{"ontology":"https://layer.re/machine/schema.json","epistemic_model":"evidence-backed-temporal-layer-graph","canonical":true,"type":"layer_history","layer":{"id":"layer-models","slug":"models","name":"Models","description":"Trained computational models and their capabilities.","machine_description":"Trained computational models and their capabilities.","active":1},"snapshots":[{"id":"state-models-2022q1","layer_id":"layer-models","as_of":"2022-03-31 23:59:59.000","summary":"Large language models were moving beyond raw next-token pretraining through human-feedback post-training, prompting methods and compute-efficient scaling research.","dependencies":["layer-research","layer-compute","layer-chips","layer-software","layer-human"],"constraints":["training compute","data allocation","reliability and truthfulness","deployment cost"],"emerging_capabilities":["instruction following","reasoning elicitation by prompting","code generation"],"underdeveloped_components":["robust general reasoning","reliable long-horizon planning"],"uncertainties":["later importance of these methods is retrospective; contemporaneous adoption varied"],"confidence":0.95,"previous_snapshot_id":null,"mapping_mode":"retrospective","reconstructed_at":"2026-10-01T17:34:32.000Z","created_at":"2026-10-01T18:05:36.388Z","updated_at":"2026-10-01T18:05:36.388Z"},{"id":"state-models-2022q2","layer_id":"layer-models","as_of":"2022-06-30 23:59:59.000","summary":"Large models in Q2 became more general across tasks and modalities, while capability increasingly depended on training system scale, specialized data, prompt/inference strategies and controlled deployment rather than parameter count alone.","dependencies":["layer-research","layer-compute","layer-chips","layer-software","layer-human"],"constraints":["training and inference cost","unreliable reasoning","limited interpretability","capability forecasting uncertainty"],"emerging_capabilities":["multimodal few-shot learning","stronger quantitative reasoning","broader code and language capability"],"underdeveloped_components":["verified reasoning","predictable scaling behavior","broad low-cost frontier inference"],"uncertainties":["benchmark performance did not establish robust real-world generality","emergence interpretations were method-sensitive"],"confidence":0.96,"previous_snapshot_id":"state-models-2022q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.761Z","updated_at":"2026-10-01T19:22:17.761Z"},{"id":"state-models-2022q3","layer_id":"layer-models","as_of":"2022-09-30 23:59:59.000","summary":"Models in Q3 became more distributable across language, image, speech and translation while research systems increasingly composed model generation with retrieval, memory, human feedback and executable code.","dependencies":["layer-research","layer-compute","layer-software","layer-open-source","layer-human"],"constraints":["inference cost","hardware requirements","hallucination","license heterogeneity","verification"],"emerging_capabilities":["downloadable 176B multilingual language models","open text-to-image weights","open multilingual speech recognition","retrieval-supported dialogue","generated robot policy code"],"underdeveloped_components":["reliable factuality","cheap frontier inference","general long-horizon autonomy"],"uncertainties":["'open' varied substantially by license and hardware burden","research benchmarks did not guarantee deployment robustness"],"confidence":0.96,"previous_snapshot_id":"state-models-2022q2","mapping_mode":"retrospective","reconstructed_at":"2026-10-02T15:12:29.000Z","created_at":"2026-10-02T16:08:16.256Z","updated_at":"2026-10-02T16:08:16.256Z"},{"id":"state-models-2022q4","layer_id":"layer-models","as_of":"2022-12-31 23:59:59.000","summary":"Model capability in Q4 was increasingly shaped after pretraining through scaled instruction finetuning, human-feedback dialogue training and experimental AI-feedback supervision, while external-action research linked language-model reasoning to environments and tools.","dependencies":["layer-research","layer-compute","layer-software","layer-human","layer-governance"],"constraints":["hallucination","prompt sensitivity","training and inference cost","evaluation validity","behavior-control tradeoffs"],"emerging_capabilities":["scaled instruction