Start with the constraint that governs a practical deployment: documented, usable facility capacity. This reference connects leading open-weight models with cooling-aware setup estimates so buyers can screen whether a site merits utility, engineering, and counterparty diligence before pursuing a conversion.
Models reviewed Nov. 6, 2026; accelerator specifications and the power method reviewed Oct. 12, 2026. Reconfirm model availability, licenses, accelerator support, and manufacturer specifications before deployment. The separate vendor-reported efficiency explorer was reviewed Sep. 27, 2026.
New since the Oct. 12, 2026 review
Nov 5Inkling 2 Small — announced Nov 5; weights pending
Nov 5ERNIE-5.5-Flash-Base — base model weights
Nov 4Kimi K3.1 Pro — weights released
Oct 29Mistral Large 4 — weights released
Oct 23MiniMax M3.1 Flash — weights released
Oct 17DeepSeek V4.1-Flash-1017 — refresh released
Oct 16Beam 1.0 — weights released
Oct 15Muse Spark Open — weights released
Oct 15Step 5 Preview — weights released
Oct 15NVIDIA Nemotron 3.5 Super VL — weights released
Enter facility capacity and a primary accelerator type to estimate a single-model deployment envelope. Setup power figures already include cooling and distribution; this tool does not apply a second PUE adjustment.
kW
1.5 MW facility input
100 kW10 MW
Primary GPU type
Facility input
1.5 MW
H100 SXM5 capacity
861 GPUs
8-GPU node equivalent
107
Aggregate VRAM
68.9 TB
Each H100 SXM5 is budgeted at 1.74 kW of facility power: NVIDIA’s published maximum for an eight-GPU system (10.2 kW), plus 5% for networking and storage, multiplied by a 1.30× air-cooling and distribution factor.
Operational models you can screen
Models with published weights only. Each maximum is a mutually exclusive screening estimate for one model—not capacity to run every listed model at once. A setup is the smallest group of accelerators that can hold the model; production serving usually needs more for concurrent users and long contexts.
Scroll horizontally to review all model-fit columns.
Available open-weight models screened against the selected facility capacity and accelerator type.
Model
Origin
Params
Minimum GPUs
Est. power / setup
Max instances*
Setup basis
Llama 4 ScoutMeta
🇺🇸
109B
1× H100 SXM5
1.7 kW
861
Developer guidance
NVIDIA Nemotron 3.5 Super VLNVIDIA
🇺🇸
120B
2× H100 SXM5
3.5 kW
430
Developer guidance
Poolside Laguna S 2.1Poolside
🇺🇸
118B
2× H100 SXM5
3.5 kW
430
Developer guidance + UNICANN estimate
Inkling-SmallThinking Machines Lab
🇺🇸
276B
4× H100 SXM5
7.0 kW
215
Developer guidance + UNICANN estimate
Beam 1.0Reflection AI
🇺🇸
501B
8× H100 SXM5
14 kW
107
Developer guidance + UNICANN estimate
DeepSeek V4.1-Flash-1017DeepSeek
🇨🇳
552B (+ Engram memory)
8× H100 SXM5
14 kW
107
Developer guidance + UNICANN estimate
ERNIE-5.5-Flash-BaseBaidu
🇨🇳
280B
8× H100 SXM5
14 kW
107
Developer guidance + UNICANN estimate
Llama 4 MaverickMeta
🇺🇸
400B
8× H100 SXM5
14 kW
107
Developer guidance + UNICANN estimate
MiniMax M3.1 FlashMiniMax
🇨🇳
~428B
8× H100 SXM5
14 kW
107
Developer guidance
Mistral Large 3Mistral AI
🇫🇷
675B
8× H100 SXM5
14 kW
107
Developer guidance + UNICANN estimate
Tencent Hy3Tencent
🇨🇳
295B
8× H100 SXM5
14 kW
107
Developer guidance + UNICANN estimate
DeepSeek V4-ProDeepSeek
🇨🇳
1.6T
16× H100 SXM5
28 kW
53
UNICANN estimate
GLM-5.3Z.ai (Zhipu AI)
🇨🇳
744B
16× H100 SXM5
28 kW
53
Developer guidance + UNICANN estimate
Hy4 (preview)Tencent
🇨🇳
770B
16× H100 SXM5
28 kW
53
Developer guidance + UNICANN estimate
InklingThinking Machines Lab
🇺🇸
975B
16× H100 SXM5
28 kW
53
Developer guidance + UNICANN estimate
Ling-3.1-flashAnt Group (inclusionAI)
🇨🇳
560B
16× H100 SXM5
28 kW
53
Developer guidance + UNICANN estimate
MiMo-V2.6-ProXiaomi
🇨🇳
1.02T
16× H100 SXM5
28 kW
53
Developer guidance + UNICANN estimate
Muse Spark OpenMeta
