Skip to content

Model Family Catalog

Overview

This page is the human-readable rendering of contracts/model_family_catalog.json.

Use it to answer three distinct questions without weakening the public meaning of the support matrix:

  • What is supported as a public lane?
  • What families are implemented in code but not publicly supported?
  • What families or capabilities should be added next?

Support Tier vs Coverage State

Term Meaning Source of truth
support tier Public support/assurance posture for a declared lane. Values stay aligned with support_matrix.json. contracts/support_matrix.json
coverage state Repo implementation maturity outside declared public support lanes, such as profile_first_class, profile_shared_alias, auto_or_loader_only, loader_only, or backlog states. contracts/model_family_catalog.json
lifecycle classification Whether a lane, catalog family, or candidate is published, backlog, blocked, smoke-only, usage-only, or out of scope. contracts/model_classification.json

The support matrix remains strict. In this catalog, declared_support is the complete public published_basis support set and follows the support-matrix lane order. implemented_coverage is reserved for code-level visibility that does not itself create a public support claim. The model classification contract records lifecycle decisions and checkpoint status used by repo checks. Access-gated vendor checkpoints are not included as repo-shipped presets.

Catalog notes use a compact format: evidence surface first, then scope caveats only when the public evidence lane needs them. Exact metrics, hardware details, and rebuild instructions live in the public evidence artifacts and linked contracts.

Declared Support

These rows are the complete public published_basis support set from contracts/support_matrix.json, rendered in support-matrix order. A row here means the repo includes a public evidence fixture for that lane; the row notes state the evidence scope and non-goals.

