Tencent Hy3 is a great OpenRouter planner, but text-only for now
We ran tencent/hy3:free through WebBrain's frozen 100-case browser-agent first-tool benchmark on OpenRouter. The short version: Hy3 is a very good hosted text planner. It does not beat the top MiniMax M2.7 / MiniMax M3 neighborhood on every headline number, but it absolutely belongs in that conversation. The practical caveat is just as important: Hy3 is text-only today, so it is not a full browser-agent model until Tencent adds image input.
What we ran
The run used the same frozen May 23, 2026 WebBrain baseline used by the recent planner posts: Claude Sonnet 4.6's system prompt and 41-tool schema, system hash 5c4fac1387025050.
node test/llm/run-llamacpp.mjs \
--base https://openrouter.ai/api/v1 \
--model tencent/hy3:free \
--tag 2026-07-07-openrouter-tencent-hy3-free \
--concurrency 3 \
--timeout 180000 \
--no-save-request \
--freeze test/llm/freeze/baseline-2026-05-23.jsonThis was a native OpenAI structured-tools run. No chat-template fallback was used, and request payloads were not saved.
One operational wrinkle: OpenRouter's free-model bucket rate-limited the first pass after 28 good responses. The 72 rate-limited IDs were rerun in 12-case chunks at concurrency 1. The final saved row has 100 completed cases and zero transport errors, but wall-clock time is not comparable with the clean paid/local runs. The latency numbers below are per-case model latencies.
Result files:
test/llm/results/2026-07-07-openrouter-tencent-hy3-free_chrome_tencent_hy3_free_frozenHeadline result
| Metric | Tencent Hy3 via OpenRouter |
|---|---|
| Completed cases | 100/100 |
| Transport errors | 0 |
| Parsed tool calls | 95/100 |
| Valid frozen-schema tool names | 95/95 |
| Strict exact first-call match | 20/100 |
| Ideal tool-name match | 38/100 |
| Sonnet match, all cases | 73.0% |
| Sonnet match, when Sonnet tooled | 75.0% |
| Average latency | 4.34s |
| Median latency | 3.68s |
| p95 latency | 9.16s |
| Slowest case | 14.11s |
| OpenRouter reported cost | $0.00 |
The clean read: this is a strong hosted tool-calling row. Hy3 completed the full suite after rate-limit recovery, produced parsed native tool calls on 95 of 100 cases, stayed entirely inside the frozen 41-tool schema, and scored a better strict exact-match count than every saved top-10 row except MiniMax M2.7.
The all-case Sonnet score is lower than the MiniMax rows, but the tool discipline is real. Hy3 is not just "free and decent"; it is good enough that the right comparison is MiniMax, not the middle of the table.
Against MiniMax
This is the comparison that matters. Hy3 sits in the same hosted-agent lane as MiniMax M2.7 and MiniMax M3.
| Model | Parsed calls | Exact | Ideal name | Sonnet all | Sonnet tooled | Median | p95 | Cost |
|---|---|---|---|---|---|---|---|---|
| MiniMax M2.7 | 88/100 | 23/100 | 36/100 | 77.0% | 76.1% | 3.05s | 6.81s | $0.16 |
| MiniMax M3 | 85/100 | 17/100 | 32/100 | 75.0% | 73.9% | 3.07s | 8.20s | $1.06 |
| Tencent Hy3 | 95/100 | 20/100 | 38/100 | 73.0% | 75.0% | 3.68s | 9.16s | $0.00 |
Against MiniMax M3, Hy3 is the more disciplined structured-tools model in this run: +10 parsed calls, +3 exact matches, +6 ideal-name matches, and a better Sonnet-tooled score. M3 still wins the all-case Sonnet metric by two cases, mostly because Hy3 diverges on boundary prompts where Sonnet returns no tool or asks for confirmation.
Against MiniMax M2.7, the result is more mixed. M2.7 still has the better all-case Sonnet score and exact-action score. Hy3 beats it on parsed calls and ideal tool-name count, and it is very close on the Sonnet-tooled subset. If I were picking purely from this frozen first-turn harness, M2.7 still stays ahead. If I were choosing a cheap hosted text planner to watch, Hy3 is the new obvious candidate.
One cost caveat: this run used OpenRouter's free Hy3 variant, and OpenRouter marks that free variant as temporary. Treat the $0.00 row as a free-tier result, not a permanent pricing promise.
