Company view of AI distillation: who uses it, who sells it, who polices it

Key figures · 8 figures

Key findings · 8 findings

1. Distillation is now the default way every lab builds its small models

   Google's Gemini 2.5 report states plainly that 'the smaller models in the Gemini 2.5 series — Flash size and below — use distillation', and Gemma 2 and Gemma 3 are trained with knowledge distillation rather than plain next-token prediction. Meta pruned Llama 3.1 8B and distilled logits from 8B/70B to make Llama 3.2 1B/3B, then co-distilled Llama 4 Maverick from Behemoth. Qwen3, Ministral 3, Apple's on-device 3B, NVIDIA's Minitron/Nemotron Nano and DeepSeek's R1-Distill series all document the same pattern in their technical reports.

   Sources arxiv.org · ai.meta.com · arxiv.org · arxiv.org

2. Labs report distillation beats RL on cost and often on quality

   Qwen3's ablation shows on-policy distillation lifting Qwen3-8B to 74.4 on AIME'24 versus 67.6 for RL, using 1,800 rather than 17,920 GPU hours. DeepSeek reports that R1-Distill-Qwen-32B (72.6 AIME'24) beats OpenAI o1-mini (63.6) with SFT-only distillation and no RL stage. NVIDIA reports up to 40x fewer training tokens and Mistral reports Ministral 3 trained on 1-3T tokens versus 15-36T for comparable Qwen 3 / Llama 3 models.

   Sources arxiv.org · huggingface.co · developer.nvidia.com · deeplearning.ai

3. Closed labs sell distillation, but only inside their own model family

   OpenAI's Model Distillation (stored completions + evals + fine-tuning, Oct 2024) lets customers distill GPT-4o/o1-preview into GPT-4o mini; Azure mirrors it; Amazon Bedrock requires teacher and student to be from the same model family, so Nova Premier distills into Nova Pro/Lite/Micro, Claude 3.5 Sonnet v2 into Claude 3 Haiku, and Llama 3.3 70B / Llama 3.1 405B into Llama 3.2 1B/3B and Llama 3.1 70B/8B; Google's Vertex early-access service distills Gemini 3.1 Pro into Gemini 2.5 Flash. In every case the student must be a model the vendor hosts, so distillation revenue stays on-platform. Notably OpenAI's docs now say it is 'winding down the fine-tuning platform' for new users, and Azure retires stored completions on 2026-10-15.

   Sources infoworld.com · aws.amazon.com · aws.amazon.com · developers.openai.com · learn.microsoft.com

4. Terms of service, not copyright, are the main legal lever against cross-lab distillation

   OpenAI forbids using 'Output to develop models that compete with OpenAI'; Anthropic's commercial terms bar access 'to build a competing product or service, including to train competing AI models'; Google's Gemini API terms say 'You may not use the Services to develop models that compete with the Services'; xAI's terms list 'distilling' among prohibited acts; Cohere bars use 'for the purpose of building a similar or competitive product or service'. Meta's Llama 4 license takes the opposite approach: derivative models are allowed but must carry 'Llama' at the start of their name.

   Sources openai.com · anthropic.com · ai.google.dev · developer.meta.com

5. Open-weight labs explicitly invite distillation in their licenses

   DeepSeek-R1's model card states the series 'allow for any modifications and derivative works, including, but not limited to, distillation for training other LLMs' under MIT. Qwen3, Mistral 3/Ministral 3 and SmolLM3 ship under Apache 2.0; Kimi K2 and MiniMax M2 use a modified MIT that only adds an attribution requirement above 100M MAU or $20M monthly revenue; NVIDIA releases Nemotron under its Open Model License and even lists the teacher models (DeepSeek-R1, GPT-OSS-120B, Qwen) used to synthesise 3.5T of its 10.6T pre-training tokens.