Family State Representative models Notes
GPT-2 causal LM published_basis openai-community/gpt2 Evidence maps to support_matrix lane gpt2-causal-hf.
BERT / RoBERTa MLM published_basis bert-base-uncased, roberta-base Evidence maps to support_matrix lane bert-mlm-hf.
Mistral 7B causal LM published_basis mistralai/Mistral-7B-v0.1 Evidence maps to support_matrix lane mistral-7b-causal-hf and includes a container-backed public report, runtime manifest, and evidence pack.
Ministral 3 8B causal LM (text-only eval) published_basis mistralai/Ministral-3-8B-Instruct-2512-BF16 Evidence maps to support_matrix lane ministral-3-8b-text-causal-hf and includes a container-backed public report, runtime manifest, and evidence pack. The separate 14B Ministral lane has its own published evidence fixture.
Ministral 3 14B causal LM (text-only eval) published_basis mistralai/Ministral-3-14B-Instruct-2512-BF16 Evidence maps to support_matrix lane ministral-3-14b-text-causal-hf and includes a container-backed public report, runtime manifest, and evidence pack.
Qwen2 7B causal LM published_basis Qwen/Qwen2-7B Evidence maps to support_matrix lane qwen2-7b-causal-hf and includes a container-backed public report, runtime manifest, and evidence pack.
Qwen2.5 7B causal LM published_basis Qwen/Qwen2.5-7B Evidence maps to support_matrix lane qwen2-5-7b-causal-hf and includes a container-backed public report, runtime manifest, and evidence pack.
Qwen2.5 14B causal LM published_basis Qwen/Qwen2.5-14B Evidence maps to support_matrix lane qwen2-5-14b-causal-hf and includes a container-backed public report, runtime manifest, and evidence pack.
Qwen3 causal LM published_basis Qwen/Qwen3-8B Evidence maps to support_matrix lane qwen3-causal-hf and includes a container-backed public report, runtime manifest, and evidence pack.
DeepSeek-R1-Distill-Qwen causal LM published_basis deepseek-ai/DeepSeek-R1-Distill-Qwen-7B Evidence maps to support_matrix lane deepseek-r1-distill-qwen-causal-hf and includes a container-backed public report, runtime manifest, and evidence pack.
Phi-4 causal LM (text-only eval) published_basis microsoft/Phi-4-reasoning-plus Evidence maps to support_matrix lane phi-4-text-causal-hf and includes a container-backed public report, runtime manifest, and evidence pack. The public fixture is text-only and skips guard-overhead measurement by preset policy.
Gemma 4 E2B causal LM (text-only eval) published_basis google/gemma-4-E2B-it Evidence maps to support_matrix lane gemma4-e2b-text-causal-hf and includes a release-profile container-backed public report, runtime manifest, and evidence pack. Public support is text-only; image-text evaluation remains on the explicit hf_multimodal + vision_text path.
TinyLlama 1.1B causal LM published_basis TinyLlama/TinyLlama-1.1B-Chat-v1.0 Evidence maps to support_matrix lane tinyllama-1-1b-causal-hf and includes a container-backed public report, runtime manifest, and evidence pack.
OLMo 2 7B causal LM published_basis allenai/OLMo-2-1124-7B Evidence maps to support_matrix lane olmo-2-7b-causal-hf and includes a container-backed public report, runtime manifest, and evidence pack.
OLMo 2 13B causal LM published_basis allenai/OLMo-2-1124-13B-Instruct Evidence maps to support_matrix lane olmo-2-13b-causal-hf and includes a container-backed public report, runtime manifest, and evidence pack.
OpenLLaMA 7B causal LM published_basis openlm-research/open_llama_7b Evidence maps to support_matrix lane open-llama-7b-causal-hf and includes a release-profile container-backed public report, runtime manifest, and evidence pack.
Falcon 7B causal LM published_basis tiiuae/falcon-7b Evidence maps to support_matrix lane falcon-7b-causal-hf and includes a release-profile container-backed public report, runtime manifest, and evidence pack.
Qwen3.5 causal LM published_basis Qwen/Qwen3.5-9B Evidence maps to support_matrix lane qwen3-5-causal-hf and includes a container-backed public report, runtime manifest, and evidence pack.
Granite 4.1 3B causal LM published_basis ibm-granite/granite-4.1-3b Evidence maps to support_matrix lane granite-4-1-3b-causal-hf and includes a release-profile container-backed public report, runtime manifest, and evidence pack.
FLAN-T5 base seq2seq LM published_basis google/flan-t5-base Evidence includes release-profile container-backed public report, runtime manifest, signed evidence pack, and pinned CNN/DailyMail validation data through hf_seq2seq.
Gemma 4 12B any-to-any LM published_basis google/gemma-4-12B-it Evidence includes release-profile container-backed public VQAv2 image-text evidence. The no-op report passes strict policy with 0.565 final accuracy over 400 examples, ratio_vs_baseline 1.009, and no guard warnings. Gemma 4 12B uses the explicit hf_multimodal path and requires the multimodal runtime stack (transformers>=5.12.0 and torchvision>=0.26.0). This is no-op preservation/null-behavior evidence; audio and broader any-to-any behavior remain out of scope, and it is not guard-value proof.
Gemma 4 E4B image-text LM published_basis google/gemma-4-E4B-it Evidence includes release-profile container-backed public VQAv2 image-text evidence. The no-op report passes strict policy with 0.500 final accuracy over 400 examples, ratio_vs_baseline 1.010, and no guard warnings. This is no-op preservation/null-behavior evidence; audio and broader any-to-any behavior remain out of scope, and it is not guard-value proof.
Gemma 4 E2B image-text LM published_basis google/gemma-4-E2B-it Evidence includes release-profile container-backed public VQAv2 image-text evidence. The no-op report passes strict policy with 0.388 final accuracy over 400 examples, ratio_vs_baseline 1.013, and no guard warnings. This is no-op preservation/null-behavior evidence; audio and broader multimodal behavior remain out of scope, and it is not guard-value proof.