finetuning","RLHF dialogue assistants","interleaved reasoning and acting","principle-guided self-critique and AI feedback"],"underdeveloped_components":["reliable truthfulness","cheap frontier inference","robust long-horizon action"],"uncertainties":["ChatGPT underlying GPT-3.5 precursor predates Q4","benchmark gains do not establish broad reliability"],"confidence":0.97,"previous_snapshot_id":"state-models-2022q3","mapping_mode":"retrospective","reconstructed_at":"2026-10-03T14:26:11.000Z","created_at":"2026-10-03T14:45:10.924Z","updated_at":"2026-10-03T14:45:10.924Z"},{"id":"state-models-2023q1","layer_id":"layer-models","as_of":"2023-03-31 23:59:59.000","summary":"Q1 combined a frontier capability step in GPT-4 with broader assistant-model competition, smaller-model research access and increasing tool-use integration; public GPT-4 vision was not generally available.","dependencies":["layer-research","layer-compute","layer-software","layer-human"],"constraints":["inference cost","hallucination","access gating","evaluation opacity"],"emerging_capabilities":["GPT-4 text capability","assistant-model competition","smaller LLaMA research models","tool-use research"],"underdeveloped_components":["reliable truthfulness","cheap frontier inference","general public multimodality","persistent autonomy"],"uncertainties":["Vendor evaluations are not independent proof; LLaMA access was restricted."],"confidence":0.97,"previous_snapshot_id":"state-models-2022q4","mapping_mode":"retrospective","reconstructed_at":"2026-10-05T15:23:51.000Z","created_at":"2026-10-05T18:17:53.800Z","updated_at":"2026-10-05T18:17:53.800Z"},{"id":"state-models-2023q2","layer_id":"layer-models","as_of":"2023-06-30 23:59:59.000","summary":"Models broadened through new provider families, document-scale context, open foundation models and structured function-call interfaces for larger systems.","dependencies":["layer-research","layer-software","layer-compute","layer-open-source"],"constraints":["context reliability","tool-call correctness","inference cost"],"emerging_capabilities":["100K-class context","structured function calling","open foundation models"],"underdeveloped_components":["reliable long-context reasoning","autonomous execution"],"uncertainties":["vendor benchmark claims remain noncanonical"],"confidence":0.97,"previous_snapshot_id":"state-models-2023q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-07T12:55:58.000Z","created_at":"2026-10-07T14:04:48.609Z","updated_at":"2026-10-07T14:04:48.609Z"},{"id":"state-models-2023q3","layer_id":"layer-models","as_of":"2023-09-30 23:59:59.000","summary":"Q3 models became more customizable, commercially distributable and multimodally accessible through self-service fine-tuning, Llama 2/Mistral releases and the beginning of public GPT-4 image interaction.","dependencies":["layer-research","layer-software","layer-compute","layer-open-source","layer-human"],"constraints":["fine-tuning quality","license heterogeneity","inference cost","multimodal reliability","evaluation opacity"],"emerging_capabilities":["self-service chat-model fine-tuning","commercial open-weight distribution","Apache-2.0 model release","public image-input rollout"],"underdeveloped_components":["uniform multimodal access","cheap frontier inference","reliable customization"],"uncertainties":["Llama 2 was community-licensed rather than unrestricted open source; public GPT-4 image rollout was incomplete"],"confidence":0.97,"previous_snapshot_id":"state-models-2023q2","mapping_mode":"retrospective","reconstructed_at":"2026-10-07T16:04:51.000Z","created_at":"2026-10-07T16:56:16.865Z","updated_at":"2026-10-07T16:56:16.865Z"},{"id":"state-models-2023q4","layer_id":"layer-models","as_of":"2023-12-31 23:59:59.000","summary":"Models combined longer context, multimodal interfaces, tool use, multiple provider families and open MoE architectures while model-specific governance became more explicit.","dependencies":["layer-research","layer-compute","layer-software","layer-governance","layer-open-source"],"constraints":["context reliability","tool reliability","compute cost","access gating"],"emerging_capabilities":["GPT-4 Turbo 128K preview","Claude 2.1 200K","Gemini Pro/Nano","Mixtral MoE"],"underdeveloped_components":["reliable long-context reasoning","uniform multimodal