Frontier-lab and regional census
As of: 2026-08-26
Parent issues: #9300, #9313
Authority: ../index.json
Result first
The current index contains 59 frontier-lab entries within a 207-entry
corpus. This pass adds primary-source coverage for major U.S./Canadian labs and for
China, Israel, the Middle East, India, Japan, Korea, and Southeast Asia. It reduces a
strong OpenAI/Anthropic bias, but it does not make the global census complete.
The new evidence strengthens four expectations:
- One endpoint is often several workload classes. Reasoning controls, modality,
language, and agent/tool modes change output length, latency, memory, and batchability.
- Active parameters are not a deployment receipt. MoE expert count/routing,
quantization, total parameters, host boundary, KV compression, and network topology
all matter.
- Regional and sovereign labs optimize different envelopes. Data locality,
language/tokenization, edge/private deployment, and national compute programs can be
first-order constraints.
- Training disclosure is uneven. Ai2 gives a physical 1,280-H100 cluster and MFU;
many other labs publish only model architecture or benchmark claims. Unknown stays
unknown.
A checked row means the index contains at least one dated source with the fields below.
It does not mean complete organizational coverage.
| Field |
What to capture |
| Identity |
Lab, parent, headquarters/region, and relevant cloud or sovereign program. |
| Model/workload |
Model family, modality, reasoning/tool mode, context, and target product workload. |
| Physical envelope |
Accelerator SKU/count, node/host boundary, topology, memory/quantization, power, and region when disclosed. |
| Scale |
Total/active parameters, experts, training tokens, users, requests, or production traffic—never substitute one for another. |
| Lifecycle |
Research, preview, released, API, on-device, private-cloud, in production, or planned. |
| Evidence |
Primary source first; retain official statement, vendor claim, measurement, report, or rumor labels. |
| Limits |
Missing denominator, undisclosed cluster, author benchmark, maximum-versus-prevalence, and traffic uncertainty. |
| fak use |
Concrete benchmark, scheduler, routing, cache, hardware, or trust-boundary implication. |
Checked labs and current evidence
United States and Canada
| Status |
Lab |
Evidence now indexed |
What remains unknown |
| Checked |
OpenAI |
Stargate/site, Oracle/NVIDIA, user-scale, and capacity announcements. |
Physical installed/healthy fleet, workload mix, tenant concentration, and plan-to-goodput conversion. |
| Checked |
Anthropic |
AWS/Google/Azure partnerships, multicloud failure evidence, U.S. infrastructure investment, and usage research. |
Provider traffic shares, physical fleet, cache/batch distributions, and regional failover. |
| Checked |
xAI |
Colossus/Grok model and financing claims. |
Physical SKU census behind H100-equivalent totals, healthy capacity, utilization, and request/token distributions. |
| Checked |
Google DeepMind |
Gemini 3.1 Flash-Lite: TPU/JAX/Pathways training, 1M input, 64K output, high-volume/latency-sensitive positioning, and five distribution channels. |
TPU generation/count, pod topology, production model mix, batch/concurrency, and neutral goodput. |
| Checked |
Meta |
Llama 4 MoE active/total parameter and host-fit envelope; existing datacenter and capex evidence. |
Production expert skew, fleet share, Meta AI request/token distribution, and quality-constrained serving economics. |
| Checked |
Amazon Nova |
Nova 2 family, configurable extended thinking, multimodality, and up-to-1M context. |
Training/serving hardware, mode-selection shares, Bedrock traffic, and output-length tails. |
| Checked |
Apple |
Roughly 3B on-device model, 2-bit QAT, KV sharing, server Parallel-Track MoE, and Private Cloud Compute split. |
Server size/fleet, request routing share, concurrency, and production latency. |
| Checked |
Cohere |
Command A enterprise/agentic and 256K-context positioning. |