   Sources huggingface.co · huggingface.co · huggingface.co · mistral.ai

6. The accusation cycle escalated from one lab to a US government policy in 15 months

   OpenAI first alleged DeepSeek distillation in January 2025; the House Select Committee on the CCP's April 2025 report called it 'highly likely'. In February 2026 OpenAI told the committee DeepSeek used 'obfuscated third-party routers' and Anthropic published per-lab exchange counts (DeepSeek 150K, Moonshot 3.4M, MiniMax 13M). The framing is not purely US-versus-China: on 30 April 2026 Elon Musk conceded under oath in Musk v. OpenAI that xAI had 'partly' used OpenAI's technology to train its models. By April 2026 OpenAI, Anthropic and Google were also sharing threat intelligence through the Frontier Model Forum and the White House issued NSTM-4 calling the campaigns 'deliberate, industrial-scale'. In July 2026 OSTP director Kratsios accused Moonshot of distilling Anthropic's Fable model to build Kimi K3.

   Sources assets.bwbx.io · anthropic.com · nextgov.com · cyberscoop.com · forbesafrica.com

7. Defensive measures now shape products: ownership bans, hidden fingerprints, CoT hiding

   Anthropic barred entities more than 50% owned by companies in unsupported regions in September 2025, citing that they 'could also potentially use our models to advance their own AI development through techniques like distillation'. OpenAI's memo describes classifiers for 'reinforcement learning-style grading behavior', models 'trained not to reveal reasoning traces', and account bans. Anthropic admitted a March 2026 Claude Code 'experiment' that embedded identifying markers to protect against distillation; Alibaba responded by banning Claude Code for staff from 10 July 2026.

   Sources anthropic.com · assets.bwbx.io · techcrunch.com

8. Accused labs have not answered, and an illicit reseller market has emerged

   As of the Feb 2026 reports, DeepSeek, Moonshot and MiniMax had not responded to Anthropic's allegations; Moonshot did not respond to the July 2026 K3 claim. OpenAI's memo says Chinese companies 'rely on networks of unauthorized resellers of OpenAI's services to evade our platform's controls', and on 3 September 2026 Anthropic's head of threat intelligence Jacob Klein described 'an entire illicit ecosystem' on the dark web spinning up accounts at scale. Meanwhile Qwen and Zhipu were conspicuously absent from Anthropic's February list, before Alibaba was named in June.

   Sources latent.space · cnbc.com · assets.bwbx.io

Charts · 7 charts

Tables · 7 tables

Timeline · 40 events

1. 14 May 2024 product

   Google says Gemini 1.5 Flash was distilled from 1.5 Pro

   Google I/O blog: Flash was 'trained by 1.5 Pro through a process called distillation'. First frontier lab to publicly label a production tier as distilled.

   Source: blog.google

2. 31 July 2024 research

   Gemma 2 2B/9B trained with knowledge distillation

   Gemma 2 report: 'We also train the 2B and 9B models with knowledge distillation (Hinton et al., 2015) instead of next token prediction.'

   Source: arxiv.org

3. 1 August 2024 product

   Arcee AI open-sources DistillKit

   Logit-based and hidden-state distillation toolkit; 1.5B student taught by 7B Arcee-Agent.

   Source: arcee.ai

4. 14 August 2024 research

   NVIDIA publishes Minitron prune-and-distill recipe

   Llama-3.1-Minitron 4B from Llama 3.1 8B using 94B tokens; up to 40x fewer training tokens, 2.7x throughput with TensorRT-LLM.

   Source: developer.nvidia.com

5. 25 September 2024 product

   Meta releases Llama 3.2 1B/3B via pruning + logit distillation

   Logits from Llama 3.1 8B and 70B 'were used as token-level targets' after single-shot structured pruning of the 8B.

   Source: ai.meta.com

6. 1 October 2024 product

   OpenAI launches Model Distillation in the API

   Stored Completions, Evals (beta) and fine-tuning to distill GPT-4o / o1-preview into GPT-4o mini; available to all developers.