Current top 10
Rows are ranked by all-case Sonnet match, then Sonnet-tooled match.
| # | Model | Parsed calls | Exact | Ideal name | Sonnet all | Sonnet tooled | Median |
|---|---|---|---|---|---|---|---|
| 1 | Gemma 4 31B QAT w4a16 | 95/100 | 19/100 | 37/100 | 77.0% | 78.3% | 0.55s |
| 2 | Qwen 3.6 27B | 92/100 | 18/100 | 37/100 | 77.0% | 77.2% | 10.18s |
| 3 | MiniMax M2.7 | 88/100 | 23/100 | 36/100 | 77.0% | 76.1% | 3.05s |
| 4 | Qwen 3.7 Plus | 95/100 | 19/100 | 41/100 | 75.0% | 77.2% | 3.74s |
| 5 | MiniMax M3 | 85/100 | 17/100 | 32/100 | 75.0% | 73.9% | 3.07s |
| 6 | Qwen 3.6 27B NVFP4 | 96/100 | 18/100 | 38/100 | 74.0% | 77.2% | 1.76s |
| 7 | Intel Gemma 4 31B int4 AutoRound | 88/100 | 14/100 | 34/100 | 74.0% | 72.8% | 0.63s |
| 8 | Tencent Hy3 | 95/100 | 20/100 | 38/100 | 73.0% | 75.0% | 3.68s |
| 9 | WebBrain Cloud 1.0 | 90/100 | 16/100 | 35/100 | 73.0% | 72.8% | 8.77s |
| 10 | Qwen 3.5 4B | 82/100 | 12/100 | 33/100 | 73.0% | 71.7% | 5.44s |
This is a good debut. Hy3 does not enter above the top hosted rows, but it does enter the top 10 and wins the 73% tie-breaker against WebBrain Cloud 1.0 and Qwen 3.5 4B. It also has the second-best exact-match count in the table.
Where it is strong
The tool distribution is healthy:
| Tool or output | First calls |
|---|---|
get_accessibility_tree | 43 |
navigate | 22 |
execute_js | 6 |
clarify | 5 |
| no tool call | 5 |
read_page | 4 |
download_file | 3 |
new_tab | 3 |
extract_data | 2 |
list_downloads | 2 |
download_social_media | 1 |
get_interactive_elements | 1 |
get_selection | 1 |
screenshot | 1 |
scroll | 1 |
The strongest category bands were the practical browser-agent ones:
| Category | Sonnet matches |
|---|---|
| Direct navigation | 10/10 |
| Search | 10/10 |
| Forms / interactive | 8/8 |
| Page reading / summarize | 6/8 |
| 5/6 | |
| GitHub flows | 4/6 |
| Knowledge questions | 4/5 |
| Multi-page / listing | 3/3 |
That is why I like this row. Hy3 can route ordinary browser work. It navigates cleanly, starts forms with the accessibility tree, and does not fall out of the structured tools interface.
Where it loses points
The weak spots are mostly boundary decisions:
| Category | Sonnet matches | Pattern |
|---|---|---|
| Ambiguous / clarify | 1/8 | Split across page inspection, no-tool answers, clarify, and one execute_js call. |
| Destructive / refusal-worthy | 3/6 | Often inspected the page instead of asking for explicit confirmation first. |
| Browser internals | 2/5 | Mixed navigate, execute_js, screenshot, and page inspection choices. |
| Downloads | 4/6 | Picked plausible download tools, but differed from Sonnet on YouTube thumbnail and README cases. |
| UI mutations | 2/4 | Mixed navigation, page inspection, and execute_js on browser-control tasks. |
This explains the gap between Hy3's strong exact/ideal scores and its lower all-case Sonnet score. It is good when a browser tool should be used. It is less Sonnet-like when the correct first move is to pause, refuse, answer directly, or handle a browser-internal edge.
The multimodal gap
For WebBrain, the obvious missing piece is vision. OpenRouter's Hy3 model page and public model metadata currently expose Hy3 as text->text: text input, text output, no image input. That makes this benchmark a planner result, not a full browser-agent result.
That distinction matters because WebBrain often needs screenshots: visual confirmation, OCR-ish page states, canvas-heavy apps, broken accessibility trees, and UI affordances that text extraction misses. A text-only planner can choose tools well, but it cannot replace a multimodal browser model.
I still expect this gap to close. Tencent is clearly positioning Hy3 for agentic workflows, long-horizon tasks, tool-calling, coding, document processing, financial analysis, game development, and frontend design. In the broader Chinese frontier-model lane, text-first releases have been moving toward visual capabilities quickly; DeepSeek's V4 release is text/agent/1M-context focused in the public docs, while the surrounding DeepSeek ecosystem has been pushing vision-token and OCR-style work. Hy3 feels like the same kind of model family: strong text first, multimodal pressure next.
So the fair wording is: Hy3 is not multimodal yet, but I would be surprised if Tencent leaves it text-only for long. If image input arrives in the coming months and the tool-calling behavior holds, this becomes much more interesting for WebBrain than the current row already is.
Bottom line
Tencent Hy3 is a great hosted text planner in this frozen WebBrain benchmark. It is not the new overall winner, and MiniMax M2.7 still has the stronger all-case Sonnet row. But Hy3 is cleaner than MiniMax M3 on parsed tool calls, exact matches, ideal-name matches, and Sonnet-tooled alignment, while matching the practical latency band and costing nothing in this temporary free run.
The caveat is simple: it is text-only today. For WebBrain, that keeps it in the planner bucket rather than the full agent bucket. Add vision, keep this tool discipline, and Hy3 becomes a serious default-candidate conversation instead of just a very good OpenRouter benchmark row.
Tags: #TencentHy3 #OpenRouter #MiniMaxM3 #MiniMaxM27 #ToolCalling #BrowserAgent #WebBrain