and had about 20 employees.[24] When Magic announced the $320 million round the following month, TechCrunch wrote that the startup's valuation "couldn't be ascertained."[4] Magic has not disclosed a valuation for any round.

By mid-September 2026 Magic's website said the company had raised $768 million from investors including Nat Friedman, Daniel Gross, CapitalG, Elad Gil, Sequoia, Jane Street and Eric Schmidt; Internet Archive captures show the site gave the figure as $515 million until at least 8 September 2026.[2][25][26] Magic has not published a funding announcement explaining the difference, and its blog lists no funding post after August 2024.[35]

In June 2025, CNBC and Axios reported that Meta planned to hire Friedman and Gross for its AI effort and take a stake in their NFDG venture fund, a plan first reported by The Information.[14][15] NFDG had led Magic's Series B and continued as an investor in the August 2024 round; Magic's website still listed Friedman and Gross among its investors in October 2026.[2][3]

What is Magic's partnership with Google Cloud?

Magic's most prominent disclosed partnership is the Google Cloud agreement announced concurrently with its August 2024 funding, under which Magic is building the Magic-G4 and Magic-G5 supercomputers on Google Cloud infrastructure.[4][6][11] The Magic-G5 system, based on NVIDIA Blackwell GB200 NVL72 hardware, is intended to scale to tens of thousands of GPUs.[3][4][6] Google Cloud's Amin Vahdat said Magic would use "a variety of leading NVIDIA chips and AI tooling from Vertex AI" to build and train its next models, and NVIDIA's Ian Buck described Magic-G5 as "one of the largest installations of the NVIDIA GB200 NVL72 rack-scale design to date."[3]

Atlassian was among the new investors in the August 2024 round; no product integration between the two companies has been announced.[3][4] CapitalG, which led the 2023 Series A, gives Magic a second tie to Alphabet.[10]

How does Magic compare to competitors?

Magic operates in one of the most heavily funded segments of the generative-AI industry. The company's most direct rival on the "autonomous software engineer" framing is Cognition AI, the maker of Devin, which was reported in mid-2024 to have raised $175 million at a $2 billion valuation, a higher valuation than the one Magic was then seeking despite Cognition being a younger company.[8][24] Coverage of Magic's 2024 fundraising placed it among better-funded AI coding startups including Codeium, Cognition, Poolside, Cursor maker Anysphere and Augment, all competing with Microsoft's GitHub Copilot.[4][24]

Sequoia's summary of its interview with Steinberger lists among his "counterintuitive ideas" the need to train proprietary large models and the view "that value will not accrue in the application layer."[5] In the same interview he described software engineering as "the minimum viable AGI, because then it creates everything else," and argued that the cost of building AGI, while very high, is "not $100 billion."[5]

Magic remains an unusually secretive company by the standards of well-funded AI startups. Its blog contains four posts between February 2023 and September 2026; it has not released model weights or a product, and it has not published peer-reviewed evaluations of its LTM models.[21][35] Reuters noted in 2024 that investors were backing Magic despite its having no revenue and no product for sale, and coverage of its 2026 research update stressed that its claims were unverified.[22][24]

Magic's competitive narrative has also rested on the claim that ultra-long context is a category-defining advantage for code-focused models. When it was announced in August 2024, LTM-2-mini's 100-million-token window was about 50 times larger than the 2-million-token windows of Google's Gemini models, the largest offered commercially at the time.[4] Its September 2026 post added pretraining efficiency to that pitch, describing long context as the first of three ingredients it believes are sufficient "to build superhuman coding agents and automate AI R&D."[21]