🇺🇸
816B
16× H100 SXM5
28 kW
53
Developer guidance + UNICANN estimate
NVIDIA Nemotron 3 UltraNVIDIA
🇺🇸
550B
16× H100 SXM5
28 kW
53
Developer guidance
Qwen3.5-397BAlibaba (Qwen)
🇨🇳
397B
16× H100 SXM5
28 kW
53
Developer guidance + UNICANN estimate
Step 5 PreviewStepFun
🇨🇳
600B
16× H100 SXM5
28 kW
53
Developer guidance + UNICANN estimate
Trinity-Large-ThinkingArcee AI
🇺🇸
~398B
16× H100 SXM5
28 kW
53
UNICANN estimate
Kimi K3.1 ProMoonshot AI
🇨🇳
2.8T
32× H100 SXM5
56 kW
26
Developer guidance + UNICANN estimate
Mistral Large 4Mistral AI
🇫🇷
1.05T
32× H100 SXM5
56 kW
26
Developer guidance + UNICANN estimate
Qwen3.8-2.4TAlibaba (Qwen)
🇨🇳
2.4T
32× H100 SXM5
56 kW
26
Developer guidance + UNICANN estimate
*Per-setup estimates include GPU and host system power, networking, cooling (air or direct liquid), and distribution losses. “Developer guidance” means the count comes from the model card or the developer’s official inference recipe; “UNICANN estimate” means the count is sized from the published checkpoint with about 25% memory headroom. Actual draw, physical layout, interconnection availability, redundancy, and performance need site-specific validation. Upcoming watchlist entries are intentionally excluded from these operational results.
Model catalog
Frontier open-weight models
Models with downloadable weights appear by default; announced releases without weights stay on a separate watchlist. Total parameters drive memory needs, while active parameters drive per-token compute. Each hardware panel shows whether its GPU count comes from the developer or from UNICANN’s memory-fit estimate.
🇨🇳
Kimi K3.1 Pro
Moonshot AI · Nov 4, 2026 (API since Oct 23)
Total params
2.8T
Active params
104B
Context
1M tokens
License
Kimi K3.1-Pro License
License terms: A separate agreement is required for model-as-a-service above $20M in revenue over 12 months; UI attribution is required above 100M MAU or $20M in monthly revenue.
Architecture: MoE (16 of 896 experts + 2 shared), Kimi Delta Attention + sparse attention
Modalities: Text, images in (video as keyframes); text out
Hardware requirement
Developer guidance + UNICANN estimate
Official vLLM recipe: 8× B300 or 8× GB300 (or 8× MI355X).
Checkpoint: MXFP4 experts · ~1.56 TB
8× B300
18 kW
16× H200
28 kW
32× H100
56 kW
Est. power per setup · B300 assumes direct liquid cooling; H200 and H100 assume air cooling.
Largest open-weight model available; succeeds Kimi K3 (Jul 27, 2026) at the same size
License terms: A separate license is required for model-as-a-service or AI-assistant use above $50M in revenue over 12 months; UI attribution is required above 100M MAU or $20M in monthly revenue.
Architecture: MoE (10 of 512 experts + 1 shared), hybrid linear attention
Modalities: Text
Hardware requirement
Developer guidance + UNICANN estimate
Official vLLM recipe: NVFP4 on 8× B300 or 16× H200; FP8 on 16× B300 or 32× H200; BF16 on 24× B300 or 48× H200.
Checkpoint: BF16 · ~4.9 TB (FP8 and NVFP4 recipes published)
8× B300
18 kW
16× H200
28 kW
32× H100
56 kW
Est. power per setup · B300 assumes direct liquid cooling; H200 and H100 assume air cooling.
Text-only open checkpoint of the Qwen3.8-Max family
License terms: Products must display “Built with Muse Spark Open”; model-as-a-service providers above $20M in revenue over 12 months need a separate agreement with Meta.
Architecture: MoE (8 of 256 experts)
Modalities: Text, images, video, audio in; text out
Hardware requirement
Developer guidance + UNICANN estimate
Official recipes: NVFP4 on 4× GB300, 8× B200, or one 8× H200 node; BF16 on 16× H200. Model card: at least 700 GB of GPU memory for NVFP4.
Checkpoint: BF16 · ~1.64 TB; NVFP4 · ~0.53 TB
4× B300
9.1 kW
8× H200
14 kW
16× H100
28 kW
Est. power per setup · B300 assumes direct liquid cooling; H200 and H100 assume air cooling.
