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

TermMeaningSource of truth
support tierPublic support/assurance posture for a declared lane. Values stay aligned with support_matrix.json.contracts/support_matrix.json
coverage stateRepo implementation maturity outside the public support matrix, such as profile_first_class, profile_shared_alias, auto_or_loader_only, loader_only, or backlog states.contracts/model_family_catalog.json

The support matrix remains strict. The model family catalog is broader by design and records code-level visibility, usage-only checkpoints, and recommended additions. Access-gated vendor checkpoints are intentionally kept out of declared support lanes and included preset inventory.

Declared Support

FamilyStateRepresentative modelsNotes
GPT-2 causal LMpublished_basisopenai-community/gpt2Public lane derived from gpt2-causal-hf.
BERT / RoBERTa MLMpublished_basisbert-base-uncased, roberta-basePublic lane derived from bert-mlm-hf.
Mistral 7B causal LMsupported_experimentalmistralai/Mistral-7B-v0.1Pilot preset and calibration config are included.
Ministral 3 causal LM (text-only eval)supported_experimentalmistralai/Ministral-3-8B-Instruct-2512-BF16, mistralai/Ministral-3-14B-Instruct-2512-BF16Text-only pilot presets and calibration configs are included for both 8B and 14B checkpoints.
Qwen2 7B causal LMsupported_experimentalQwen/Qwen2-7BPilot preset and calibration config are included.
Qwen2.5 7B causal LMsupported_experimentalQwen/Qwen2.5-7BPilot preset and calibration config are included.
Qwen2.5 14B causal LMsupported_experimentalQwen/Qwen2.5-14BPilot preset and calibration config are included.
Qwen3 causal LMsupported_experimentalQwen/Qwen3-8BPilot preset and calibration config are included.
DeepSeek-R1-Distill-Qwen causal LMsupported_experimentaldeepseek-ai/DeepSeek-R1-Distill-Qwen-7BPilot preset and calibration config are included.
Phi-4 causal LM (text-only eval)supported_experimentalmicrosoft/Phi-4-reasoning-plusText-only pilot preset and calibration config are included. Current HF runtime validation closes cleanly when the lane opts into trust_remote_code.
Gemma 4 E2B causal LM (text-only eval)supported_experimentalgoogle/gemma-4-E2B-itText-only pilot preset and calibration config are included. Image-text evaluation uses the explicit hf_multimodal + vision_text path.
TinyLlama 1.1B causal LMsupported_experimentalTinyLlama/TinyLlama-1.1B-Chat-v1.0Ungated Llama-family pilot lane with included preset and calibration config.
OLMo 2 causal LMsupported_experimentalallenai/OLMo-2-1124-7B, allenai/OLMo-2-1124-13B-InstructPilot presets and calibration configs are included for both 7B and 13B scale points.
Qwen3.5 causal LMsupported_experimentalQwen/Qwen3.5-9BPilot preset and calibration config are included.
Seq2Seq / local pairscommunity_experimentalt5-small, facebook/bart-baseGeneric seq2seq lane without a published-basis claim.

Implemented Coverage

FamilyCoverage stateRepresentative modelsNotes
Mixtralprofile_first_classmistralai/Mixtral-8x7B-v0.1Profile and loader code recognize the family directly.
Llamaprofile_first_classopenlm-research/open_llama_7b, TinyLlama/TinyLlama-1.1B-Chat-v1.0Generic Llama-family profile handling is first-class. TinyLlama provides the ungated declared support lane, while access-gated vendor checkpoints remain omitted.
Qwen family aliases (Qwen1.5/Qwen2.5/Qwen3 naming)profile_first_classQwen/Qwen2.5-14B, Qwen/Qwen3.5-9BShared qwen-family heuristics cover aliases beyond the declared Qwen2, Qwen2.5 14B, Qwen3, and Qwen3.5 lanes, including usage-only Qwen2.5 checkpoints.
Yiprofile_first_class01-ai/Yi-34BTreated as a RoPE decoder family in profile logic.
Phi familyprofile_first_classmicrosoft/Phi-3-mini-4k-instruct, microsoft/Phi-4-reasoning-plusDedicated phi-family selectors exist. Phi-4 has a declared text-only lane, while multimodal Phi-4 remains backlog-only.
Gemma familyprofile_first_classgoogle/gemma-3-4b-it, google/gemma-4-E2B-itGemma 3/4 selectors and loaders are first-class. Gemma 4 E2B has a declared text-only lane, image-text evaluation uses hf_multimodal + vision_text, and audio remains deferred.
OPT / GPT-NeoX / GPT-Jprofile_shared_aliasfacebook/opt-1.3b, EleutherAI/gpt-neox-20bAvailable through shared GPT-style paths.
GPT-OSSprofile_first_classopenai/gpt-oss-20bDedicated profile selectors and HF causal decoder spec now cover the open-weight checkpoint directly.
Falconauto_or_loader_onlytiiuae/falcon-7bVisible through adapter-auto heuristics only.
GLMauto_or_loader_onlyTHUDM/glm-4-9b-chatVisible through adapter-auto heuristics only.
DeepSeekprofile_first_classdeepseek-ai/DeepSeek-R1-Distill-Qwen-7BDeepSeek distill checkpoints share the qwen-family route. Oversized FP8 checkpoint-specific repo hooks and included configs are omitted because they do not fit the supported hardware/runtime path.
Broader BERT-like MLMs (DistilBERT/ALBERT/DeBERTa/ELECTRA)auto_or_loader_onlydistilbert-base-uncased, microsoft/deberta-v3-baseLoader/auto support exceeds the public BERT / RoBERTa lane.
Broader seq2seq families (mBART/PEGASUS/Marian)auto_or_loader_onlyfacebook/mbart-large-50, Helsinki-NLP/opus-mt-en-deLoader support is broader than the generic seq2seq public lane.

