Compare models with charts, understand changes through developments, and explore projects through maintained profiles.
FieldToFit Daily
The ServiceNow plugin is live, warehouse robotics remains a proof of concept, and New York brings child protection into chatbot feature rules. Ten reports connect products, compute and business delivery, clearly labeling new developments, recent progress and observations.
OpenAI and ServiceNow expanded their partnership. The ServiceNow MCP plugin is available in the OpenAI marketplace. Customers choose enabled capabilities, with Action Fabric applying existing business rules and permissions; richer interactive experiences remain in preview.
New York’s governor signed S.9051B/A.10379C, restricting chatbot features for minors that impersonate humans, foster emotional dependency, encourage self-harm or provide sexual content. It requires multiple age-verification methods, including one not solely dependent on government ID, and deletion of age information after use.
Microsoft reported fiscal Q1 2027 results for the quarter ended September 30: revenue of $92.3 billion, up 19%; Microsoft Cloud revenue of $63.9 billion, up 30%; and Azure revenue growth of 48%. GAAP net income was $36.6 billion, versus $36.4 billion excluding OpenAI investment effects.
Carbon initially rolls out to selected cyber defenders and Fairwind partners. Introductory input/output prices are $2/$10 per million tokens, later $4/$20; general developer, business and consumer access is still forthcoming.
2026-10-28
Eligible paid Pro and Business subscriptions can activate credits for one API organization after at least 7 days. First grants are prorated; they neither increase ChatGPT or signed-in Codex allowances nor cover Codex API-key usage.
2026-10-28
The ServiceNow MCP plugin is available in the OpenAI marketplace. Enhanced capabilities are expected by late 2026; voice, security and Codex in AI Workflow Factory follow separate 2027 schedules.
2026-10-28
Gemini 4 Carbon; announced 2026-10-28, initially available to selected cyber defenders and Fairwind partners.
2026-10-28
OpenAI API, Codex, ChatGPT Work and Oracle AI Workspace; partnership announcements dated 2026-10-27, with no single software version.
2026-10-27
GPT-6.1 Terra / gpt-6.1-terra; check adoption conditions separately for Work, Codex and the API.
2026-10-26
CRD derives dense-caption supervision from high-quality images. The generator stays frozen while the enhancer learns lighting, composition, layout and detail.
Report first released: 2026-10-24 · Selected here:
Non-streaming Flash and Next support speaker-structured transcription. Only Next adds general audio understanding; Streaming and Message support neither extension.
Report first released: 2026-10-20 · Selected here:
22 flags with severity at least 3 occur across 48,690 matched tasks; this is not an external-user incident rate.
Report first released: 2026-10-26 · Selected here:
Flagged trajectories update only FORGET; deployment removes that branch and retains RETAIN.
Report first released: 2026-10-12 · Selected here:
2017—2026 · How capabilities are learned, used and run efficiently.
Catalog updated 2026-10-24
Update model parameters through data, learning objectives and optimization to form and change capabilities, including pretraining, post-training and adaptation.
2017 · Sparse expert routing · Sparsely-gated MoE
2017 · Attention architecture · Transformer
2022 · Compute allocation · Chinchilla
2024 · Group-relative RL · DeepSeekMath
2024 · Reasoning RL · OpenAI · o1
2025 · Verifiable rewards · R1 / R1-Zero
2025 · Policy optimization · DAPO
2025 · Scaling sparse experts · Qwen3 MoE
2025 · Agent environments · Kimi K2
2025 · Constrained connections · mHC
2026 · Sparse / linear hybrid with mHC · GLM-5.3-Flash
2026 · Gated recurrence / attention with MoE · Qwen3.8-2.4T-A95B
2026 · Post-training on a shared base · GLM-5.3
2026 · Parallel agent RL · Kimi K2.5 · PARL
2026 · Unified multimodal prediction · ERNIE 5.0
2026 · KDA with global attention · Kimi K3
2026 · Constitution-guided post-training · Claude Opus 5.5
2026 · Joint pretraining and RL scaling · Muse Wonder / Full
Full map →Use trained models to process inputs and generate outputs, organizing search, verification and tools to solve tasks with existing capabilities.