Meta’s first open-weight Muse model; U.S.-developed
License terms: A custom Apache-style license. “Powered by StepFun AI” attribution is required above 100M MAU or $20M in monthly revenue; model-as-a-service above $20M in revenue needs a separate agreement.
Architecture: Sparse MoE
Modalities: Text, images in; text out
Hardware requirement
Developer guidance + UNICANN estimate
Model card: FP8 on 4× B300 or 8× H200; BF16 on 8× B300 or 16× H200.
License terms: Commercial use requires “Built with MiniMax M3.1-Flash” attribution and a one-time notice; prior authorization is required above $20M in annual revenue.
Architecture: MoE (same parameter count as MiniMax M3)
Modalities: Text, images, video in; text out
Hardware requirement
Developer guidance
Official vLLM recipe: 4× B200, B300, or GB300; 8× H100 or H200.
Checkpoint: MXFP8 dense + NVFP4 experts · ~252 GB
4× B300
9.1 kW
8× H200
14 kW
8× H100
14 kW
Est. power per setup · B300 assumes direct liquid cooling; H200 and H100 assume air cooling.
Refresh of MiniMax M3 (June 2026) with a much smaller checkpoint
License terms: Not licensed in the EU, UK, or South Korea. Above $50M in related annual revenue, written authorization is required; below it, a one-time notice and “Powered by ERNIE” attribution.
Architecture: MoE base model (no post-training)
Modalities: Text
Hardware requirement
Developer guidance + UNICANN estimate
SGLang-verified recipes: FP8 on 2× B300, 4× B200, or 8× H100.
Checkpoint: BF16 · ~573 GB; FP8 · ~304 GB
2× B300
4.6 kW
4× H200
7.0 kW
8× H100
14 kW
Est. power per setup · B300 assumes direct liquid cooling; H200 and H100 assume air cooling.
Base model for fine-tuning; the post-trained ERNIE 5.5 Flash is not yet released
Each hardware requirement is the smallest accelerator group that can hold a model’s published checkpoint. Production deployments typically use 2–4× the minimum for concurrent requests and KV-cache headroom, and long-context serving can need more. Some licenses add revenue thresholds, attribution, or separate commercial agreements; review the full license before deployment. Upcoming entries are watchlist material, not operational recommendations.
Estimate methodology
Power per setup is an engineering estimate inclusive of GPU TDP, host system overhead, networking, cooling method (air or direct liquid), and power distribution losses. Actual draw varies with workload intensity, ambient conditions, hardware generation, and facility design. These figures are for capacity planning only and do not constitute a guarantee of performance or consumption.
American open-weight renaissance
U.S. labs are competing for open-weight leadership
American labs are expanding the range of open-weight options alongside Chinese developers. Thinking Machines, NVIDIA, Meta, Arcee, Poolside, Reflection AI, and AI2 illustrate a broad ecosystem with different licenses, modalities, and deployment assumptions. Buyers should confirm current weights, license terms, and support directly with each developer.
Inkling (July 15, 2026) and Inkling-Small (July 30) are available under Apache 2.0; Inkling-Small fits on a single B300 in NVFP4. Inkling 2 Small, announced Nov. 5, is hosted on Tinker now, with Apache 2.0 weights promised after external safety testing.
Nemotron 3 Ultra (550B) · Nemotron 3.5 Super VL (120B)
Open weights, open data, and open training recipes on a hybrid Mamba-2 + MoE architecture. Nemotron 3.5 Super VL (Oct. 15) adds image and video input and runs on a single H200 in NVFP4.
U.S.-trained reasoning model under OpenMDW-1.1, with no revenue threshold. Arcee is also working with the U.S. Department of Energy on Genesis-Science-1, an open-weight science model.
Weights released Oct. 16, 2026, under Apache 2.0: a 501B-total, 23B-active sparse MoE text model. Reflection’s reference setup runs NVFP4 on four B300 GPUs.
Fully open research models with public code and data under Apache 2.0. AI2’s largest current model is 32B dense; AI2 says its next generation will use a mixture-of-experts design.
Muse Spark Open, released Oct. 15, 2026, is Meta’s first open-weight Muse model: 816B total, 42B active, with text, image, video, and audio input. Its license requires attribution and a separate agreement for large model-as-a-service providers.
Meta’s Llama 4 models also remain open-weight and appear in the catalog above.
On the horizon
Upcoming open-weight releases
Announced models whose weights were not yet downloadable as of Nov. 6, 2026. The trend is toward larger MoE systems that need more aggregate GPU memory — and more facility power. Dates come from each developer and can slip.
🇺🇸
Inkling 2 Small
Weights expected mid-Nov 2026
Thinking Machines Lab
594B total / 17B active, up to 1M context, text and image input. Hosted on Tinker since Nov. 5; Thinking Machines says the full Apache 2.0 weights follow after external safety testing and red-teaming.