Qwen3.5 4B image-text LM published_basis Qwen/Qwen3.5-4B Evidence includes release-profile container-backed public VQAv2 image-text evidence. The no-op report passes strict policy with 0.855 final accuracy over 400 examples, ratio_vs_baseline 1.036, and no guard warnings. This is no-op preservation/null-behavior evidence; it is not guard-value proof.
Qwen3.5 2B image-text LM published_basis Qwen/Qwen3.5-2B Evidence maps to support_matrix lane qwen3-5-2b-image-text-hf and includes release-profile container-backed public VQAv2 image-text evidence, runtime manifest, and signed evidence pack. This is no-op preservation/null-behavior evidence, not guard-value proof.
Qwen3.5 27B image-text LM (scoped) published_basis Qwen/Qwen3.5-27B Evidence includes release-profile container-backed public VQAv2 image-text evidence using scoped self-attention and MLP guard scans. The no-op report passes strict policy with 0.8975 final accuracy over 400 examples and no guard warnings. Linear-attention module coverage remains captured as strict spectral-cap findings in the categorized larger-model lane outcomes; this is no-op preservation/null-behavior evidence and not guard-value proof.
Qwen3.6 27B image-text LM (scoped) published_basis Qwen/Qwen3.6-27B Evidence includes release-profile container-backed public VQAv2 image-text evidence using scoped self-attention and MLP guard scans. The no-op report passes strict policy with 0.8825 final accuracy over 400 examples and no guard warnings. Linear-attention module coverage remains captured as strict spectral-cap findings in the categorized larger-model lane outcomes; this is no-op preservation/null-behavior evidence and not guard-value proof.
Gemma 4 26B-A4B MoE image-text LM published_basis google/gemma-4-26B-A4B-it Evidence includes release-profile container-backed public VQAv2 image-text evidence. The no-op report passes strict policy with 0.555 final accuracy over 400 examples, ratio_vs_baseline 1.009, and no guard warnings. This is no-op preservation/null-behavior evidence; audio, exhaustive expert-bank behavior, and MoE routing quality remain out of scope, and it is not guard-value proof.
Gemma 4 31B image-text LM published_basis google/gemma-4-31B-it Evidence includes release-profile container-backed public VQAv2 image-text evidence. The no-op report passes strict policy with 0.610 final accuracy over 400 examples and no guard warnings. This is no-op preservation/null-behavior evidence; audio and broader any-to-any behavior remain out of scope, and it is not guard-value proof.
Mixtral 8x7B MoE causal LM published_basis mistralai/Mixtral-8x7B-v0.1 Evidence includes a Mixtral MoE causal LM published basis with explicit hf_causal preset, release-profile public report, runtime manifest, and signed evidence pack. The preset keeps the full guard scan surface while using BF16, automatic device placement, low-CPU-memory loading, disabled optional HF loading-info collection, and a skipped optional guard-overhead check for large-model resource discipline.
Qwen3 30B-A3B MoE causal LM published_basis Qwen/Qwen3-30B-A3B-Instruct-2507 Evidence includes a recent Qwen3 MoE causal LM published basis with explicit hf_causal preset, public WikiText-103 release-profile report, runtime manifest, and signed evidence pack. The fixture uses BF16/low-CPU-memory loading, disabled optional Hugging Face loading-info collection, automatic device placement, and scoped attention/router/shared-expert guard scans. This is no-op preservation evidence; it is not a benchmark-quality, fine-tuning, deployment, exhaustive expert-bank, or MoE routing-quality claim.
GPT-OSS 20B causal LM published_basis openai/gpt-oss-20b Evidence includes release-profile container-backed public WikiText-2 validation evidence for the pinned fixture. The no-op report passes strict policy with final perplexity 86.273 over 400 paired windows and no guard warnings. This is no-op preservation evidence; it is not a fine-tuning, benchmark-quality, deployment, safety, or alternate-seed/window robustness claim.
OLMoE 1B-active/7B-total causal LM published_basis allenai/OLMoE-1B-7B-0924 Evidence includes release-profile container-backed public report, runtime manifest, and signed evidence pack. This fixture is the smaller MoE validation basis; larger MoE assurance is covered by Mixtral, Qwen3 30B-A3B, and Gemma 4 26B-A4B published bases. Gemma 4 26B-A4B is image-text preservation/null-behavior evidence, not audio, exhaustive expert-bank, or MoE routing-quality evidence.
Ministral 3 3B causal LM (text-only eval) published_basis mistralai/Ministral-3-3B-Instruct-2512-BF16 Evidence maps to support_matrix lane ministral-3-3b-text-causal-hf and includes a container-backed public report, runtime manifest, and evidence pack.
Granite 4.1 8B causal LM published_basis ibm-granite/granite-4.1-8b Public release-profile container-backed evidence fixture is included.
SmolLM3 3B causal LM published_basis HuggingFaceTB/SmolLM3-3B Evidence includes release-profile container-backed public report, runtime manifest, signed evidence pack, and recorded 400-window preservation evidence. The preset explicitly skips guard-overhead measurement, so this is preservation evidence and not guard-overhead evidence.
Phi-4 mini causal LM published_basis microsoft/Phi-4-mini-instruct Evidence includes release-profile container-backed public report, runtime manifest, signed evidence pack, and recorded 400-window preservation evidence.
DeepSeek-R1-Distill-Qwen 14B causal LM published_basis deepseek-ai/DeepSeek-R1-Distill-Qwen-14B Evidence maps to support_matrix lane deepseek-r1-distill-qwen-14b-causal-hf and includes a container-backed public report, runtime manifest, and evidence pack; the fixture uses baseline-report reuse recovery.
DeepSeek-R1-0528-Qwen3 8B causal LM published_basis deepseek-ai/DeepSeek-R1-0528-Qwen3-8B Evidence maps to support_matrix lane deepseek-r1-0528-qwen3-8b-causal-hf and includes a container-backed public report, runtime manifest, and evidence pack.