Neutral efficiency reproduction, private-deployment fleet, customer/tenant traffic, and regional demand. |
| Checked |
Ai2 |
OLMo 2 32B trained to 6T tokens on 160 × 8 H100 nodes; >1,800 tokens/s/GPU and ~38% MFU. |
Net wall-clock availability/cost and inference demand; training throughput is not serving goodput. |
| Checked |
Microsoft Research Phi |
Phi-4-reasoning 14B inference-time-compute model; compact open-weight reasoning envelope. |
Physical training/serving cluster, first-party traffic, reasoning-length distribution, and neutral goodput. |
| Unchecked |
NVIDIA model research |
NVIDIA appears as platform/partner, not as a separately audited model lab. |
Nemotron model training and deployed inference envelope. |
| Unchecked |
Databricks/MosaicML, Essential AI, Reflection AI, Safe Superintelligence, Thinking Machines, World Labs |
No dedicated primary-source entries. |
Model status, cluster/funding-to-capacity conversion, product workload, and deployment state. |
China
| Status |
Lab |
Evidence now indexed |
What remains unknown |
| Checked |
DeepSeek |
V3 training cluster/economics and hardware/architecture reflections. |
Current fleet, export-control response, production traffic, and post-release capacity. |
| Checked |
Alibaba Qwen |
Qwen3 reasoning-budget evidence plus Alibaba Cloud/cache entries. |
Physical Qwen training/serving cluster, product traffic, and reasoning-mode shares. |
| Checked |
Moonshot AI |
Kimi K2 model/cluster evidence. |
Production request distribution, utilization, cache behavior, and financing-to-capacity lifecycle. |
| Checked |
MiniMax |
M3 1M-context coding/multimodal release. |
Context prevalence, cluster hardware, traffic, compaction, and serving cost. |
| Checked |
ByteDance Seed |
Seed1.5-VL 532M vision encoder + 20B-active MoE, GUI/game agent tasks, and Volcano Engine availability. |
Total parameters, cluster hardware, Doubao/Volcano traffic, modality share, and production routing. |
| Checked |
Baidu |
ERNIE 5.0 PaddlePaddle hybrid parallelism, FP8 activation storage, adaptive offload, and separated tokenizer/backbone nodes. |
Cluster size/SKU, net training efficiency, serving topology, and ERNIE traffic. |
| Checked |
Tencent |
Hunyuan-Large 389B total/52B active, 256K context, mixed routing, and KV compression. |
Expert-load distribution, cluster, deployed fleet, and service economics. |
| Checked |
Z.ai / Zhipu AI |
GLM-4.5 355B total/32B active, 23T training tokens, and hybrid thinking/direct modes. |
Cluster hardware, mode-selection rates, API traffic, and independent benchmark reproduction. |
| Checked |
Huawei Cloud Pangu |
Pangu 5.0 tiers above 1B/10B/100B/1T parameters and >400 scenarios across >30 industries; the separate Atlas 900 A3 SuperPoD vendor specification names 384 Ascend NPUs, 300 TB HBM, 48 PFLOPS dense BF16, 16.1 PB/s HBM bandwidth, 7.2 PB/s scale-up bandwidth, and 96 cabinets. |
Built/installed/healthy/schedulable deployment, power/cooling, model-tier traffic, requests/users, batch/cache behavior, and neutral goodput. |
| Partial |
01.AI |
Yi-34B-200K model card: 34B parameters, 200K advertised context, 5B-token long-context continuation. |
Physical compute, neutral long-context serving results, production traffic, and lifecycle evidence. |
| Partial |
StepFun |
Step3 release: 321B total / 38B active multimodal MoE and attention/FFN disaggregation design. |
Training cluster, neutral matched efficiency, shipped serving topology, adoption, and traffic. |
| Covered |
Baichuan, iFLYTEK, Meituan |
Baichuan 2 run-specific A800/token envelope; iFLYTEK typed ecosystem denominators; Meituan LongCat, DORA, MTServe, and MTGenRec records. |
iFLYTEK hardware/Ascend receipts, absolute traffic, user/session distributions, and neutral production performance remain missing. |
| Partial |
Shanghai AI Lab / InternLM |
InternLM3-8B release: 4T training tokens and first-party training/deployment toolchain. |
Physical compute, cost accounting, production deployment, and traffic. |
| Partial |
SenseTime |