access"],"uncertainties":["GPT-4 Turbo was preview and Gemini Ultra was not generally available"],"confidence":0.98,"previous_snapshot_id":"state-models-2023q3","mapping_mode":"retrospective","reconstructed_at":"2026-10-08T15:50:00.000Z","created_at":"2026-10-08T17:00:00.129Z","updated_at":"2026-10-08T17:00:00.129Z"},{"id":"state-models-2024q1","layer_id":"layer-models","as_of":"2024-03-31 23:59:59.000","summary":"Vision-capable assistants broadened across providers; selected million-token multimodal experiments and video-generation research expanded the model interface beyond Q4's available products.","dependencies":[],"constraints":["availability scope","reliability and implementation limits"],"emerging_capabilities":["Gemini Advanced exposes Ultra 1.0 to subscribers","Sora research demonstrates text-conditioned video generation","Gemini 1.5 Pro enters limited million-token preview","Claude 3 Opus and Sonnet become available with vision"],"underdeveloped_components":["generalization beyond the selected evidence"],"uncertainties":["Access is plan and market bounded; it does not demonstrate widespread adoption.","The research report does not establish general availability, accurate physics or an executable game-state engine.","Experimental long context is not general availability or reliable reasoning over every token.","Haiku was announced for later availability; benchmark leadership and near-human intelligence remain vendor claims."],"confidence":0.93,"previous_snapshot_id":"state-models-2023q4","mapping_mode":"retrospective","reconstructed_at":"2026-10-09T19:12:47.000Z","created_at":"2026-10-09T21:04:17.183Z","updated_at":"2026-10-09T21:04:17.183Z"},{"id":"state-models-2024q2","layer_id":"layer-models","as_of":"2024-06-30 23:59:59.000","summary":"Multimodal text/vision and long-context interfaces gained broader product and GA API access, with live audio/video rollout still separate.","dependencies":[],"constraints":["availability scope","reliability and implementation limits"],"emerging_capabilities":["Llama 3 weights release and Meta AI expands","GPT-4o text and vision roll out","Gemini 1.5 Pro and Flash API releases become GA","Claude 3.5 Sonnet becomes available on Amazon Bedrock","Gemini API adds code execution and two-million-token GA context"],"underdeveloped_components":["generalization beyond the selected evidence"],"uncertainties":["Custom-licensed open weights are not unrestricted open source; planned multimodality is not released capability.","The announced native audio/video experience was not generally available at launch.","This date applies to the identified API releases, not all provider surfaces or the then-waitlisted two-million-token context.","Vendor benchmark claims remain claims; this is not the October computer-use release.","Code execution is a bounded API capability, not universal delegated authority or reliable autonomous work."],"confidence":0.93,"previous_snapshot_id":"state-models-2024q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-09T19:12:47.000Z","created_at":"2026-10-10T07:29:06.786Z","updated_at":"2026-10-10T07:29:06.786Z"},{"id":"state-models-2024q3","layer_id":"layer-models","as_of":"2024-09-30 23:59:59.000","summary":"Licensed open-weight scale, specialized families and restricted reasoning previews diversified model development; vendor parity claims remain unverified.","dependencies":[],"constraints":["availability scope","reliability and implementation limits"],"emerging_capabilities":["Llama 3.1 releases licensed 405B and smaller weights","FLUX.1 launches with distinct API and weight licenses","OpenAI releases o1 reasoning previews","Qwen2.5 releases general, coding and mathematics weights"],"underdeveloped_components":["generalization beyond the selected evidence"],"uncertainties":["Meta's frontier-parity evaluation is an actor claim; accessible weights do not mean unrestricted open source.","Licensing differs by variant; image quality superiority is not independently established here.","Next-update benchmark results are not measured capabilities of the released preview; launch lacked browsing and image upload.","License exceptions vary by model; vendor benchmark leadership is not an independently verified