Implemented Coverage

These rows are implementation, adapter, profile, or loader coverage that exists outside declared public support lanes. They are not public support claims unless they later move into declared_support and support_matrix.json.

Family Coverage state Representative models Notes
Llama profile_first_class TinyLlama/TinyLlama-1.1B-Chat-v1.0 Generic Llama-family profile handling is first-class. OpenLLaMA and TinyLlama now provide ungated declared support lanes, while access-gated vendor checkpoints remain omitted.
Qwen family aliases (Qwen1.5/Qwen2.5/Qwen3 naming) profile_first_class Qwen/Qwen2.5-14B, Qwen/Qwen3.5-9B, Qwen/Qwen3.5-4B, Qwen/Qwen3.5-27B, Qwen/Qwen3.6-27B Shared qwen-family heuristics cover aliases beyond the declared text-only Qwen2, Qwen2.5 14B, Qwen3, and Qwen3.5 9B public lanes. Qwen3.5 2B, Qwen3.5 4B, and scoped Qwen3.5/Qwen3.6 27B have published image-text evidence through hf_multimodal and pinned public VQAv2 materialization.
Yi profile_first_class 01-ai/Yi-34B Yi is treated as a RoPE decoder family in profile and adapter-auto logic.
Phi family profile_first_class microsoft/Phi-3-mini-4k-instruct, microsoft/Phi-4-reasoning-plus Phi-family selectors are first-class. Phi-4 now has a declared text-only lane via the reasoning-plus pilot, while multimodal Phi-4 remains backlog-only.
Gemma family profile_first_class google/gemma-4-E2B-it, google/gemma-4-E4B-it, google/gemma-4-12B-it, google/gemma-4-26B-A4B-it, google/gemma-4-31B-it Gemma-family selectors and loaders are first-class for compatible local or user-supplied checkpoints. Repo-declared published Gemma support includes Gemma 4 E2B text-only plus Gemma 4 E2B, E4B, 12B, 26B-A4B, and 31B image-text published-basis evidence.
OPT / GPT-NeoX / GPT-J profile_shared_alias EleutherAI/gpt-neox-20b These families are available through shared GPT-style loader and profile paths rather than dedicated public support lanes.
Falcon profile_shared_alias tiiuae/falcon-7b Falcon 7B now has a declared support lane; remaining Falcon-family coverage is available through adapter-auto heuristics and variant-path tests.
GLM auto_or_loader_only local-glm-compatible-checkpoint Family visibility currently comes from adapter-auto heuristics only.
DeepSeek profile_first_class deepseek-ai/DeepSeek-R1-Distill-Qwen-7B DeepSeek distill checkpoints continue to share the qwen-family route. Oversized FP8 checkpoint-specific repo hooks and shipped configs were removed after bring-up showed that they do not fit the supported hardware/runtime path.
Broader BERT-like MLMs (DistilBERT/ALBERT/DeBERTa/ELECTRA) auto_or_loader_only distilbert-base-uncased, microsoft/deberta-v3-base Loader and adapter-auto support exceeds the public BERT / RoBERTa support lane.
Broader seq2seq families (mBART/PEGASUS/Marian) auto_or_loader_only facebook/mbart-large-50 Concrete loader families exist beyond the public FLAN-T5 base seq2seq basis; mBART/PEGASUS/Marian still need dedicated presets and public evidence before they become declared support lanes.

