sync-models

Sync Models

Add descriptions for new models available in the HuggingFace router to chat-ui's prod.yaml and dev.yaml. Also flag models that support the OpenAI-compatible reasoning_effort parameter so chat-ui shows the thinking-effort selector for them, and enable artifacts for models with 32B or more total parameters. Finally, prune deprecated models — entries in the config whose ids the router no longer returns.

The comparison runs in two directions:

• Missing = in the router but not in the config → add them (steps 4–8).
• Deprecated = in the config but no longer in the router → remove them (step 9).

Both directions share the same source of truth: whatever https://router.huggingface.co/v1/models returns right now.

Workflow

1. Fetch models from router

   WebFetch https://router.huggingface.co/v1/models
   

   Extract all model IDs from the response.

2. Read current configuration

   • Read chart/env/prod.yaml (and chart/env/dev.yaml — the two share the same model set)
   • Extract model IDs from the MODELS JSON array in envVars

3. Identify missing and deprecated models Compare the router list against the config in both directions:

   • Missing = in router but not in the config → candidates to add (continue to step 4).
   • Deprecated = in the config but not in the router → candidates to remove (jump to step 9).

   Compute both sets with a script so the diff is exact rather than eyeballed — the list is long and near-duplicate ids (GLM-4.7 vs GLM-4.7-FP8, -Instruct vs -Thinking) are easy to miss:

   python3 - <<'EOF'
   import json, re, subprocess
   raw = subprocess.check_output(["curl","-sS","https://router.huggingface.co/v1/models"]).decode()
   router = {m["id"] for m in json.loads(raw)["data"]}
   txt = open("chart/env/prod.yaml").read()
   block = re.search(r"MODELS:\s*>\s*\n(.*?)\n\S", txt, re.S).group(1)
   ids = [e["id"] for e in json.loads(block)]
   keep = {"omni"}  # never treat the router alias (PUBLIC_LLM_ROUTER_ALIAS_ID) as deprecated
   print("MISSING (add):   ", sorted(router - set(ids)))
   print("DEPRECATED (rm): ", [i for i in ids if i not in router and i not in keep])
   EOF
   

   Only operate on the missing set for the add/research steps (4–8). Never edit, re-flag, or re-describe entries that already exist in prod.yaml / dev.yaml — even if you think their reasoning capability or description could be improved. Existing entries are intentionally curated and may have been hand-tuned for known quirks. Out of scope unless the user explicitly asks for a re-audit.

   Exclude the router alias from the deprecated set. The omni entry (whatever PUBLIC_LLM_ROUTER_ALIAS_ID is set to) is a synthetic alias, not a router model, so it never appears in /v1/models. Never remove it. The same goes for any other intentionally-synthetic id that isn't meant to come from the router.

4. Research each missing model For each missing model, search the web for its specifications:

   • Model architecture (dense, MoE, parameters)
   • Key capabilities (coding, reasoning, vision, multilingual, etc.)
   • Target use cases
   • Whether it's a reasoning model (see step 5)

5. Decide if the model is reasoning-capable The supportsReasoning flag controls TWO behaviors, and both must be appropriate before flagging:

   1. chat-ui renders the thinking-effort dropdown and forwards reasoning_effort to the router;
   2. chat-ui echoes the model's prior reasoning back as reasoning_content on past assistant messages (cross-turn "preserved thinking" — see prepareFiles.ts / endpointOai.ts).

   A model qualifies for (1) if it accepts the OpenAI-style reasoning_effort: low|medium|high parameter via the HF router and meaningfully changes its chain-of-thought depth in response. Whether that holds depends on both the model and the providers serving it — the router is a transparent proxy, so behavior comes from each provider's implementation. Don't decide from the name alone.