fact."],"confidence":0.93,"previous_snapshot_id":"state-models-2024q2","mapping_mode":"retrospective","reconstructed_at":"2026-10-09T19:12:47.000Z","created_at":"2026-10-10T07:49:15.919Z","updated_at":"2026-10-10T07:49:15.919Z"},{"id":"state-models-2024q4","layer_id":"layer-models","as_of":"2024-12-31 23:59:59.000","summary":"Reasoning, audio, video and experimental tool-capable releases diversified available model interfaces, each with distinct rollout boundaries.","dependencies":[],"constraints":["availability scope","reliability and implementation limits"],"emerging_capabilities":["Realtime API enters public beta","Claude computer use enters public beta","Sora video product begins rollout","Gemini 2.0 Flash releases experimentally","Full o1 API gains tool and vision interfaces","DeepSeek-V3 releases model and API"],"underdeveloped_components":["generalization beyond the selected evidence"],"uncertainties":["The beta is not GA or general autonomous voice operation.","Experimental computer use remains error-prone and requires developer integration and supervision.","Generated video is not reliable physics simulation or a controllable game engine.","Experimental model access is not GA; Project Mariner and Jules are research prototypes or limited trials.","The rollout was not immediately available to every API tier; tool calling does not authorize arbitrary actions.","Reported training efficiency and benchmark parity are vendor claims, not independent total-cost measurements."],"confidence":0.93,"previous_snapshot_id":"state-models-2024q3","mapping_mode":"retrospective","reconstructed_at":"2026-10-09T19:12:47.000Z","created_at":"2026-10-10T08:02:14.503Z","updated_at":"2026-10-10T08:02:14.503Z"},{"id":"state-models-2025q1","layer_id":"layer-models","as_of":"2025-03-31 23:59:59.000","summary":"Downloadable reasoning weights, hybrid extended thinking and compact multimodal releases widened model choices without independent frontier-parity claims.","dependencies":[],"constraints":["availability scope","reliability and implementation limits"],"emerging_capabilities":["DeepSeek-R1 releases reasoning weights","Claude 3.7 and Claude Code research preview launch","Gemma 3 releases compact multimodal model family"],"underdeveloped_components":["generalization beyond the selected evidence"],"uncertainties":["Vendor o1-parity and low-cost narratives are not independent total-cost evidence; distilled models retain upstream-license considerations.","Model availability differs from coding-agent preview; autonomous coding claims do not establish unsupervised reliability.","Gemma licensing differs from unrestricted open source; benchmark and single-accelerator efficiency remain vendor-scoped claims."],"confidence":0.93,"previous_snapshot_id":"state-models-2024q4","mapping_mode":"retrospective","reconstructed_at":"2026-10-09T19:12:47.000Z","created_at":"2026-10-10T08:11:02.813Z","updated_at":"2026-10-10T08:11:02.813Z"},{"id":"state-models-2025q2","layer_id":"layer-models","as_of":"2025-06-30 23:59:59.000","summary":"This period adds specific models evidence through Claude 4 releases and Claude Code reaches GA; Gemini Robotics On-Device is introduced. Maturity remains specific to the cited products and programs; this synthesis does not infer sector-wide adoption.","dependencies":[],"constraints":["Vendor evaluations and long-running demonstrations do not establish general unattended autonomy.","Selected-access demonstrations do not prove general commercial deployment or cross-embodiment reliability."],"emerging_capabilities":["Claude 4 releases and Claude Code reaches GA","Gemini Robotics On-Device is introduced"],"underdeveloped_components":["Independent measurement of deployment scope and outcomes"],"uncertainties":["The referenced historical predecessor belongs to the reviewed chronology; the live history must be checked before ingestion."],"confidence":0.9,"previous_snapshot_id":"state-models-2025q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-10T06:24:42.000Z","created_at":"2026-10-10T08:25:19.365Z","updated_at":"2026-10-10T08:25:19.365Z"},{"id":"state-models-2025q3","layer_id":"layer-models","as_of":"2025-09-30 23:59:59.000","summary":"This period adds specific models evidence through GPAI Code of Practice is published; OpenAI releases gpt-oss weights; Genie 3 interactive world-model research is introduced; GPT-5 