See also

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12"Code-generating platform Magic challenges GitHub's Copilot with $23M in VC backing", TechCrunch, 6 February 2023. techcrunch.com/...de-generating-startup-raises-23m
  2. ^1 ^2 ^3 ^4 ^5 ^6 ^7Magic, "Magic" company website. magic.dev
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30Magic, "100M Token Context Windows", company blog, 29 August 2024. magic.dev/...100m-token-context-windows
  4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22"Generative AI coding startup Magic lands $320M investment from Eric Schmidt, Atlassian and others", TechCrunch, 29 August 2024. techcrunch.com/...ric-schmidt-atlassian-and-others
  5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8"Founder Eric Steinberger on Magic's Counterintuitive Approach to Pursuing AGI", Sequoia Capital "Training Data" podcast. sequoiacap.com/...training-data-eric-steinberger
  6. ^1 ^2 ^3 ^4 ^5 ^6"Magic Secures $320M Funding, Partners with Google Cloud for AI Supercomputers", Maginative, August 2024. maginative.com/...ogle-cloud-for-ai-supercomputers
  7. ^1 ^2 ^3 ^4 ^5"Magic AI Secures $117 Million to Build an AI Software Engineer", Maginative, 16 February 2024. maginative.com/...to-build-an-ai-software-engineer
  8. ^1 ^2"AI Coding Software Magic Looking To Conjure Up $200M At $1.5B - Report", Crunchbase News, 3 July 2024. news.crunchbase.com/...c-cognition-builder-unicorn
  9. ^1 ^2"Sebastian De Ro - Co-founder at Magic.Dev", The Org. theorg.com/...sebastian-de-ro
  10. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8"Magic.dev Raises $28 Million To Build AI Software Engineer", PR Newswire, 6 February 2023. prnewswire.com/...d-ai-software-engineer-301738983
  11. ^1 ^2 ^3"AI coding assistant startup Magic closes $320M round as rival Codeium nabs $150M", SiliconANGLE, 29 August 2024. siliconangle.com/...-round-rival-codeium-nabs-150m
  12. ^Magic, "Careers". magic.dev/careers
  13. ^"Magic AI Raises $117M for Generative AI Coding 'Coworker' Not 'Copilot'", Voicebot.ai, 20 February 2024. voicebot.ai/...tive-ai-coding-coworker-not-copilot
  14. ^"Meta tried to buy Safe Superintelligence, hired CEO Daniel Gross", CNBC, 19 June 2025. cnbc.com/...perintelligence-hired-ceo-daniel-gross
  15. ^"Meta reportedly plans big new AI hires", Axios, 20 June 2025. axios.com/...meta-zucker-berg-ai-daniel-gross
  16. ^"Magic: Reimagining Software Engineering with AI", CapitalG. capitalg.com/...ining-software-engineering-with-ai
  17. ^Magic, "Introducing LTM-1", company blog, 6 June 2023. magic.dev/...ltm-1
  18. ^Magic (@magicailabs), "Meet LTM-1: LLM with 5,000,000 prompt tokens", post on X, 6 June 2023. x.com/...1666116935904292869
  19. ^1 ^2 ^3Magic (@magicailabs), thread posts on LTM-1 ("How?" and "What's next? More compute."), X, 6 June 2023. x.com/...1666116949560967168 and x.com/...1666116951226163203
  20. ^1 ^2 ^3 ^4 ^5 ^6Eric Steinberger, "Magic's $23M Series A and a note on finding meaning in an automated world", Magic blog, 6 February 2023. magic.dev/...series-a
  21. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18Magic, ">10x More Efficient Pretraining", company blog, 8 September 2026. magic.dev/...pretraining
  22. ^1 ^2 ^3 ^4 ^5 ^6 ^7Jakob Steinschaden, "Magic: AI Startup Claims Frontier-Level Pretraining for a Few Million Dollars", Trending Topics, 8 September 2026. trendingtopics.eu/...ing-for-a-few-million-dollars
  23. ^Jakob Steinschaden, "Magic: Der Überraschungsgegner von OpenAI kommt aus einer Wiener Schule" (in German), Trending Topics, 16 February 2024. trendingtopics.eu/magic-openai-challenger
  24. ^1 ^2 ^3 ^4 ^5 ^6Reuters, "Magic AI Startup Aims for $1.5 Billion Valuation Despite No Revenue", syndicated by Devdiscourse, 3 July 2024. devdiscourse.com/...n-valuation-despite-no-revenue
  25. ^Internet Archive, snapshot of magic.dev, 8 September 2026 ("raised $515 million"). web.archive.org/...magic.dev
  26. ^1 ^2Internet Archive, snapshot of magic.dev, 13 September 2026 ("raised $768 million"). web.archive.org/...magic.dev
  27. ^1 ^2Magic, "AGI Readiness Policy", version 1.0, 2 July 2024. magic.dev/agi-readiness-policy
  28. ^1 ^2Magic, "Safety". magic.dev/safety
  29. ^Eric Steinberger, "Single Deep Counterfactual Regret Minimization", arXiv:1901.07621, 2019. arxiv.org/...1901.07621
  30. ^Eric Steinberger, Adam Lerer, Noam Brown, "DREAM: Deep Regret minimization with Advantage baselines and Model-free learning", arXiv:2006.10410, 2020. arxiv.org/...2006.10410
  31. ^Magic, "hash-hop: Long context evaluation for large language models", GitHub. github.com/...hash-hop
  32. ^Asankhaya Sharma, "Reverse Engineering a $500M Mystery: From HashHop to Memory-Augmented Language Models", Hugging Face community blog, 23 January 2026. huggingface.co/...reverse-engineering-magic-hashhop
  33. ^Internet Archive, snapshot of magic.dev, 5 March 2026 ("thousands of GB200s"). web.archive.org/...magic.dev
  34. ^1 ^2Magic, "LTM-1: an LLM with a 5,000,000 token context window", company blog, 6 June 2023 (Internet Archive capture). web.archive.org/...ltm-1
  35. ^1 ^2 ^3Magic, "Blog" index. magic.dev/blog
  36. ^Internet Archive, snapshot of magic.dev, 31 December 2025 ("8,000 H100s"). web.archive.org/...magic.dev

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

10 revisions by 1 contributor · v10 · 4,369 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Funding history, model announcements and company statements checked against Magic's blog and website, its posts on X, archived snapshots, and TechCrunch, Reuters and CNBC coverage, October 2026.

Cite this page: AI Wiki. "Magic (AI software engineer)." aiwiki.ai, updated 28 Oct 2026, fact-checked 27 Oct 2026. CC BY 4.0. https://aiwiki.ai/wiki/magic_dev

Suggest edit

Where to go next

Topics:AI Code GenerationAI Companies