   For (2), check the vendor's preserved-thinking / multi-turn guidance, because it can point in either direction:

   • Flag-strengthening: the vendor documents that prior reasoning_content must or should be passed back in multi-turn or tool-calling flows. Known examples: Moonshot thinking.keep (Kimi K2.6+/K3), MiniMax "Interleaved Thinking" ("must preserve the model's thinking content completely"), DeepSeek V4 thinking mode (hard 400 if reasoning_content is missing on tool-call turns), Z.ai "Preserved Thinking" (clear_thinking: false), Qwen3.6 preserve_thinking.
   • Flag-blocking: the vendor documents that historical thoughts must be STRIPPED across completed turns. Known example: the Gemma family — Google requires removing thoughts from previous turns ("historical model output must only include the final response") while preserving them only inside a single turn's tool loop, which chat-ui handles automatically without the flag. Do not flag such models even though they emit reasoning and may accept an effort knob — flagging would make chat-ui echo reasoning the vendor says to strip.

   Vendor doc entry points for the preserved-thinking check: https://platform.kimi.ai/docs/guide/use-kimi-k2-thinking-model, https://platform.minimax.io/docs/guides/text-m3-function-call, https://api-docs.deepseek.com/guides/thinking_mode/, https://docs.z.ai/guides/capabilities/thinking-mode, https://ai.google.dev/gemma/docs/capabilities/thinking, plus the model card's own multi-turn/best-practices section (Qwen cards document what their chat template does with historical <think> blocks).

   Heuristic shortlist (candidates worth verifying):

   • Name contains gpt-oss, -Thinking, -thinking, -Reasoning, -reasoning, QwQ, R1, MiniMax-M, Kimi-K2-Thinking, cogito-
   • Hybrid models with a thinking switch: DeepSeek V3.1+, GLM-4.5 / 4.6 / 4.7 / 5.x, Qwen3 thinking variants
   • Model card mentions "thinking mode", "reasoning traces", "extended thinking", "test-time compute", or shows <think>...</think> examples

   Skip without further checking:

   • Generic "good at reasoning" marketing copy — every modern LLM claims this. Only flag when reasoning is the mode of operation.
   • Non-thinking siblings (Qwen3-235B-A22B-Instruct-2507 ≠ Qwen3-235B-A22B-Thinking-2507).
   • Translation / vision-only / guard / coder-only models with no documented thinking mode.

   Verify each candidate via provider docs before flagging:

   For each model on the heuristic shortlist, look up its live providers in the /v1/models payload, then check those providers' chat-completions documentation for reasoning_effort, reasoning_content, enable_thinking, or a thinking parameter. If at least one live provider documents it for this model (or for the model family in general), flag it as reasoning-capable. The HF router will proxy the parameter to whichever provider it picks.

   Provider docs to consult (use WebFetch / WebSearch):

   • fireworks-ai: https://docs.fireworks.ai/api-reference/post-chatcompletions
   • groq: https://console.groq.com/docs/reasoning
   • cerebras: https://inference-docs.cerebras.ai/capabilities/reasoning
   • together: https://docs.together.ai/docs/gpt-oss
   • novita: https://novita.ai/docs/guides/llm-interleaved-thinking
   • sambanova: search "sambanova reasoning_effort"
   • deepinfra: https://docs.deepinfra.com/chat/overview
   • nscale, scaleway, ovhcloud, hyperbolic, zai-org, cohere, featherless-ai: search " reasoning_effort" or check their model catalog pages

   If none of the live providers document reasoning support for the model, don't flag it — even if the name pattern-matches. If documentation is ambiguous, lean toward not flagging and mention it in the commit so it can be revisited.

6. Decide if the model gets artifacts Enable artifacts for any new model with 32B or more total parameters by appending "supportsArtifacts": true to its entry. This makes chat-ui instruct the model to emit <artifact> blocks rendered in the side panel.

   • Use the total parameter count, not active parameters. A 35B-A3B MoE qualifies (35B total ≥ 32B) even though only 3B are active.
   • The count is usually in the model name (Qwen3.6-27B, 550B-A55B). When it isn't, use the parameter count found while researching the model in step 4.
   • This is independent of reasoning capability — a model can have both flags, either one, or neither.
   • Models under 32B don't get the flag; users can still enable artifacts per-model via settings overrides.