2022 · Draft and verify · Speculative decoding
2024 · Latent KV compression · DeepSeek-V2 · MLA
2025 · Feature-based drafts · EAGLE-3
2026 · Sparse / linear hybrid with mHC · GLM-5.3-Flash
2026 · Block diffusion drafts · DFlash
2026 · Adaptive thinking · Claude Opus 4.6
2026 · Block sparse attention · MiniMax M3
2026 · KDA with global attention · Kimi K3
2026 · Learned internal reasoning · GPT-6 Astra / GPT-6.1 Sol
2026 · Cross-layer KV and index reuse · DeepSeek V4.1-Flash
Full map →Organize computation, memory and communication for training and inference through parallelism, kernels and scheduling on real hardware.
2019 · Training-state sharding · ZeRO
2022 · IO-aware attention · FlashAttention
2024 · Prefill/decode split · DistServe
2025 · Distributed PD serving · NVIDIA Dynamo
2026 · Blackwell kernels · FlashAttention-4
2026 · Block sparse attention · MiniMax M3
2026 · Encode/prefill/decode · Dynamo · EPD
2026 · Cross-layer KV and index reuse · DeepSeek V4.1-Flash
2026 · Asynchronous long-trajectory RL · Mistral Large 4
2026 · Distributed training on TPUs · Gemini 4 Argon
Full map →Selected milestones; chronology does not imply inheritance.
Explore full maps18 illustrated explanations with manual browser reading records.
Open full directory →15 companies · Checked 2026-10-23 · No link does not imply non-use. Source dates are not necessarily launch dates.
General / reasoning · Officially released
Training and internal reasoning are disclosed; architecture and production systems are not established.
Source date / precision as disclosed · 2026-09-03 · Checked 2026-10-23
GPT-6 Astra System CardSpecific disclosures appear below; other implementation details remain unknown.
reasoning_rl-o1 · The current card discloses reasoning RL (Sol by explicit reference); the optimizer is unspecified.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
GPT-6 Astra System Card§2 Model Data and Training · Relevant text reviewed · Checked 2026-10-23
Specific disclosures appear below; other implementation details remain unknown.
thinking-internal · Internal reasoning precedes the answer; prompted CoT and training are not equated.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
GPT-6 Astra System Card§2 Model Data and Training · Relevant text reviewed · Checked 2026-10-23
The reviewed material does not establish the implementation; no link does not imply non-use.
General / reasoning · Officially released
The addendum explicitly references Astra training types, not a distillation relationship.
Source date / precision as disclosed · 2026-09-29 · Checked 2026-10-23
GPT-6.1 Sol System Card AddendumSpecific disclosures appear below; other implementation details remain unknown.
reasoning_rl-o1 · The current card discloses reasoning RL (Sol by explicit reference); the optimizer is unspecified.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
GPT-6.1 Sol System Card Addendum§2 Model Data and Training · Relevant text reviewed · Checked 2026-10-23
Specific disclosures appear below; other implementation details remain unknown.
thinking-internal · Internal reasoning precedes the answer; prompted CoT and training are not equated.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
GPT-6.1 Sol System Card Addendum§2 Model Data and Training · Relevant text reviewed · Checked 2026-10-23
The reviewed material does not establish the implementation; no link does not imply non-use.
Agent / coding · Officially released
Thinking is always on. The system-card PDF remains unavailable; API documentation was reviewed.
Source date / precision as disclosed · 2026-09 · Checked 2026-10-23
What’s new in Opus 5.5Constitution-guided post-training is described; the exact recipe is not disclosed.
post-opus55 · Substantial constitution-guided post-training; no undisclosed optimizer is assumed.
This does not establish every step of the historical CAI recipe.
Claude Transparency HubClaude Opus 5.5: training methodology · Relevant text reviewed · Checked 2026-10-23
Specific disclosures appear below; other implementation details remain unknown.
thinking-adaptive · Adaptive thinking is always on, with depth controlled by effort rather than the old manual budget.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
What’s new in Opus 5.5New model / Breaking changes · Relevant text reviewed · Checked 2026-10-23
The reviewed material does not establish the implementation; no link does not imply non-use.
