* fix(inferx): remove stale/retired model TOMLs
* fix(inferx): rename model TOMLs to match InferX's exact dashboard model names
* feat(inferx): add 9 missing models currently live on InferX dashboard
* fix(inferx): correct schema validation errors in new model TOMLs (base_model links, reasoning_options, family enums, missing output limits)
* fix(inferx): remove unverified reasoning_options, document the one confirmed toggle
Per review feedback: reasoning_options=[{type=toggle}] was applied to
6 models (Agents-A1, Hy3-295B-NVFP4, Ornith-1.0-35B-FP8,
Step-3.7-Flash-NVFP4, deepseek-v4-flash, mimo-v25) without individual
verification. Only Qwen3.6-35B-A3B-FP8 was actually tested against
InferX's live API (chat_template_kwargs.enable_thinking).
- Set reasoning_options = [] on the 6 unverified models
- Added a sourced comment documenting the one verified toggle mechanism
* fix(inferx): add missing [cost] blocks, fix Devstral output limit
Per review feedback:
- Added [cost] input=0/output=0 to all 10 new models, matching the
pattern used by every existing InferX entry (still free tier)
- Fixed Devstral-2-123B-Instruct-2512-int4-AutoRound: context override
(128_000) left output inherited at 262_144 from base_model, exceeding
context. Added explicit output=128_000 override to match.
* fix(inferx): document verified reasoning toggle for deepseek-v4-flash
Tested both reasoning_effort (low/high — no measurable behavior
difference, ~2% token variance) and chat_template_kwargs.enable_thinking
(toggle — confirmed working, reasoning drops to null and completion
tokens drop ~70% when disabled). InferX supports the toggle mechanism,
not upstream DeepSeek's effort levels.
* fix(inferx): use preview's documented output limit for unpublished Hy3-295B-NVFP4
Model isn't live on InferX yet, so limit.output can't be verified via
API test. Using tencent/hy3-preview's documented 64_000 (same 256k
context) as a labeled estimate rather than context=output guess, until
real values can be confirmed post-publish.
* fix(inferx): correct verified reasoning/output limits based on live tests
* fix(inferx): remove unpublished Hy3, correct embedding output limit
* fix(inferx): document verified 27B toggle, move rationale comments to file headers
* fix(inferx): remove unpublished Step-3.7-Flash-NVFP4, verify output limits for deepseek-v4-flash and mimo-v25
* fix(inferx): restore deepseek-v4-flash reasoning toggle documentation lost in previous edit
Prashanth-InferX
295e59c6fd