7. Write descriptions Match existing style:

   • 8-12 words
   • Sentence fragments (no period needed)
   • No articles ("a", "the") unless necessary
   • Focus on: architecture, specialization, key capability

   Examples:

   • "Flagship GLM MoE for coding, reasoning, and agentic tool use."
   • "MoE agent model with multilingual coding and fast outputs."
   • "Vision-language Qwen for documents, GUI agents, and visual reasoning."
   • "Mobile agent for multilingual Android device automation."

8. Add new models to both files Add new models at the TOP of the MODELS array in:

   • chart/env/prod.yaml
   • chart/env/dev.yaml

   Base format:

   { "id": "org/model-name", "description": "Description here." }
   

   Append "supportsReasoning": true for reasoning-capable models (step 5) and "supportsArtifacts": true for 32B+ models (step 6). A model can carry both:

   {
   	"id": "org/model-name",
   	"description": "Description here.",
   	"supportsReasoning": true,
   	"supportsArtifacts": true
   }
   

   supportsReasoning makes chat-ui render the Thinking-effort dropdown in the chat footer, forward reasoning_effort to the router, AND echo the model's prior reasoning back as reasoning_content on past assistant messages (preserved thinking). Models whose vendor requires stripping historical thoughts (Gemma family) must stay unflagged — see step 5. supportsArtifacts enables the artifacts side panel for the model.

9. Remove deprecated models from both files Delete the full entry line for every id in the deprecated set (step 3) from both chart/env/prod.yaml and chart/env/dev.yaml. Match on the exact "id" value so near-duplicate ids aren't removed by accident, and keep the removal symmetric — the two files must end with the same model set.

   Removing a deprecated entry is safe and low-risk: MODELS is an overrides map, not the model list. src/lib/server/models.ts builds the catalog from the router's /v1/models response and only applies a MODELS entry when its id is present in that response (it maps over the router models and looks each up in the override map). An entry whose id the router no longer serves is a dead override — it never renders in the UI — so pruning it changes nothing at runtime; it just keeps the config honest and readable.

   After editing, re-parse the MODELS block in each file as JSON to confirm it's still valid and that no deprecated id remains (reuse the script from step 3 — the deprecated set should now be empty).

   Do not touch models that are merely referenced by env vars but still present in the router (e.g. TASK_MODEL, LLM_ROUTER_TOOLS_MODEL, LLM_ROUTER_MULTIMODAL_MODEL). If a deprecated id is referenced by one of those env vars, stop and flag it to the user instead of silently removing it — that indicates a config that needs a replacement model, not just a pruned line.

10. Commit changes In the commit message, mention how many models were added (and how many of those are reasoning-capable / get artifacts) and how many deprecated models were removed, so it's easy to review.

    git add chart/env/prod.yaml chart/env/dev.yaml
    git commit -m "feat: sync models from router (+N added, M reasoning-capable, K artifacts, -D removed)"
    

Notes

• Prune-only runs are normal: the router may drop models without adding any. If the missing set is empty but the deprecated set isn't, skip steps 4–8 entirely and just do the removal (step 9) plus commit.
• Add vs. remove are judged per-id: the router routinely keeps a base model while dropping its FP8/variant sibling (or the reverse). Treat GLM-4.7 and GLM-4.7-FP8, or -Instruct and -Thinking, as independent ids — presence of one says nothing about the other. Trust the script's diff, not the family name.
• FP8 variants: describe as "FP8 [base model] for efficient inference with [key capability]". If the base model is reasoning-capable, the FP8 variant is too — flag both. Same for artifacts: quantization doesn't change the parameter count, so a 32B+ base means the FP8 variant gets supportsArtifacts too.
• Vision models: mention "vision-language" and key visual tasks. A vision model can still be reasoning-capable (e.g. Qwen3-VL-*-Thinking) — judge by the same rules.
• Agent models: mention "agent" and automation capabilities.
• Regional models: mention language focus (e.g., "European multilingual", "Southeast Asian").