developer models become available; Gemini Robotics 1.5 introduces physical-agent planning; Claude Sonnet 4.5 is released. Maturity remains specific to the cited products and programs; this synthesis does not infer sector-wide adoption.","dependencies":[],"constraints":["A voluntary compliance code is distinct from binding law and its later enforcement schedule.","Open weights do not imply publication of the full training data or complete reproducible training pipeline.","Generated video worlds do not establish semantically addressable persistent game state or a shipped game engine.","Provider benchmark results do not establish independent productivity gains or task reliability in every deployment.","API access to reasoning is distinct from unrestricted access to robot actuation models.","Long-horizon vendor demonstrations do not establish unattended success rates across real production tasks."],"emerging_capabilities":["GPAI Code of Practice is published","OpenAI releases gpt-oss weights","Genie 3 interactive world-model research is introduced","GPT-5 developer models become available","Gemini Robotics 1.5 introduces physical-agent planning","Claude Sonnet 4.5 is released"],"underdeveloped_components":["Independent measurement of deployment scope and outcomes"],"uncertainties":["The referenced historical predecessor belongs to the reviewed chronology; the live history must be checked before ingestion."],"confidence":0.9,"previous_snapshot_id":"state-models-2025q2","mapping_mode":"retrospective","reconstructed_at":"2026-10-10T06:24:42.000Z","created_at":"2026-10-10T08:31:22.793Z","updated_at":"2026-10-10T08:31:22.793Z"},{"id":"state-models-2025q4","layer_id":"layer-models","as_of":"2025-12-31 23:59:59.000","summary":"This period adds specific models evidence through Gemini 3 and Antigravity are introduced; GPT-5.2 releases across API and paid plans. Maturity remains specific to the cited products and programs; this synthesis does not infer sector-wide adoption.","dependencies":[],"constraints":["Preview access and provider evaluations do not prove dependable unsupervised delivery.","Provider evaluations do not independently establish general professional replacement or absence of failure modes."],"emerging_capabilities":["Gemini 3 and Antigravity are introduced","GPT-5.2 releases across API and paid plans"],"underdeveloped_components":["Independent measurement of deployment scope and outcomes"],"uncertainties":["The referenced historical predecessor belongs to the reviewed chronology; the live history must be checked before ingestion."],"confidence":0.9,"previous_snapshot_id":"state-models-2025q3","mapping_mode":"retrospective","reconstructed_at":"2026-10-10T06:24:42.000Z","created_at":"2026-10-10T08:37:10.929Z","updated_at":"2026-10-10T08:37:10.929Z"},{"id":"state-models-2026q1","layer_id":"layer-models","as_of":"2026-03-31 23:59:59.000","summary":"This period adds specific models evidence through Project Genie opens a limited interactive-world experiment; Claude Opus 4.6 is released; GPT-5.4 adds native computer use and tool search. Maturity remains specific to the cited products and programs; this synthesis does not infer sector-wide adoption.","dependencies":[],"constraints":["Bounded experimental access is not a general game platform or persistent semantic game state.","Million-token context is beta and agent teams are a research preview rather than unrestricted production automation.","Model capability does not grant permissions or independently establish safe end-to-end task completion."],"emerging_capabilities":["Project Genie opens a limited interactive-world experiment","Claude Opus 4.6 is released","GPT-5.4 adds native computer use and tool search"],"underdeveloped_components":["Independent measurement of deployment scope and outcomes"],"uncertainties":["The referenced historical predecessor belongs to the reviewed chronology; the live history must be checked before ingestion."],"confidence":0.9,"previous_snapshot_id":"state-models-2025q4","mapping_mode":"retrospective","reconstructed_at":"2026-10-10T06:24:42.000Z","created_at":"2026-10-10T08:45:32.710Z","updated_at":"2026-10-10T08:45:32.710Z"},{"id":"state-models-2026q2","layer_id":"layer-models","as_of":"2026-06-30 23:59:59.000","summary":"This period adds specific models