   Source: infoworld.com

7. 3 December 2024 product

   Amazon announces Bedrock Model Distillation (preview) alongside Nova

   Teacher-student pairs within the Nova, Claude and Llama families (teacher and student must be from the same family); promises up to 500% faster, 75% cheaper, <2% accuracy loss.

   Source: aws.amazon.com

8. 12 December 2024 research

   Microsoft Phi-4 report: earlier Phi models 'largely distill' GPT-4

   14B model on synthetic data; 'phi-4 substantially surpasses its teacher model on STEM-focused QA'.

   Source: arxiv.org

9. January 2025 legal

   OpenAI alleges DeepSeek distilled its models

   OpenAI says it has evidence DeepSeek queried its models at scale to train competitors in breach of ToS; Microsoft investigates API exfiltration.

   Source: sites.law.berkeley.edu

10. January 2025 research

    Hugging Face launches Open R1

    Fully open reproduction of the R1 distillation pipeline; OpenR1-Math-220k released Feb 2025.

    Source: github.com

11. 20 January 2025 product

    DeepSeek releases R1 plus six R1-Distill models under MIT

    ~800K R1 samples SFT'd into Qwen 1.5B-32B and Llama 8B/70B; license expressly allows 'distillation for training other LLMs'. Released 20 January 2025; the accompanying paper (arXiv 2501.12948) was submitted 22 January 2025.

    Source: arxiv.org

12. 12 March 2025 research

    Gemma 3: all sizes trained with distillation

    256 sampled logits per token; 1B/4B/12B/27B on 2T/4T/12T/14T tokens; post-training distills from 'a large IT teacher'.

    Source: arxiv.org

13. April 2025 policy

    House Select Committee report: DeepSeek 'highly likely' used unlawful distillation

    Committee report on DeepSeek cites OpenAI's account of DeepSeek employee accounts; R1 base model often self-identifies as ChatGPT.

    Source: fdd.org

14. 5 April 2025 product

    Meta codistills Llama 4 Maverick from Behemoth; license requires 'Llama' naming

    Behemoth ~2T total / 288B active; 'novel distillation loss function that dynamically weights the soft and hard targets'.

    Source: ai.meta.com

15. 30 April 2025 research

    Phi-4-reasoning trained on o3-mini reasoning traces

    Microsoft's 14B reasoning model uses SFT on demonstrations generated by OpenAI's o3-mini.

    Source: arxiv.org

16. 1 May 2025 product

    Amazon Bedrock Model Distillation reaches GA

    Adds Nova Premier (teacher) -> Nova Pro, Claude 3.5 Sonnet v2 and Llama 3.3 70B teachers, Llama 3.2 1B/3B students.

    Source: aws.amazon.com

17. 14 May 2025 research

    Qwen3 report formalises 'strong-to-weak distillation'

    Off-policy then on-policy distillation from Qwen3-235B-A22B/32B; beats RL with 1/10 the GPU hours.

    Source: arxiv.org

18. 3 June 2025 market

    Researchers claim DeepSeek R1-0528 resembles Gemini 2.5 Pro

    Stylistic and trace similarities flagged by EQ-Bench and SpeechMap developers; not conclusive.

    Source: winbuzzer.com

19. 17 June 2025 product

    Gemini 2.5 Flash / Flash-Lite ship; report confirms distillation

    'The smaller models in the Gemini 2.5 series — Flash size and below — use distillation' with a k-sparse teacher distribution.

    Source: arxiv.org

20. July 2025 research

    Apple reports distilling its on-device 3B model from a 64-expert MoE

    Last 10% of tokens (~1.4T) retrained with a distillation loss; teacher training cost cut 90%; no structural pruning needed.

    Source: arxiv.org

21. 8 July 2025 product

    Hugging Face SmolLM3 uses Qwen3-32B synthetic reasoning traces

    3B model, 11.2T tokens; reasoning SFT data generated by prompting Qwen3-32B.