SenseNova 5.0 vendor release: MoE, >10TB token data, ~200K effective context, cloud-device-edge matrix. |
Neutral benchmark, physical compute, production workload, and lifecycle evidence. |
| Partial |
Xiaomi MiMo |
MiMo-V2-Pro release: >1T total / 42B active parameters, 1M context, tiered API pricing above 256K. |
Physical compute, traffic/context distribution, neutral serving performance, and adoption. |
Europe
| Status |
Lab |
Evidence now indexed |
What remains unknown |
| Checked |
Mistral AI |
Mistral Compute, model training, environmental LCA, and cross-cloud distribution. |
Physical cluster and utilization, API/private mix, tenant concentration, and sovereign-cloud economics. |
| Unchecked |
Aleph Alpha, Black Forest Labs, Helsing, Kyutai, LightOn, Poolside, Silo AI/AMD |
No dedicated primary-source model/infrastructure entries. |
Training/serving envelopes, acquisitions/ownership effects, and production traffic. |
| Unchecked |
UK labs beyond DeepMind |
No dedicated census for Stability AI, Synthesia, Wayve, or university/sovereign programs. |
Model scope, compute, financing, and deployment state. |
Israel
| Status |
Lab |
Evidence now indexed |
What remains unknown |
| Checked |
AI21 Labs |
Jamba 1.5 hybrid SSM-Transformer MoE, 256K effective context, single-GPU/one-node fit claims, and quantized deployment envelope. |
Production concurrency, traffic, hardware mix, and neutral long-context throughput. |
| Unchecked |
AI research/product labs at NVIDIA Israel and other private firms |
No dedicated entries. |
Lab boundaries, model programs, and physical deployment. |
Middle East
| Status |
Lab |
Evidence now indexed |
What remains unknown |
| Checked |
Technology Innovation Institute (UAE) |
Falcon Arabic and Falcon-H1 sovereign-language, hybrid Transformer/Mamba, multilingual, edge-to-enterprise positioning. |
Training cluster, production traffic, per-tier throughput/energy, and independent performance. |
| Partial |
G42 / Inception / MBZUAI (UAE) |
Stargate UAE plus MBZUAI annual-review evidence for NANDA, LLM360, and K2-65B. |
Dedicated Jais evidence, model training/serving envelope, national cluster allocation, and product demand. |
| Unchecked |
Saudi SDAIA/ALLaM and other sovereign programs |
No dedicated entry. |
Primary model sources, compute allocation, language traffic, and deployment. |
India
| Status |
Lab |
Evidence now indexed |
What remains unknown |
| Checked |
Sarvam AI |
30B/105B 128-expert models trained in India on IndiaAI compute, 16T/12T tokens, with named production products. |
Physical cluster/allocation, request volumes, traffic mix, goodput, and independent performance. |
| Unchecked |
Krutrim, BharatGen, Soket AI, Gnani, CoRover, IIT/research consortia |
No dedicated entries. |
Model families, IndiaAI allocation, language/token distributions, and production state. |
Japan
| Status |
Lab |
Evidence now indexed |
What remains unknown |
| Checked |
Sakana AI |
Composition/model-merging, automated AI science, edge-efficient model strategy, and enterprise deployment claims. |
Model and cluster scale, measured energy, customer traffic, and repeatable production outcomes. |
| Checked |
NTT tsuzumi |
0.6B CPU and 7B single-GPU Japanese enterprise deployment tiers. |
Hardware SKU, quantization, context, production latency/concurrency, and demand. |
| Unchecked |
Preferred Networks, Fujitsu, SoftBank/SAKURA, AIST and university programs |
No dedicated model-lab entries. |
Domestic accelerator/cloud use, training runs, sovereign programs, and service traffic. |
Korea
| Status |
Lab |
Evidence now indexed |
What remains unknown |
| Checked |
NAVER Cloud |
HyperCLOVA X THINK: 6T Korean/English tokens, 128K context, reasoning/length-control training, and multimodal/agentic targets. |
Parameter/cluster size, production demand, language mix, and serving economics. |
| Checked |
LG AI Research |
EXAONE 4.0 32B professional + 1.2B on-device hybrid reasoning/direct family with tool use. |
Cluster, traffic, mode selection, device latency, and neutral efficiency. |
| Partial |
Samsung Electronics |