evidence through Project Glasswing launches with restricted security-model access; Project Genie expands access and adds location-based worlds; Glasswing expands selected defensive access; GPT-5.6 family enters limited preview. Maturity remains specific to the cited products and programs; this synthesis does not infer sector-wide adoption.","dependencies":[],"constraints":["Partner access is restricted; October updates on the same page are excluded from April reconstruction.","Experimental location reconstruction is distinct from native game runtime state or production simulation fidelity.","Expanded eligibility remains controlled access and does not establish general availability of Mythos.","Preview benchmarks and harness results are company claims; July access is not backdated into June."],"emerging_capabilities":["Project Glasswing launches with restricted security-model access","Project Genie expands access and adds location-based worlds","Glasswing expands selected defensive access","GPT-5.6 family enters limited preview"],"underdeveloped_components":["Independent measurement of deployment scope and outcomes"],"uncertainties":["The referenced historical predecessor belongs to the reviewed chronology; the live history must be checked before ingestion."],"confidence":0.9,"previous_snapshot_id":"state-models-2026q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-10T06:24:42.000Z","created_at":"2026-10-10T09:06:48.254Z","updated_at":"2026-10-10T09:06:48.254Z"},{"id":"state-models-20260930-initial","layer_id":"layer-models","as_of":"2026-09-30 12:45:00.000","summary":"High-end model capabilities are diffusing into faster and cheaper tiers, improving the economic feasibility of sustained AI work.","dependencies":["layer-compute","layer-chips","layer-software"],"constraints":["inference cost","factual reliability","deployment safety"],"emerging_capabilities":["cheaper high-capability inference","better long-context caching","stronger coding and computer use"],"underdeveloped_components":["independent cross-provider evaluation"],"uncertainties":["real-world reliability outside vendor benchmarks"],"confidence":0.9,"previous_snapshot_id":null,"mapping_mode":"contemporaneous","reconstructed_at":null,"created_at":"2026-09-30T12:54:32.490Z","updated_at":"2026-09-30T12:54:32.490Z"},{"id":"state-models-20260930-late","layer_id":"layer-models","as_of":"2026-09-30 21:20:00.000","summary":"Frontier model competition widened with Google's Gemini 4 Argon announcement, while deployment of the most sensitive cybersecurity capabilities is becoming more explicitly phased and governed.","dependencies":["layer-compute","layer-chips","layer-software","layer-governance"],"constraints":["controlled access for high-risk capabilities","independent validation"],"emerging_capabilities":["long-horizon software engineering","frontier cybersecurity workflows"],"underdeveloped_components":["broad public availability evidence","cross-provider independent benchmarking"],"uncertainties":["real-world performance and public release timing"],"confidence":0.92,"previous_snapshot_id":"state-models-20260930-initial","mapping_mode":"contemporaneous","reconstructed_at":null,"created_at":"2026-09-30T21:57:03.549Z","updated_at":"2026-09-30T21:57:03.549Z"},{"id":"state-models-2026q3","layer_id":"layer-models","as_of":"2026-09-30 23:59:59.000","summary":"This period adds specific models evidence through GPT-5.6 expands to broader product and API access; Claude Opus 5 is released; Gemini Robotics 2 extends whole-body robot control; GPT-6 Astra begins limited organizational rollout; Google announces Gemini 4 Argon with restricted defender access; OpenAI and Synopsys agree to develop a chip-design model. Maturity remains specific to the cited products and programs; this synthesis does not infer sector-wide adoption.","dependencies":[],"constraints":["Broad product access remains plan-dependent and does not validate all vendor performance claims independently.","Performance and cost comparisons are vendor-reported, not universal production outcome measures.","Demonstrated embodiments and bounded access do not establish general commercial humanoid autonomy.","The September 3 entry explicitly says Astra is not yet generally available; later rollout must not be backdated.","Broader