Sovereign Open‑Weight Model Ranking — Final Scoring & Analysis

Date: 2026‑03‑05 Scope: 17 candidates Output: HTML/CSS single page

A single, navigable page that preserves the complete scoring framework, the final ranked table, and the full model‑by‑model rationale. The evaluation lens is intentionally sovereignty‑maximalist (open weights, self‑hosting realism, forkability, and resistance to centralized stack capture).

Composite formula 0.25·C + 0.15·H + 0.30·S + 0.15·T + 0.10·L + 0.05·A Score range 0–100 Interpretation comparative inside this set

1. Scoring framework

The composite score is a weighted sum across six axes, each scored 0–100. Weights are fixed and applied identically to every candidate.

C — Capability / Intelligence (25%)
Reasoning, coding, tool-use/agentic robustness, multilingual competence, long-context stability.
H — Hardware / Self‑hosting viability (15%)
Practical runnability on sovereign hardware (laptop, phone, single GPU, commodity nodes), quantization availability, local tooling.
S — Sovereignty / Stack entanglement (30%)
Forkability, independence from state/hyperscaler control grids, likelihood of supply-chain pressure, and how “capturable” the ecosystem is.
T — Transparency / Base access (15%)
Availability of base checkpoints, technical reports, documented training choices, and practical modifiability for re-alignment.
L — License (10%)
Permissive open-weight licensing (MIT / Apache 2.0), commercial use, redistribution, and derivative rights.
A — Alignment / Gating friction (5%)
Degree of refusal behavior and safety scaffolding in common checkpoints; higher = fewer hard rails and easier re-tuning.
Anchor-SKU rule. Several entries are “families.” Scoring is anchored to a concrete, representative checkpoint (e.g., Trinity Mini, Ministral 3B, Jamba2 3B, Phi‑4), rather than averaging across every SKU.

2. Final ranking table

Composite = 0.25·C + 0.15·H + 0.30·S + 0.15·T + 0.10·L + 0.05·A. Higher composite implies higher suitability as a sovereign-root model inside this set.

Rank Model Composite C H S T L A
1 Trinity Mini (Arcee Trinity family) 93.09292909810090
2 Ministral 3B 85.58698758810072
3 Jamba2 3B 84.759097708810070
4 Ministral 8B 84.49090728810072
5 Ministral 14B 82.79280708810070
6 Jamba2 Mini 81.459370708810070
7 Phi‑4 80.59095608510060
8 Mistral Large 3 76.759745658510070
9 IBM Granite 4.0 Micro 75.558692459010065
10 Snowflake Arctic 71.49250508610060
11 Step‑3.5‑Flash 71.09675259510080
12 DeepSeek‑R1 70.559480259210075
13 Qwen 3.5 SLM family 68.39092208510065
14 GLM‑5 65.59940259010075
15 DeepSeek‑V3.2 65.259840259010075
16 DeepSeek‑V3.2‑Exp 65.09740259010075
17 Seed‑OSS‑36B 64.59445209510080
Why some frontier models rank lower (read)

Several entries (e.g., GLM‑5, DeepSeek‑V3.2, Step‑3.5‑Flash) score extremely high on capability but are penalized on S (stack entanglement) and H (datacenter‑level hardware) under sovereignty‑maximalist weighting.

3. Banded interpretation

A. Core sovereign edge band

Trinity Mini, Ministral 3B/8B/14B, Jamba2 3B/Mini, Phi‑4

Local-first, permissive weights, realistic runnability, strong modifiability, and comparatively higher forkability.

B. Frontier-adjacency band

Mistral Large 3, IBM Granite 4.0 Micro, Snowflake Arctic

Open weights and strong engineering; heavier infra gravity and/or enterprise governance orientation.

C. PRC frontier band (high-power, high-risk)

Step‑3.5‑Flash, DeepSeek (R1, V3.2, V3.2‑Exp), Qwen 3.5, GLM‑5, Seed‑OSS‑36B

Technically exceptional open weights; discounted on sovereignty due to stack entanglement and/or datacenter hardware requirements.

4. Model-by-model analysis

Each section includes: (1) the axis scores used in the composite, (2) a plain-language rationale, and (3) direct links to primary artifacts (model cards, licenses, blogs, technical reports, and deployment recipes) embedded inline.