    Source: huggingface.co

22. 20 August 2025 research

    NVIDIA Nemotron Nano 2: 12B pruned and distilled to 9B

    Minitron strategy on a 20T-token hybrid Mamba-Transformer; up to 6x throughput vs Qwen3-8B.

    Source: arxiv.org

23. 4 September 2025 policy

    Anthropic bars Chinese-controlled entities, citing distillation risk

    Entities >50% owned by companies in unsupported regions blocked worldwide; 'could also potentially use our models to advance their own AI development through techniques like distillation'.

    Source: anthropic.com

24. 19 September 2025 product

    xAI ships Grok 4 Fast, framed as RL not distillation

    $0.20/$0.50 per 1M tokens; '98% reduction in price' vs Grok 4; '40% fewer thinking tokens'.

    Source: x.ai

25. 15 October 2025 product

    Anthropic releases Claude Haiku 4.5 at $1/$5

    'one-third the cost and more than twice the speed' of Sonnet 4; training method undisclosed.

    Source: anthropic.com

26. 2 December 2025 product

    Mistral 3: Ministral 3 built by cascade distillation from Mistral Small 3.1

    3B/8B/14B under Apache 2.0; 1-3T training tokens vs 15-36T for Qwen 3 / Llama 3 peers.

    Source: mistral.ai

27. 15 December 2025 product

    NVIDIA Nemotron 3 Nano lists DeepSeek-R1, GPT-OSS-120B and Qwen as synthetic-data teachers

    ~3.5T of 10.6T pre-training tokens distilled from other labs' open models.

    Source: huggingface.co

28. 12 February 2026 policy

    OpenAI memo to House Select Committee on DeepSeek's 'obfuscated' distillation

    Alleges DeepSeek employee accounts used 'obfuscated third-party routers' and 'code to access US AI models and obtain outputs for distillation'; describes CoT-hiding and RL-grading classifiers.

    Source: assets.bwbx.io

29. 23 February 2026 legal

    Anthropic: 24,000 fraudulent accounts, 16M+ exchanges by DeepSeek, Moonshot, MiniMax

    DeepSeek 150K+, Moonshot 3.4M+, MiniMax 13M+ exchanges; MiniMax pivoted half its traffic to a new Claude model within 24h.

    Source: anthropic.com

30. 7 April 2026 market

    OpenAI, Anthropic and Google agree to share distillation threat intel via Frontier Model Forum

    Bloomberg reports US officials estimate unauthorised distillation costs labs 'billions of dollars' annually.

    Source: techbrew.com

31. 23 April 2026 policy

    White House memorandum NSTM-4 on 'industrial-scale' distillation

    OSTP director Kratsios: foreign entities 'principally based in China' run 'deliberate, industrial-scale campaigns'; commits to intel sharing and accountability measures.

    Source: nextgov.com

32. 24 April 2026 product

    DeepSeek V4 preview released despite allegations

    V4-Flash followed 2026-07-31 and V4-Pro 2026-08-13 (per Wikipedia summary; primary release notes not fetched).

    Source: en.wikipedia.org

33. 30 April 2026 legal

    Musk concedes under oath that xAI 'partly' used OpenAI technology

    Cross-examined in Musk v. OpenAI, Elon Musk said xAI had partly used OpenAI's technology to train its models and that 'Generally A.I. companies distill other A.I. companies' - the first sworn distillation admission by a US lab.

    Source: forbesafrica.com

34. 10 June 2026 legal

    Anthropic tells US Senate Alibaba ran 'the largest known distillation attack'

    Letter of 10 June 2026 to Senate Banking Committee Chair Tim Scott and Ranking Member Elizabeth Warren: ~25,000 fraudulent accounts and 28.8M interactions between 22 April and 5 June 2026 attributed to Alibaba's Qwen effort; first reported by CNBC on 2026-06-24.

    Source: cnbc.com

35. 4 July 2026 market

    Anthropic admits Claude Code anti-distillation fingerprinting 'experiment'

    Engineer Thariq Shihipar: markers added in March 'to prevent account abuse from unauthorized resellers and protect against distillation'.