Separate capability tier · Identity / scope under review
Listed in the official system-card index; independent training evidence and availability remain under review.
Source date / precision as disclosed · 2026-09 · Checked 2026-10-23
Claude system cardsThe reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
Separate capability tier · Identity / scope under review
Listed in the official system-card index; independent training evidence and availability remain under review.
Source date / precision as disclosed · 2026-09 · Checked 2026-10-23
Claude system cardsThe reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
General / reasoning · Officially released
A new model, explicitly not a fine-tune of a prior model; the training stack is disclosed.
Source date / precision as disclosed · 2026-10-13 · Checked 2026-10-23
Gemini 4 Argon Model CardSpecific disclosures appear below; other implementation details remain unknown.
moe-sparse · Explicitly a sparse MoE Transformer with native multimodal support.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
Gemini 4 Argon Model CardArchitecture · Relevant text reviewed · Checked 2026-10-23
reasoning_rl-o1 · RL uses multi-step reasoning, problem-solving and theorem data.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
Gemini 4 Argon Model CardTraining Dataset · Relevant text reviewed · Checked 2026-10-23
The reviewed material does not establish the implementation; no link does not imply non-use.
Specific disclosures appear below; other implementation details remain unknown.
system-argon · TPU, JAX and ML Pathways are explicitly used for training.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
Gemini 4 Argon Model CardHardware / Software · Relevant text reviewed · Checked 2026-10-23
Agent / coding · Officially released
Post-training on the 5.2 base; the exact release date remains to be pinned.
Source date / precision as disclosed · 2026 · Checked 2026-10-23
GLM-5.3 / GLM-5.3-Flash official repositorySpecific disclosures appear below; other implementation details remain unknown.
post-glm53 · Uses the 5.2 base with improvements from post-training.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
GLM-5.3 / GLM-5.3-Flash official repositoryIntroduction: GLM-5.3 · Relevant text reviewed · Checked 2026-10-23
The reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
Agent / coding · Officially released
A new base and hybrid architecture; not all GLM-5 systems details can be inherited.
Source date / precision as disclosed · 2026 · Checked 2026-10-23
GLM-5.3 / GLM-5.3-Flash official repositorySpecific disclosures appear below; other implementation details remain unknown.
hybrid-glm53flash · The new base combines sparse / linear attention and mHC.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
GLM-5.3 / GLM-5.3-Flash official repositoryIntroduction: GLM-5.3-Flash · Relevant text reviewed · Checked 2026-10-23
repr-mhc · The official description explicitly names mHC.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
GLM-5.3 / GLM-5.3-Flash official repositoryIntroduction: GLM-5.3-Flash · Relevant text reviewed · Checked 2026-10-23
Specific disclosures appear below; other implementation details remain unknown.
hybrid-glm53flash · The same structure combines recurrent state and sparse access during inference.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
GLM-5.3 / GLM-5.3-Flash official repositoryIntroduction: GLM-5.3-Flash · Relevant text reviewed · Checked 2026-10-23
The reviewed material does not establish the implementation; no link does not imply non-use.
Agent / coding · Open weights
Report v1; the 1017 service update is separate, and the report date is not the model launch date.
Source date / precision as disclosed · 2026-09-17 · Checked 2026-10-23
DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache CompressionThe reviewed material does not establish the implementation; no link does not imply non-use.
Specific disclosures appear below; other implementation details remain unknown.
context-csa2 · Full, Reindex and Reuse control sharing and index recomputation.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression§2.3–2.3.1 · Relevant text reviewed · Checked 2026-10-23
Specific disclosures appear below; other implementation details remain unknown.
context-csa2 · Shared main KV and index keys reduce storage; index reuse avoids repeated computation.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression§2.3.1 · Relevant text reviewed · Checked 2026-10-23
General / reasoning · Model source recorded
Separate from V4.1-Flash; Flash implementation or service revisions do not establish Pro behavior.
Source date / precision as disclosed · Not stated · Checked 2026-10-23
DeepSeek-V4 model cardThe reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
Agent / coding · Open weights
Architecture sections reviewed; report revisions and model adoption dates remain distinct.