Usage Only

These checkpoints appear in repo workflows or historical validation inputs, but they are not public support lanes.

Family State Representative models Notes
Qwen2.5 32B usage_only Qwen/Qwen2.5-32B Used in evidence-pack suites and validation defaults outside the declared Qwen2.5 14B support lane.
Yi 34B usage_only 01-ai/Yi-34B Used in workshop and full evidence-pack suites without a public support lane.

<=14B Text Candidate Inventory

This section summarizes the contract-tracked <=14B text and MLM candidates that sit outside, adjacent to, or are included in declared support.

It is a catalog view, not a run ledger. Exact criterion-by-criterion status and decision codes live under published_basis_candidates_text_le_14b in contracts/model_family_catalog.json.

Family Representative model Decision status Catalog location Notes
Qwen2.5 7B causal LM Qwen/Qwen2.5-7B published_basis_complete published_basis Evidence includes container-backed Qwen2.5 7B public report, runtime manifest, and signed evidence pack.
Qwen2.5 14B causal LM Qwen/Qwen2.5-14B published_basis_complete published_basis Evidence includes container-backed Qwen2.5 14B public report, runtime manifest, and signed evidence pack.
Qwen3 8B causal LM Qwen/Qwen3-8B published_basis_complete published_basis Evidence includes container-backed Qwen3 8B public report, runtime manifest, and signed evidence pack.
DeepSeek-R1-Distill-Qwen causal LM deepseek-ai/DeepSeek-R1-Distill-Qwen-7B published_basis_complete published_basis Evidence includes container-backed DeepSeek-R1-Distill-Qwen 7B public report, runtime manifest, and signed evidence pack.
Phi-4 reasoning-plus causal LM microsoft/Phi-4-reasoning-plus published_basis_complete published_basis Evidence includes container-backed Phi-4 reasoning-plus public report, runtime manifest, and signed evidence pack. This fixture is text-only and skips guard-overhead measurement by preset policy.
OpenLLaMA 7B causal LM openlm-research/open_llama_7b published_basis_complete published_basis Evidence includes release-profile container-backed public report, runtime manifest, and signed evidence pack.
Phi-3 Mini 4K Instruct causal LM microsoft/Phi-3-mini-4k-instruct explicitly_out_of_scope implemented_coverage The public Phi surface remains the shipped Phi-4 text-only lane for this wave.
Falcon 7B causal LM tiiuae/falcon-7b published_basis_complete published_basis Evidence includes release-profile container-backed public report, runtime manifest, and signed evidence pack.
Broader BERT-like MLMs (DistilBERT/ALBERT/DeBERTa/ELECTRA) distilbert-base-uncased blocked_missing_artifacts implemented_coverage Loader and adapter tests exist for representative DistilBERT and DeBERTa checkpoints, and the repo now ships a lane preset plus calibration config with dry-run sweep coverage, but approved calibration/evaluation evidence is still missing.
mBART large 50 seq2seq facebook/mbart-large-50 explicitly_out_of_scope implemented_coverage FLAN-T5 base now supplies the concrete public seq2seq basis; this broader mBART candidate stays out of scope until it has its own preset, calibration config, smoke evidence, and public run.

The machine-readable criterion-by-criterion ledger for this candidate set lives under published_basis_candidates_text_le_14b in contracts/model_family_catalog.json.

Priority Family Planned support mode Representative models Notes
P2 Audio-text evaluation pipeline phase2_audio_eval google/gemma-4-E2B-it, google/gemma-4-E4B-it Image-text evaluation is included. The remaining multimodal backlog item is audio-capable evaluation for the smaller Gemma 4 checkpoints.

Publication Criteria

A family only moves into support_matrix.json after all of the following are present:

  1. explicit adapter/profile recognition
  2. an included preset
  3. an included calibration config
  4. targeted tests
  5. CLI smoke evidence
  6. approved calibration/evaluation evidence