 Craig Donnelly
|
1a4e693bb9
|
Add TensorX provider with 32 models (#2696)
* Add TensorX provider with 32 models
TensorX is an EU-sovereign OpenAI-compatible inference platform
(https://tensorx.ai) offering 40+ open-source and frontier models.
This adds:
- providers/tensorx/provider.toml (OpenAI-compatible, api.tensorx.ai/v1)
- providers/tensorx/logo.svg
- 32 chat model TOMLs across 8 vendors:
- z-ai (7 GLM models, base_model from zhipuai metadata)
- deepseek (7 models incl. V4 Flash/Pro, R1, Chat V3 variants)
- minimax (5 M2/M3 variants)
- moonshotai (3 Kimi K2 models)
- qwen (5 models incl. Qwen3.5, VL, Coder)
- meta-llama (2 Llama models)
- nvidia (1 Nemotron model)
- openai (2 GPT-OSS models)
21 models use base_model inheritance from existing models/ metadata.
11 models have full provider TOML definitions.
Non-chat models (embedding, TTS, STT), internal aliases, and duplicate
entries are excluded. Pricing, limits, and capabilities are sourced from
the TensorX API (https://api.tensorix.ai/v1/model/info).
* Address PR #2696 review feedback
- Fix logo viewBox origin (25 15 -> 0 0) to match repo convention
- Add provider-audited reasoning_options to all 6 reasoning models
(gpt-oss-120b/20b, deepseek-r1-0528/v3.2/chat-v3.1, qwen3.5-9b)
values audited against TensorX API; mandatory-reasoning models
omit 'none' (gpt-oss, deepseek-r1-0528)
- Correct GPT-OSS release_date from 2024-12-01 to 2025-08-05
- Factor Qwen3.5-9B through canonical base_model; create
models/alibaba/qwen3.5-9b.toml with provider-agnostic facts.
Fixes attachment/modalities (was text-only; HF confirms VL model
with image/video/audio input)
* Address second-round review feedback on PR #2696
- Remove fixed width/height from logo.svg per logo guidelines
- Reconcile model list with the advertised TensorX catalog: drop 9
delisted models (deepseek-chat-v3-0324, llama-3.3-70b, llama-4-maverick,
minimax-m2/m2.1/m2.7, gpt-oss-20b, qwen-2.5-72b, glm-4.6)
- Use canonical base_model for openai/gpt-oss-120b and
deepseek/deepseek-r1-0528; inherit canonical limits
- Add missing description to remaining inline models
- Declare reasoning_options on every reasoning model, audited per-model
against the TensorX API (validation errors + behavioural probes);
exact request syntax recorded as TOML comments
- Remove unsupported audio input modality from Qwen3.5 9B canonical
|
2026-08-01 10:55:51 -05:00 |
|