#1 Trinity Mini (Arcee Trinity family) — Composite 93.0

Open-weight MoE designed for local/VPC deployment with unusually strong base-checkpoint availability and documentation.

C
92
Capability
H
92
Hardware
S
90
Sovereignty
T
98
Transparency
L
100
License
A
90
Gating
HF: Trinity Mini HF: Trinity‑Mini‑Base HF: GGUF arXiv: Technical report Arcee: open catalog

#2 Ministral 3B — Composite 85.5

Edge-first dense model with vision and large context in the Ministral 3 line; strong “everywhere deployment” profile.

C
86
Capability
H
98
Hardware
S
75
Sovereignty
T
88
Transparency
L
100
License
A
72
Gating
Mistral: Mistral 3 HF: collection HF: 3B Instruct HF: 3B Base vLLM recipe

#3 Jamba2 3B — Composite 84.75

On-device, long-context hybrid architecture tuned for reliable instruction following with Apache‑2.0 open weights.

C
90
Capability
H
97
Hardware
S
70
Sovereignty
T
88
Transparency
L
100
License
A
70
Gating
  • Release details: AI21 introduces the Jamba2 open-source family (Apache‑2.0; 256K context) (AI21 release blog).
  • Concrete anchor: Jamba2 3B model card emphasizes on-device deployment and reliability.
  • Docs: AI21 documentation covers Jamba2 Mini and Jamba2 3B roles and positioning (AI21 Jamba docs).
HF: Jamba2 3B AI21: release blog AI21: docs

#4 Ministral 8B — Composite 84.4

Balanced 8B dense model: strong capability with local viability; anchored to the official Instruct checkpoint.

C
90
Capability
H
90
Hardware
S
72
Sovereignty
T
88
Transparency
L
100
License
A
72
Gating
HF: 8B Instruct HF: GGUF HF: Reasoning Mistral: Mistral 3

#5 Ministral 14B — Composite 82.7

High-end dense member of Ministral 3: better reasoning and breadth at a higher local cost.

C
92
Capability
H
80
Hardware
S
70
Sovereignty
T
88
Transparency
L
100
License
A
70
Gating
HF: 14B Instruct HF: GGUF HF: collection Mistral: announcement

#6 Jamba2 Mini — Composite 81.45

MoE member of the Jamba2 family; higher capability than 3B at the cost of heavier deployment requirements.

C
93
Capability
H
70
Hardware
S
70
Sovereignty
T
88
Transparency
L
100
License
A
70
Gating
  • Anchor checkpoint: AI21‑Jamba2‑Mini (Apache‑2.0).
  • Release facts: AI21 lists family fast facts (Apache‑2.0, model sizes, 256K context) (AI21 blog).
  • Docs: model positioning and benefits in AI21 docs (Jamba docs).
HF: Jamba2 Mini AI21: release blog AI21: docs

#7 Phi‑4 — Composite 80.5

MIT-licensed small model with strong reasoning for size; high local viability; more corporate safety post-training.

C
90
Capability
H
95
Hardware
S
60
Sovereignty
T
85
Transparency
L
100
License
A
60
Gating
HF: phi‑4 HF: MIT license Microsoft: announcement HF: phi‑4 mini

#8 Mistral Large 3 — Composite 76.75

Frontier-scale open MoE (41B active / 675B total) under Apache‑2.0; exceptional capability with datacenter-level hardware needs.

C
97
Capability
H
45
Hardware
S
65
Sovereignty
T
85
Transparency
L
100
License
A
70
Gating
Mistral: announcement HF: Instruct HF: Base vLLM recipe

#9 IBM Granite 4.0 Micro — Composite 75.55

Apache‑2.0 open-weight enterprise-oriented family with strong docs and edge-friendly variants.

C
86
Capability
H
92
Hardware
S
45
Sovereignty
T
90
Transparency
L
100
License
A
65
Gating
  • Anchor checkpoint: granite‑4.0‑micro (Apache‑2.0).
  • Docs: Granite 4.0 documentation highlights edge workflows and Apache‑2.0 (IBM Granite docs).
  • Repo: Granite 4.0 language models repository (Apache‑2.0) (GitHub repo).
  • Base variant: micro base card includes training strategy description (micro‑base).
HF: micro HF: micro base IBM: docs GitHub: repo

#10 Snowflake Arctic — Composite 71.4

Open Apache‑2.0 enterprise MoE hybrid with strong documentation and a clear focus on data/SQL workflows.