    Source: techcrunch.com

36. 10 July 2026 market

    Alibaba bans employees from Claude Code

    Classified as high-risk software; staff told to uninstall Anthropic models and use Alibaba's Qoder.

    Source: techcrunch.com

37. 22 July 2026 policy

    White House accuses Moonshot of distilling Anthropic's Fable into Kimi K3

    Kratsios says Moonshot used 'a sophisticated internal platform to conduct large scale distillation' and GB300 servers; K3 billed as 'first open 2.8 trillion parameter model'.

    Source: cyberscoop.com

38. 23 July 2026 product

    Google Cloud documents Gemini 3.1 Pro -> 2.5 Flash distillation service (pre-GA)

    Allowlist-only, no production use, 30-day initial access; conflicting model IDs suggest not yet a shippable product.

    Source: runtimewire.com

39. 3 September 2026 market

    Anthropic: distillation fight moves to the dark web

    Head of threat intelligence Jacob Klein describes 'an entire illicit ecosystem' spinning up accounts at extreme scale to resell Claude access.

    Source: cnbc.com

40. 15 October 2026 product

    Azure OpenAI stored completions (distillation input) scheduled to retire

    Microsoft Learn directs users to migrate to the Responses API and Agent Traces.

    Source: learn.microsoft.com

Glossary · 14 terms

Knowledge distillation (KD)

Training a smaller student model to match a larger teacher's output distribution (soft targets / logits) rather than only ground-truth labels; introduced by Hinton et al. 2015 and cited by Gemma 2, Gemma 3 and Llama 3.2 reports.

Strong-to-weak distillation

Qwen3's term for training its 0.6B-30B models from Qwen3-235B-A22B / Qwen3-32B teachers, first off-policy (on teacher outputs) then on-policy (student samples scored by teacher logits).

On-policy distillation

The student generates its own sequences and is corrected by the teacher's token-level distribution, fixing train/inference mismatch (Agarwal et al. GKD; implemented in Hugging Face TRL's GKDTrainer).

Codistillation

Meta's term for training Llama 4 Maverick and Scout concurrently with the Behemoth teacher, using a loss that dynamically weights soft and hard targets.

Cascade distillation

Mistral's iterative loop of pruning a parent model (Mistral Small 3.1) and continuing training with a distillation loss to produce Ministral 3 3B/8B/14B.

Prune-and-distill (Minitron)

NVIDIA's recipe: structured width/depth pruning of a trained model followed by KD to recover accuracy with far fewer tokens (e.g. 94B vs 15T).

k-sparse teacher distribution

Google's approximation in Gemini 2.5 that stores only the top-k teacher probabilities per token to make distillation storage tractable.

Stored completions

OpenAI / Azure feature that logs production prompt-response pairs (store=true) so they can be filtered into a fine-tuning dataset for a smaller model.

Distillation attack

Industry term (Anthropic, OpenAI, Google, White House) for unauthorised large-scale querying of a proprietary model, often via fraudulent accounts or proxies, to harvest outputs for training a competitor.

Chain-of-thought (CoT) extraction

Eliciting a model's reasoning traces to use as training targets; OpenAI now trains models 'not to reveal reasoning traces' and Google summarises them to blunt this.

Anti-competing-model clause

ToS language (OpenAI, Anthropic, Google, xAI, Cohere) forbidding use of outputs or services to build competing models; the primary legal basis for distillation disputes.

Modified MIT license

Open-weight license used by Moonshot (Kimi K2) and MiniMax that keeps MIT permissions but adds an attribution requirement above 100M MAU or $20M monthly revenue.

Frontier Model Forum (FMF)

Non-profit founded by OpenAI, Anthropic, Google and Microsoft in 2023, used from April 2026 to share distillation-attack intelligence.