Final release of Tencent’s next flagship, launched as a hosted product. Tencent says the weights and a system card will follow; final sizes and license are not yet published.
Trillion-parameter-class American open-weight science model built with the U.S. Department of Energy’s Genesis Mission. Exact size, license, and date not yet published.
MiniMax’s Oct. 19 M3.1 Flash announcement previews a larger Pro model. Size and date are unpublished, and a 2.7T model Reuters reported in July has not been confirmed as the same release.
Cooling-aware planning estimates use inclusive facility-side factors for hosts, networking, cooling, and distribution. They are screening aids, not an IT-load conversion or engineering design; avoid applying PUE again to the figures below.
1.5 MW
Cooling-aware planning input
H100 GPUs (air)
~860
B300 GPUs (liquid)
~660
Example workloads
Dedicated single-tenant inference deployments
Inference capacity for smaller or quantized model configurations
Development, evaluation, or fine-tuning environments subject to model fit
Workloads that merit utility, engineering, and counterparty diligence
Sovereign-compute capacity planning subject to diligence
Phased build-outs with engineering validation of rack and cooling design
The structural advantage
Existing service and metering can make a former cultivation facility worth screening, but they do not establish usable capacity, a change-of-use path, interconnection rights, or a conversion schedule. Utility agreements, load studies, equipment condition, local approvals, and counterparty requirements must all be verified before a site can be treated as deployable.
As Neel Somani writes in Power 2026: “The real constraint to expanding AI capacity is power” — and “the real demand for power is now,” not years from now when greenfield plants finish construction.
Accelerator reference
Data center GPU landscape
Current and upcoming accelerators relevant to open-weight model deployment. Power per rack is climbing with each generation, and the newest racks require direct liquid cooling — making well-documented facility power more valuable over time.
Scroll horizontally to review all accelerator specifications.
Accelerator specifications and stated market status.
GPU
Architecture
VRAM
Bandwidth
Power
Status
H100 SXM5
Hopper
80 GB HBM3
3.35 TB/s
700W
Shipping
H200 SXM
Hopper
141 GB HBM3e
4.8 TB/s
700W
Shipping
B200 SXM
Blackwell
192 GB HBM3e
8 TB/s
1,000W
Shipping
B300 (HGX / DGX)
Blackwell Ultra
288 GB HBM3e
8 TB/s
14.5 kW per 8-GPU system
Shipping
GB300 NVL72
Blackwell Ultra
20 TB (rack)
576 TB/s
Up to 142 kW/rack
Shipping
MI300X
CDNA 3 (AMD)
192 GB HBM3
5.3 TB/s
750W
Shipping
MI355X
CDNA 4 (AMD)
288 GB HBM3E
8 TB/s
1,400W
Shipping
Vera Rubin NVL72
NVIDIA Rubin
288 GB HBM4/GPU
19.2 TB/s per GPU
~190–230 kW/rack (est.)
Shipping
Helios (MI455X)
CDNA 5 (AMD)
432 GB HBM4/GPU
Up to 23.3 TB/s
Not published
2H 2026
Vera Rubin Ultra
NVIDIA Rubin Ultra
Not final
Not final
Not published
2027 (NVIDIA)
Sources: product documentation from NVIDIA, AMD, and Supermicro, reviewed Oct. 12, 2026. Single-GPU rows list thermal design power or typical board power; system and rack rows list published maximum system power. NVIDIA has not published Vera Rubin NVL72 rack power; the range shown is an independent estimate for its two power profiles, and Supermicro sizes its liquid cooling for 227 kW per rack. AMD Helios is a partner reference design with volume deployments expected in 2H 2026. Industry reports in July 2026 said the Kyber rack planned for Rubin Ultra could slip to 2028; NVIDIA says its roadmap is intact. Upcoming specifications may change.
Why this matters
Open weights create infrastructure demand
Data sovereignty
Some regulated or sensitive workloads may need data-handling and residency options that make dedicated or self-managed deployments worth evaluating.
No deprecation risk
Hosting or self-managing a permitted model version can reduce dependence on a provider’s product lifecycle, subject to license terms, source availability, and ongoing operational support.
Customization
Fine-tuning, distillation, and quantization may benefit from reserved accelerator capacity; feasibility depends on the model, the workload, the provider’s offering, and operating requirements.
Cost at scale
At sustained utilization, buyers can compare dedicated deployments with hosted alternatives using their own workload, contract, utilization, and financing assumptions.
Deployment choice alone does not establish compliance, rights, availability, support, reliability, or operating controls. Buyers must confirm data-handling, contractual, residency, license, hardware, and service requirements.