C
92
Capability
H
50
Hardware
S
50
Sovereignty
T
86
Transparency
L
100
License
A
60
Gating
Snowflake: blog HF: instruct GitHub: repo

#11 Step‑3.5‑Flash — Composite 71.0

Open Apache‑2.0 MoE (~196B total / ~11B active) built for agentic workloads; high capability, discounted on sovereignty.

C
96
Capability
H
75
Hardware
S
25
Sovereignty
T
95
Transparency
L
100
License
A
80
Gating
  • Model card: Step‑3.5‑Flash (Apache‑2.0; BF16 weights; ~199B params).
  • Code repo: architecture overview and “11B active” framing (GitHub repo).
  • Technical paper: arXiv 2602.10604 describing the RL framework and eval results.
  • Base checkpoint access: mid-train base is published (Base‑Midtrain).
HF: model GitHub: repo arXiv: paper HF: base-midtrain

#12 DeepSeek‑R1 — Composite 70.55

MIT-licensed open reasoning model series; strong capability and modifiability, discounted for stack-entanglement risk.

C
94
Capability
H
80
Hardware
S
25
Sovereignty
T
92
Transparency
L
100
License
A
75
Gating
  • Model card: DeepSeek‑R1 explicitly states MIT licensing and derivative rights.
  • License file: MIT license text in-repo (LICENSE).
  • Official repo: DeepSeek’s GitHub release notes include distillation/derivative permissions (GitHub repo).
  • Policy pressure example: Reuters reporting on censorship-focused derivatives illustrates geopolitical alignment pressure (Reuters: R1‑Safe).
HF: model HF: license GitHub: repo Reuters: R1‑Safe

#13 Qwen 3.5 SLM family — Composite 68.3

Small open weights with strong performance and extensive tooling; discounted on sovereignty due to deep platform entanglement.

C
90
Capability
H
92
Hardware
S
20
Sovereignty
T
85
Transparency
L
100
License
A
65
Gating
  • Official blog: Qwen3.5 “Towards Native Multimodal Agents” (Qwen blog).
  • License statement: QwenLM states all open-weight Qwen3.5 models are Apache‑2.0 (GitHub: Qwen3.5).
  • Anchor small checkpoints: 0.8B, 2B, 4B, 9B.
  • Local packaging: Ollama model page indicates easy local pulls for small variants (Ollama: qwen3.5:2b).
Qwen: blog GitHub: Qwen3.5 HF: 9B HF: 4B HF: 2B HF: 0.8B

#14 GLM‑5 — Composite 65.5

Frontier-scale open-weight MoE (up to 744B total / 40B active) with MIT licensing; datacenter hardware profile.

C
99
Capability
H
40
Hardware
S
25
Sovereignty
T
90
Transparency
L
100
License
A
75
Gating
HF: model Z.ai: blog GitHub: repo

#15 DeepSeek‑V3.2 — Composite 65.25

MIT-licensed reasoning-first successor to V3; frontier-level capability with datacenter deployment characteristics.

C
98
Capability
H
40
Hardware
S
25
Sovereignty
T
90
Transparency
L
100
License
A
75
Gating
HF: model HF: license vLLM recipe DeepSeek: news

#16 DeepSeek‑V3.2‑Exp — Composite 65.0

Experimental sparse-attention version; strong capability, similar sovereignty profile to V3.2 with long-context efficiency focus.

C
97
Capability
H
40
Hardware
S
25
Sovereignty
T
90
Transparency
L
100
License
A
75
Gating
HF: model HF: license vLLM recipe Tom’s Hardware

#17 Seed‑OSS‑36B — Composite 64.5

Apache‑2.0 open dense 36B with long context and strong reasoning; penalized for heavier hardware and strong platform entanglement.

C
94
Capability
H
45
Hardware
S
20
Sovereignty
T
95
Transparency
L
100
License
A
80
Gating
Seed: blog HF: instruct GitHub: repo

End of model cards. The page deliberately embeds primary links within each model section to avoid a “links-only appendix”.

Built as a single HTML file with embedded CSS. No external libraries required.