NSTM-4

April 2026 White House national security memorandum committing the US government to share intelligence on foreign distillation campaigns and explore accountability measures.

Sources · 75 sources

Every figure on this page comes from one of these primary sources. Compiled 3 September 2026.

1. Model Distillation in the APIOpenAI · 1 October 2024 · blog
2. OpenAI updates API with model distillation, prompt caching abilitiesInfoWorld · 3 October 2024 · news
3. OpenAI Terms of Use (Rest of World)OpenAI · 2025 · law
4. Be Careful With OpenAI's Terms of UseOSPOCO · 2024 · blog
5. OpenAI API pricingOpenAI · September 2026 · pricing
6. Supervised fine-tuning: distilling from a larger modelOpenAI · 2026 · docs
7. Memo to US House Select Committee: Updated Stakes for American-Led, Democratic AIOpenAI (via Bloomberg) · 12 February 2026 · filing
8. OpenAI Accuses China's DeepSeek of Distilling US AI Models to Gain an EdgeBloomberg · 12 February 2026 · news
9. OpenAI accuses DeepSeek of malpractice ahead of AI launchRest of World · 12 February 2026 · news
10. The Innovation Dilemma: AI Distillation in OpenAI v. DeepSeekBerkeley Law · 30 March 2025 · blog
11. OpenAI Alleges China's DeepSeek Stole its IP to Train its Own ModelsFDD · 13 February 2026 · news
12. Detecting and preventing distillation attacksAnthropic · 23 February 2026 · blog
13. Anthropic accuses Chinese labs of trying to illicitly take Claude's capabilitiesCyberScoop · 23 February 2026 · news
14. Updating restrictions of sales to unsupported regionsAnthropic · 4 September 2025 · blog
15. Anthropic tightens AI access rules, targeting Chinese-controlled entitiesCRN Asia · 5 September 2025 · news
16. Anthropic Commercial Terms of ServiceAnthropic · 17 June 2025 · law
17. Anthropic Consumer Terms of ServiceAnthropic · 8 October 2025 · law
18. Introducing Claude Haiku 4.5Anthropic · 15 October 2025 · blog
19. Anthropic accuses Alibaba of campaign to 'brazenly' and 'illicitly' distill ClaudeCNBC · 24 June 2026 · news
20. Alibaba reportedly bans employees from using Claude CodeTechCrunch · 4 July 2026 · news
21. White House official accuses Chinese startup of distilling Anthropic's modelCyberScoop · 22 July 2026 · news
22. Anthropic's distillation battle turns to the dark web as China concerns swellCNBC · 3 September 2026 · news
23. White House accuses China of 'deliberate, industrial-scale' campaigns to steal US AI modelsNextgov · 23 April 2026 · news
24. White House accuses China of industrial-scale theft of US AI frontier modelsInteresting Engineering · 23 April 2026 · news
25. OpenAI, Anthropic, Google join forces against ChinaTech Brew (citing Bloomberg) · 7 April 2026 · news
26. Gemini 2.5: Pushing the Frontier with Advanced ReasoningGoogle DeepMind · July 2025 · paper
27. Gemini 1.5 Flash announcement (Google I/O 2024)Google · 14 May 2024 · blog
28. Gemma 2: Improving Open Language Models at a Practical SizeGoogle DeepMind · 31 July 2024 · paper
29. Gemma 3 Technical ReportGoogle DeepMind · 12 March 2025 · paper
30. Gemini API Additional Terms of ServiceGoogle · 28 April 2026 · law
31. Google Cloud page describes Gemini distillation service, but its release status is unclearRuntimeWire · 28 July 2026 · news
32. Google says Gemini was hit with 100,000 prompts in apparent cloning attemptNBC News · February 2026 · news
33. Is DeepSeek Training its AI with Data from Google Gemini?WinBuzzer · 3 June 2025 · news