Source date / precision as disclosed · 2026-07-27 · Checked 2026-10-23
Kimi K3: Open Frontier IntelligenceSpecific disclosures appear below; other implementation details remain unknown.
hybrid-k3 · KDA and Gated MLA mix at 3:1, with an additional final global layer.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
Kimi K3: Open Frontier Intelligence§2.1 · Relevant text reviewed · Checked 2026-10-23
Specific disclosures appear below; other implementation details remain unknown.
hybrid-k3 · Recurrent state updates complement global access; this is not a purely linear model.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
Kimi K3: Open Frontier Intelligence§2.1 · Relevant text reviewed · Checked 2026-10-23
The reviewed material does not establish the implementation; no link does not imply non-use.
Max-class open weights · Open weights
An exact open-weight identity, not automatically the same as the API Max deployment.
Source date / precision as disclosed · 2026 · Checked 2026-10-23
Qwen3.8-2.4T-A95B model cardSpecific disclosures appear below; other implementation details remain unknown.
hybrid-q38 · Specifies Gated DeltaNet, Gated Attention, MoE and multi-step MTP.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
Qwen3.8-2.4T-A95B model cardModel Overview · Relevant text reviewed · Checked 2026-10-23
Specific disclosures appear below; other implementation details remain unknown.
thinking-adaptive · reasoning_effort adjusts reasoning and thinking history can be preserved; no claim of identical internals to Claude.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
Qwen3.8-2.4T-A95B model cardHighlights / API Usage · Relevant text reviewed · Checked 2026-10-23
The reviewed material does not establish the implementation; no link does not imply non-use.
General / reasoning · API / weights mapping under review
The card references Max, but the exact mapping to the API deployment is not verified.
Source date / precision as disclosed · 2026 · Checked 2026-10-23
Qwen3.8-2.4T-A95B model cardThe reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
Efficiency branch · Report abstract reviewed
Next is a separate branch; abstract-only details are not promoted to reviewed adoption links.
Source date / precision as disclosed · 2026 · Checked 2026-10-23
Qwen3.8-Next architecture reportThe reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
Agent / coding · Officially released
Release and role are clear; direct evidence for MSA or its training recipe was not obtained.
Source date / precision as disclosed · 2026-10-19 · Checked 2026-10-23
MiniMax M3.1 FlashThe reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
Agent / coding · Officially released
MSA and kernel disclosures apply to M3, not automatically M3.1 Flash.
Source date / precision as disclosed · 2026-06-01 · Checked 2026-10-23
MiniMax M3The reviewed material does not establish the implementation; no link does not imply non-use.
Specific disclosures appear below; other implementation details remain unknown.
context-msa · Block-selected sparse attention.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
MiniMax M3MSA · Relevant text reviewed · Checked 2026-10-23
Specific disclosures appear below; other implementation details remain unknown.
context-msa · A KV-outer kernel gathers matching queries to reduce repeated reads.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
MiniMax M3MSA · Relevant text reviewed · Checked 2026-10-23
Agent / coding · Officially released
The official post describes pretraining and RL scaling; Llama architecture is not presumed for Muse.
Source date / precision as disclosed · 2026-10-12 · Checked 2026-10-23
Muse Wonder and Muse Wonder FullSpecific disclosures appear below; other implementation details remain unknown.
post-wonder · The official post explicitly describes scaling pretraining and reinforcement learning; the optimizer is unspecified.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
Muse Wonder and Muse Wonder FullScaling Toward Personal Superintelligence · Relevant text reviewed · Checked 2026-10-23
The reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
Agent / coding · Available to Muse Code subscribers
The official post describes pretraining and RL scaling; Llama architecture is not presumed for Muse.
Source date / precision as disclosed · 2026-10-12 · Checked 2026-10-23
Muse Wonder and Muse Wonder FullSpecific disclosures appear below; other implementation details remain unknown.
post-wonder · The official post explicitly describes scaling pretraining and reinforcement learning; the optimizer is unspecified.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
Muse Wonder and Muse Wonder FullScaling Toward Personal Superintelligence · Relevant text reviewed · Checked 2026-10-23
The reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
General / reasoning · Officially released
Release material reviewed; task performance is not an algorithm or architecture disclosure.
