feat(models): add Tencent Hy3 model

[vicovaro]

Summary

Adds Tencent's Hy3 model to the catalogue with provider mappings for DeepInfra and NovitaAI.

Hy3 is Tencent's 295B Mixture-of-Experts model (21B active params) with native 256K context and three reasoning modes, targeting complex reasoning and agentic tasks.

Provider mappings

Field	DeepInfra	NovitaAI
External ID	tencent/Hy3	tencent/hy3
Input price	$0.14/M	$0.14/M
Cached input price	$0.035/M	$0.035/M
Output price	$0.58/M	$0.58/M
Context size	262144	262144
Max output	131072	262144
Quantization	fp8	—
Streaming	yes	yes
Reasoning	yes (none, low, high)	yes (none, low, high)
Vision	no	no
Tools	yes	yes
JSON output	yes	yes

reasoningEfforts: ["none", "low", "high"] on both. DeepInfra's max output is 131072 (max_output_tokens from its model metadata; the page's max_tokens: 262144 is the full context window, not the output cap). Novita's max_output_tokens is 262144.

References

• https://deepinfra.com/tencent/Hy3
• https://novita.ai/models/model-detail/tencent-hy3
• https://openrouter.ai/tencent/hy3
• https://github.com/Tencent-Hunyuan/Hy3

🤖 Generated with Claude Code

Summary by CodeRabbit

• New Features
  • Added Tencent’s HY3 model to the available model catalog.
  • Added support for streaming, reasoning, tool use, vision, and JSON output capabilities.
  • Included provider availability, context limits, quantization, and pricing details.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 9217cef5-eac4-42d4-b733-b9adc136abec

📥 Commits

Reviewing files that changed from the base of the PR and between 0de5b76 and 2f0de7e.

📒 Files selected for processing (1)

• packages/models/src/models/tencent.ts

---

Walkthrough

Adds Tencent’s HY3 model with DeepInfra and Novita provider configurations, then registers the definitions in the shared models array.

Changes

Tencent model registry

Layer / File(s)	Summary
Define and register Tencent HY3 packages/models/src/models/tencent.ts, packages/models/src/models.ts	Defines HY3 metadata and provider capabilities for DeepInfra and Novita, then adds the Tencent model list to the exported registry.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related PRs

• theopenco/llmgateway#3188: Adds separate provider-specific model definitions to the shared models registry.

Suggested reviewers: steebchen

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the primary change: adding Tencent's Hy3 model to the models catalogue.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@packages/models/src/models/tencent.ts`:
- Around line 21-25: Add the shared reasoningEfforts values ["no_think", "low",
"high"] to both Hy3 model mappings alongside reasoning: true, and update the Hy3
request-building path to forward the selected reasoning effort as the
OpenAI-compatible reasoning_effort body field.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: a5ed714d-2cba-4e70-b992-2415bbc29939

📥 Commits

Reviewing files that changed from the base of the PR and between 1358c5e and ef591f0.

⛔ Files ignored due to path filters (1)

• pnpm-lock.yaml is excluded by !**/pnpm-lock.yaml

📒 Files selected for processing (2)

• packages/models/src/models.ts
• packages/models/src/models/tencent.ts

[steebchen]

Ran the scoped e2e suite for both mappings (TEST_MODELS="deepinfra/hy3,novita/hy3" FULL_MODE=true pnpm test:e2e).

One real catalogue bug found and fixed (pushed as 2f0de7e): Novita's tencent/hy3 does not support the json_object response format — it 400s with:

Model 'tencent/hy3' does not support 'json_object' response format. Supported formats: json_schema.

It does support json_schema. DeepInfra's tencent/Hy3 supports both. So the flags are now:

	jsonOutput	jsonOutputSchema
deepinfra	✅	✅
novita	—	✅

Both verified with direct provider calls, not just through the gateway.

Results after the fix

Tests 122 passed | 87 skipped — every chat/streaming/reasoning/tool-call/JSON/prompt-caching test passes for both mappings.

The only remaining failure is responses tool calls 'deepinfra/hy3' (404 on the stored-response GET), which is a pre-existing test-harness race, unrelated to this PR: responses.e2e.ts runs its tests concurrently, and the harness beforeEach calls clearCache() → redisClient.flushdb(), wiping the Redis-backed responses store out from under a concurrently in-flight test. It passes when hy3 runs alone, and it reproduces identically on models this PR doesn't touch (novita/gemma-4-31b-it failed the same assertion twice in two control runs).

Other metadata verified against the provider APIs

Prices ($0.14 / $0.035 cached / $0.58 per M), context 262144, max output (131072 DeepInfra / 262144 Novita), fp8 quantization, and the tools/reasoning/json capability flags all match what DeepInfra and Novita report. 👍