34. Llama 3.2: Revolutionizing edge AI and vision with open, customizable modelsMeta · 25 September 2024 · blog
35. The Llama 4 herd: The beginning of a new era of natively multimodal AI innovationMeta · 5 April 2025 · blog
36. Llama 4 Community License AgreementMeta · 5 April 2025 · law
37. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement LearningDeepSeek · 22 January 2025 · paper
38. DeepSeek-R1 model card and licenseDeepSeek / Hugging Face · January 2025 · docs
39. DeepSeek (Wikipedia) - V4 release datesWikipedia · 2026 · docs
40. Qwen3 Technical ReportAlibaba Qwen Team · 14 May 2025 · paper
41. Qwen3-235B-A22B model cardAlibaba / Hugging Face · May 2025 · docs
42. Phi-4 Technical ReportMicrosoft · 12 December 2024 · paper
43. Phi-4-reasoning Technical ReportMicrosoft · 30 April 2025 · paper
44. Phi-4-mini-instruct model cardMicrosoft / Hugging Face · February 2025 · docs
45. How to use Azure OpenAI stored completions & distillationMicrosoft Learn · 6 July 2026 · docs
46. Introducing Model Distillation in Azure OpenAI ServiceMicrosoft · 2024 · blog
47. How to Prune and Distill Llama-3.1 8B to an NVIDIA Llama-3.1-Minitron 4B ModelNVIDIA · 14 August 2024 · blog
48. LLM Pruning and Distillation in Practice: The Minitron ApproachNVIDIA · August 2024 · paper
49. NVIDIA Nemotron Nano 2NVIDIA · 20 August 2025 · paper
50. NVIDIA-Nemotron-3-Nano-30B-A3B model cardNVIDIA / Hugging Face · 15 December 2025 · docs
51. Amazon Bedrock Model Distillation (preview)AWS · 3 December 2024 · blog
52. Amazon Bedrock Model Distillation is now generally availableAWS · 1 May 2025 · docs
53. The Amazon Nova Family of Models: Technical Report and Model CardAmazon · 3 December 2024 · paper
54. Introducing Mistral 3Mistral AI · 2 December 2025 · blog
55. Ministral 3 (paper)Mistral AI · 13 January 2026 · paper
56. Mistral Uses Cascade Distillation on Mistral 3 To Build Ministral FamilyDeepLearning.AI The Batch · 6 February 2026 · news
57. Generalized Knowledge Distillation TrainerHugging Face · 2026 · docs
58. Open R1: fully open reproduction of DeepSeek-R1Hugging Face · January 2025 · docs
59. SmolLM3: smol, multilingual, long-context reasonerHugging Face · 8 July 2025 · blog
60. Announcing DistillKitArcee AI · 1 August 2024 · blog
61. arcee-ai/DistillKitArcee AI · 2026 · docs
62. Grok 4 FastxAI · 19 September 2025 · blog
63. xAI Terms of Service - ConsumerxAI · undisclosed · law
64. xAI Acceptable Use PolicyxAI · undisclosed · law
65. Musk admits distilling OpenAI data for his xAIForbes Africa · 1 May 2026 · news
66. Kimi-K2-Instruct model cardMoonshot AI / Hugging Face · July 2025 · docs
67. Kimi K2 Modified MIT LicenseMoonshot AI · July 2025 · law
68. MiniMax-M2 model cardMiniMax / Hugging Face · 2025 · docs
69. GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation ModelsZhipu AI / Z.ai · 8 August 2025 · paper
70. GLM-4.5 model card and licenseZhipu AI / Hugging Face · July 2025 · docs
71. AINews: Anthropic accuses DeepSeek, Moonshot, and MiniMaxLatent Space · 23 February 2026 · news
72. Apple Intelligence Foundation Language Models: Tech Report 2025Apple · July 2025 · paper
73. Command A: An Enterprise-Ready Large Language ModelCohere · 1 April 2025 · paper
74. Cohere Terms of UseCohere · 7 September 2022 · law
75. C4AI Command R7B model cardCohere Labs / Hugging Face · December 2024 · docs