Usage Only

FamilyStateRepresentative modelsNotes
Qwen2.5 32Busage_onlyQwen/Qwen2.5-32BUsed in evidence-pack suites and validation defaults outside the declared Qwen2.5 14B support lane.
Qwen1.5 72Busage_onlyQwen/Qwen1.5-72BUsed concretely in evidence-pack suites.
Yi 34Busage_only01-ai/Yi-34BUsed in workshop and full evidence-pack suites.
Mixtral 8x7Busage_onlymistralai/Mixtral-8x7B-v0.1Used in evidence-pack flows 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 have recently graduated into declared support.

It is a catalog view, not a run ledger. Exact criterion-by-criterion status and decision codes live under promotion_candidates_text_le_14b in contracts/model_family_catalog.json.

FamilyRepresentative modelPromotion statusCatalog locationNotes
Qwen2.5 7B causal LMQwen/Qwen2.5-7Bpromote_nowsupported_experimentalThis candidate already graduated into declared support and remains listed here because the contract tracks the reviewed candidate set.
OpenLLaMA 7B causal LMopenlm-research/open_llama_7bblocked_missing_artifactsimplemented_coverageRecognition and targeted tests exist, and the repo ships a lane preset plus calibration config with dry-run sweep coverage, but approved calibration/evaluation evidence is still missing.
Phi-3 Mini 4K Instruct causal LMmicrosoft/Phi-3-mini-4k-instructexplicitly_out_of_scopeimplemented_coverageThe current declared Phi support surface remains the shipped Phi-4 text-only lane.
Gemma 3 4B ITgoogle/gemma-3-4b-itexplicitly_out_of_scopeimplemented_coverageThe broader Gemma family remains multimodal/audio-capable, so this text-only candidate is not part of the tracked promotion set.
OPT 1.3B causal LMfacebook/opt-1.3bblocked_missing_artifactsimplemented_coverageShared GPT-style recognition exists, and the repo ships a lane preset plus calibration config with dry-run sweep coverage, but approved calibration/evaluation evidence is still missing.
Falcon 7B causal LMtiiuae/falcon-7bblocked_missing_artifactsimplemented_coverageRecognition and targeted variant-path tests exist, and the repo ships a lane preset plus calibration config with dry-run sweep coverage, but approved calibration/evaluation evidence is still missing.
GLM 4 9B ChatTHUDM/glm-4-9b-chatblocked_missing_artifactsimplemented_coverageRecognition and targeted variant-path tests exist, and the repo ships a lane preset plus calibration config with dry-run sweep coverage, but approved calibration/evaluation evidence is still missing.
Broader BERT-like MLMs (DistilBERT/ALBERT/DeBERTa/ELECTRA)distilbert-base-uncasedblocked_missing_artifactsimplemented_coverageLoader and adapter tests exist for DistilBERT and DeBERTa, and the repo ships a lane preset plus calibration config with dry-run sweep coverage, but approved calibration/evaluation evidence is still missing.
mBART large 50 seq2seqfacebook/mbart-large-50explicitly_out_of_scopeimplemented_coverageGeneric seq2seq and community lanes remain outside this tracked candidate set.

The machine-readable criterion-by-criterion ledger for this candidate set lives under promotion_candidates_text_le_14b in contracts/model_family_catalog.json.

PriorityFamilyPlanned support modeRepresentative modelsNotes
P2Audio-text evaluation pipelinephase2_audio_evalgoogle/gemma-4-E2B-it, google/gemma-4-E4B-itImage-text evaluation is included. Audio-capable evaluation for the smaller Gemma 4 checkpoints remains deferred.

Promotion 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