Source date / precision as disclosed · 2026-09-21 · Checked 2026-10-23
Grok 4.7The reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
Agent / coding · Officially released
The official page identifies both sizes; the model-card PDF could not be read and the training recipe is unverified.
Source date / precision as disclosed · Not stated · Checked 2026-10-23
Seed2.1 Pro and TurboThe reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
Agent / coding · Officially released
The official page identifies both sizes; the model-card PDF could not be read and the training recipe is unverified.
Source date / precision as disclosed · Not stated · Checked 2026-10-23
Seed2.1 Pro and TurboThe reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
General / reasoning · Public preview
Currently a public preview; planned end-of-month weights are not marked as released.
Source date / precision as disclosed · 2026-10-06 · Checked 2026-10-23
Introducing Mistral Large 4Specific disclosures appear below; other implementation details remain unknown.
agent-k2 · RL spans tool environments with tests, judges and static checks; this is not adoption of Kimi’s recipe.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
Introducing Mistral Large 4Reinforcement learning at scale · Relevant text reviewed · Checked 2026-10-23
The reviewed material does not establish the implementation; no link does not imply non-use.
Specific disclosures appear below; other implementation details remain unknown.
system-ml4 · Rollout actors and training execute asynchronously with staleness control.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
Introducing Mistral Large 4Reinforcement learning at scale · Relevant text reviewed · Checked 2026-10-23
General / reasoning · Reported model; latest status under review
Unified multimodal architecture reviewed; report date does not establish launch or latest-flagship status.
Source date / precision as disclosed · 2026-02-04 · Checked 2026-10-23
ERNIE 5.0 Technical ReportSpecific disclosures appear below; other implementation details remain unknown.
train-ernie5 · Unified next-group-of-tokens training with modality-agnostic expert routing.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
ERNIE 5.0 Technical Report§2–2.1 · Relevant text reviewed · Checked 2026-10-23
The reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
Agent / coding · Public preview
Tencent platform documentation confirms MoE and preview status; integration is not proof of model internals.
Source date / precision as disclosed · 2026-08-28 · Checked 2026-10-23
Hy4 Preview on Tencent ADPSpecific disclosures appear below; other implementation details remain unknown.
moe-sparse · Officially identified as sparse MoE; the exact routing strategy is not established.
Supports the described mechanism or application, not every historical implementation detail or weight lineage.
Hy4 Preview on Tencent ADP核心Capability与技术参数 / Architecture · Relevant text reviewed · Checked 2026-10-23
The reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
Current flagship identity unresolved · Identity / scope under review
Service documentation found, but no exact current flagship pinned; historical Pro MoE is not substituted.
Source date / precision as disclosed · 2026-10-14 · Checked 2026-10-23
Pangu model service overviewThe reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
The reviewed material does not establish the implementation; no link does not imply non-use.
2026-10-26 · Official Day 22 · Codex / Work
GPT-6.1 Terra reaches Work, Codex and the API. Standard input / cached-input / output prices are $1 / $0.05 / $5 per million tokens. Full system-card analysis separates internal deployment simulation from authorization tests.
2026-10-26 · Other OpenAI
Other OpenAI: three research disclosures; artifact access and review distinguished
2026-10-26 · Other OpenAI
Other OpenAI: Signals work-activity data with explicit coverage limits
Intelligence Index v4.3 · USD per benchmark task · 172 configurations
Source date: Not provided · Checked: 2026-10-26
Original sourceText Arena · Overall · Style Control · USD per million output tokens · 99 configurations
Source date: 2026-10-22 · Checked: 2026-10-26
Original source2026-10-28 · ChatGPT / Codex: ServiceNow MCP is available, with phased voice, security and integration plans
2026-10-28 · ChatGPT / Codex: ServiceNow MCP is available, with phased voice, security and integration plans
2026-10-28 · Gemini 4 Carbon: limited defender rollout, with introductory and later prices