Gauss2 variants, MoE Supreme tier, internal developer and call-center use. |
Model sizes, physical cluster, request/token distribution, latency, and external product deployment. |
| Partial |
SK Telecom |
A.X 3.1/4.0 dual strategy, 72B/7B variants, and A. call-summarization application. |
Training/serving hardware, traffic, latency, batching, and external adoption. |
| Partial |
Kakao |
Kanana technical report and open 2.1B Nano release across a 2.1B-32.5B model family. |
Physical compute, production deployment, service traffic, and lifecycle. |
Southeast Asia
| Status |
Lab |
Evidence now indexed |
What remains unknown |
| Checked |
Sea AI Lab |
Sailor2: stated production preference for 8B/20B tiers, 1B speculative-decoding/research tier, ~500B continued-pretraining tokens, and 15 languages. |
Request counts, tenant/language shares, speculative acceptance, hardware, and deployed goodput. |
| Checked |
AI Singapore SEA-LION |
v4.5 multilingual/multimodal/agentic family for >11 SEA languages with custom speculative decoder. |
Model size, hardware, acceptance distribution, traffic, and language/user shares. |
| Unchecked |
Grab, GoTo, SCB10X, VinAI, regional sovereign programs |
SCB10X appears only as a Sailor2 collaborator; no dedicated lab entries. |
Model programs, compute, product demand, and country/language distributions. |
Baichuan, iFLYTEK, and Meituan now have dedicated records, but maturity differs: Baichuan
is a disclosed training run, iFLYTEK is an ecosystem/adoption filing without hardware or
absolute traffic, and Meituan spans model training plus internal production-system
measurements. These records do not turn author/vendor maxima into neutral prevalence.
Cross-lab parameter observations
| Dimension |
Evidence in this batch |
Safe expectation |
| Reasoning mode |
Amazon Nova 2, GLM-4.5, HyperCLOVA X THINK |
Treat reasoning selection as a workload variable with separate output/SLO distributions. |
| MoE |
Llama 4, Seed1.5-VL, Hunyuan-Large, GLM-4.5, Sarvam |
Capture total/active parameters, expert count, routing skew, memory, and topology; active parameters alone are insufficient. |
| Long context |
Gemini 1M/64K output; Nova up to 1M; Command A/Jamba/Hunyuan/HyperCLOVA 128K–256K |
Capability ceiling is not prevalence. Benchmark token-time, KV occupancy, compaction, and multi-tenant concurrency. |
| Device/private cloud |
Apple on-device + Private Cloud Compute; TII/Sakana edge claims |
Route by trust and resource boundary; do not collapse device and datacenter economics. |
| Regional language |
Falcon Arabic, Sarvam, HyperCLOVA, Sailor2 |
Measure tokenizer expansion, language mix, data locality, and locale-specific SLOs. |
| Physical training receipt |
Ai2 OLMo: 1,280 H100s, GPUDirect-TCPXO, >1,800 tokens/s/GPU, ~38% MFU |
Prefer physical SKU/topology/utilization receipts over “equivalent” counts. Training evidence still does not prove serving demand. |
| Speculative decoding |
Sailor2 1B tier named for speculative decoding |
Measure acceptance, verifier cost, draft placement, and net goodput under regional-language workloads. |
Remaining priority queue
- NVIDIA Nemotron plus Microsoft first-party deployment/traffic evidence.
- Huawei/iFLYTEK Ascend deployment and power/cooling receipts, 01.AI, StepFun,
Shanghai AI Lab, SenseTime, and Xiaomi; deepen Baichuan and Meituan traffic/user evidence.
- G42/Inception Jais/Nanda deployment evidence and Saudi ALLaM/SDAIA sovereign programs.
- Samsung Gauss, SKT A.X, and Kakao-related Korean programs.
- Preferred Networks, Fujitsu, SoftBank/SAKURA, and Japanese public compute.
- Grab/GoTo, SCB10X, VinAI, and regional sovereign initiatives.
- Europe beyond Mistral and DeepMind, plus private U.S./Canadian labs.
- For every checked lab: a second pass on failures, cancellations, users/requests,
physical serving hardware, geography, and lifecycle from release to active goodput.
The census remains partial until that queue and the broader
../coverage-audit.md requirements are closed with primary
sources and explicit unknowns.