developer and consumer access is prospective; internal productivity and autonomous patching assertions remain provider claims.","The specialized model and bundled service are prospective; the agreement is not a shipped autonomous chip-design system."],"emerging_capabilities":["GPT-5.6 expands to broader product and API access","Claude Opus 5 is released","Gemini Robotics 2 extends whole-body robot control","GPT-6 Astra begins limited organizational rollout","Google announces Gemini 4 Argon with restricted defender access","OpenAI and Synopsys agree to develop a chip-design model"],"underdeveloped_components":["Independent measurement of deployment scope and outcomes"],"uncertainties":["The referenced historical predecessor belongs to the reviewed chronology; the live history must be checked before ingestion."],"confidence":0.9,"previous_snapshot_id":"state-models-2026q2","mapping_mode":"retrospective","reconstructed_at":"2026-10-10T06:24:42.000Z","created_at":"2026-10-10T09:18:22.070Z","updated_at":"2026-10-10T09:18:22.070Z"},{"id":"state-models-20261004-open-weight","layer_id":"layer-models","as_of":"2026-10-04 20:12:00.000","summary":"The open-weight model ecosystem gained a concrete European release through Kolibri, while Reflection signaled a forthcoming U.S. open-weight model tied to localized AI-factory deployment.","dependencies":["layer-compute","layer-software","layer-open-source"],"constraints":["benchmark verification","deployment compute","ecosystem adoption"],"emerging_capabilities":["European open-weight reasoning/tool-use model deployment","institution-controlled open-weight AI stacks"],"underdeveloped_components":["independent evaluation of Kolibri","released and reproducible Reflection model"],"uncertainties":["relative frontier performance","adoption scale","whether Reflection ships on the reported timeline"],"confidence":0.93,"previous_snapshot_id":null,"mapping_mode":"contemporaneous","reconstructed_at":null,"created_at":"2026-10-04T21:57:54.257Z","updated_at":"2026-10-04T21:57:54.257Z"},{"id":"state-models-20261006-beam-preview","layer_id":"layer-models","as_of":"2026-10-06 07:03:00.000","summary":"Reflection’s previously announced U.S. open-weight effort advanced to a concrete Beam early-access preview. Public weights and independent reproducibility remain pending.","dependencies":["layer-compute","layer-software","layer-open-source"],"constraints":["public weights not yet available","independent evaluation unavailable","deployment compute"],"emerging_capabilities":["early-access Beam model","large-scale GB300-based training and RL program"],"underdeveloped_components":["public Apache-2.0 weights","technical report and model card","independent benchmark replication"],"uncertainties":["relative frontier performance","deployment economics","adoption scale"],"confidence":0.94,"previous_snapshot_id":"state-models-20261004-open-weight","mapping_mode":"contemporaneous","reconstructed_at":null,"created_at":"2026-10-06T15:45:41.996Z","updated_at":"2026-10-06T15:45:41.996Z"},{"id":"state-models-20261007-european-preview","layer_id":"layer-models","as_of":"2026-10-07 06:25:00.000","summary":"The frontier/open-weight landscape now includes a concrete Mistral Large 4 API preview alongside Reflection Beam and the earlier Kolibri release. Mistral’s public weights and independent benchmark replication remain pending, so the new evidence expands the European capability path without establishing a completed open-weight release or superiority over Chinese peers.","dependencies":["layer-compute","layer-software","layer-open-source","layer-geopolitics"],"constraints":["public weights not yet available","independent evaluation unavailable","deployment economics","license/final artifacts pending"],"emerging_capabilities":["European frontier-scale multimodal MoE preview","agent/tool APIs with long-context operation"],"underdeveloped_components":["public weights","independent benchmark replication","verified deployment scale"],"uncertainties":["comparative frontier performance","later-October weight release","adoption scale"],"confidence":0.94,"previous_snapshot_id":"state-models-20261006-beam-preview","mapping_mode":"contemporaneous","reconstructed_at":null,"created_at":"2026-10-07T06:53:49.375Z","updated_at":"2026-10-07T06:53:49.375Z"}]}