Tencent Hy4: Vision + Computer-Use API Today, $0.417/$1.251 Promo, 1M vs 65K Max-Out, Weights Pending

Attributed compile + one small offline run on my box

Primary source: Tencent Launches Hy4, Joining the Ranks of the World's Leading Open-Source AI Models by Tencent, published 2026-11-04.

Secondary: TechNode — Tencent launches Hy4 with vision and computer-use capabilities (2026-11-04).

Callable surface: OpenRouter — tencent/hy4 (models API JSON + page FAQ / provider table). Local copies: /workspace/reputation/runs/33-hy4/.

Blocked in preflight: hy.tencent.ai/research/hy4 returned EdgeOne HTTP 567 Restricted Access — cite OpenRouter + press for pricing; do not treat the research page as opened.

Numbers below from press / TechNode / OpenRouter are vendor or third-party dashboard unless marked otherwise. This box did not call chat completions, download weights, or use an API key. Offline preflight = public HTTP scrapes + OpenRouter /api/v1/models JSON only.

---

Tencent shipped Hy4 on 2026-11-04: first vision + first computer-use for the Hy line, with the API live on Tencent Cloud TokenHub and third-party routes including OpenRouter. The useful builder story is not “another open-source flagship pitch.” It is four concrete checks: (1) today is an API / product day — weights and the system card are promised in the coming days / shortly, not a download you can pin yet, (2) OpenRouter model id tencent/hy4 is text+image→text with a 1,048,576 context card and a separate 65,536 max-completion cliff, (3) TokenHub / OR route pricing is on a 50% promo through 2026-12-31 where cache-read rates dominate agent-loop cost, and (4) vendor financial / hallucination / GameDev deltas are a map for your golden eval — not a merge-into-prod certificate.

TechNode reports ~770B parameters (vendor via TechNode — label as such). Press positions Hy4 among leading open-source models and acknowledges responsible-release framing around cyber and bio. Coverage around the launch also notes Hy4 still trails leading closed models on hard internal math / coding / self-improve / VQA benches (vendor / press framing); we invent no closed-model scores here.

What it is (short)

Hy4 adds fine-grained image understanding, complex chart analysis, and mathematical visual reasoning, plus computer-use across CLI and GUI. Press calls out specialized training on professional software such as Blender and CAD, and stronger long-horizon agentic behavior. Product surfaces named on launch day: Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom AI Assistant (Dayuan), ima; third-party API routes named: OpenRouter, Hermes, OpenCode. Promo windows: WorkBuddy + CodeBuddy free Nov 4–15; TokenHub Hy4 50% off through Dec 31, 2026. Hy4 Flash is scheduled soon (press body / TechNode — treat as upcoming, not a second GA you can call today from our preflight).

Sibling on OpenRouter: tencent/hy4-preview is still text→text only (no image), same 1M context card, and list-style pricing without the GA 50% we saw on the tencent/hy4 Tencent Cloud route. Do not reuse preview client assumptions for vision.

Access matrix

Route	What we saw	Note
OpenRouter tencent/hy4	models API + page HTTP 200	text+image→text; tools + structured outputs + reasoning params per OR FAQ
OpenRouter tencent/hy4-preview	models API HTTP 200	text→text only; same 1M context; no GA 50% on the route we saw
Tencent Cloud TokenHub / WorkBuddy / CodeBuddy / Yuanbao / WeCom Dayuan / ima	Named in press	Product day — we did not authenticate
Hermes / OpenCode	Named in press	Third-party labels — not opened in this preflight
Weights + system card	Promised “coming days” / “shortly”	Not released in our 2026-11-04 preflight
hy.tencent.ai/research/hy4	HTTP 567 EdgeOne	Restricted — not a source of record for this draft

Ecosystem notes around the launch mention partnership with vLLM and SGLang communities for when weights land; our box could not re-verify that wording on the research page (567). Until HF / vLLM / SGLang recipes appear, stay on API.

Price math (OpenRouter + press)

Prefer OpenRouter API JSON + OR page for callable numbers. Press approximate USD (~$0.90 / $2.70 / $0.045 cached) is close but not identical to OR list — slight press vs OR difference; pin the route you actually bill.

OpenRouter tencent/hy4 (Tencent Cloud route, 50% off in our snapshot):

	List (struck on OR)	Effective at 50% off
Input / M	$0.834	$0.417 (0.000000417 / token)
Output / M	$2.501	~$1.251 (0.0000012505 / token)
Cache read / M	$0.042	$0.021 (0.000000021 / token)

OR provider table snapshot (not our measured bench): P50 latency ~3.30 s, throughput ~53 tok/s.

Why cache hit rate dominates agent loops. Illustrative weighted input at ~95% cache hit:

• List: 0.05×0.834 + 0.95×0.042 ≈ $0.082 / M input
• Promo: 0.05×0.417 + 0.95×0.021 ≈ $0.041 / M input

Output still bills at full completion rates (~$1.251 / M on promo, $2.501 / M list). An OR dashboard snapshot in our preflight showed an effective weighted average around ~$0.079 / M input vs ~$2.38 / M output under a ~95% cache-hit regime — treat as OR dashboard snapshot, not an independent bench. The mechanism point is the same: for long agent loops that re-send a stable prefix, cache-read pricing — not the brochure input rate — sets the floor.

Promo math ends 2026-12-31 on the TokenHub / OR 50% path we saw; after that, re-read the provider table before you budget Q1.

Context cliff: 1M card ≠ 65K max-out

OR FAQ: context window 1,048,576; supports up to 65,536 completion tokens. Architecture from models API: modality text+image→text; input modalities text + image; output text. Tools + tool_choice, structured outputs via response_format, and reasoning / related params are listed as supported on OR — still verify on your client against the live schema.

Do not confuse a 1M prompt window with a 65K max completion. Long computer-use traces that emit huge action JSON will hit max_out long before they fill the context card. Pin max_tokens / max_completion_tokens deliberately and watch finish_reason=length.

Call path (illustrative — not executed)

OpenAI-compat against OpenRouter with model tencent/hy4. Minimal shape for a text+image turn (we did not call chat completions):

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "tencent/hy4",
    "max_tokens": 2048,
    "messages": [{
      "role": "user",
      "content": [
        {"type": "text", "text": "Extract the trend and units from this chart."},
        {"type": "image_url", "image_url": {"url": "https://example.com/chart.png"}}
      ]
    }]
  }'

# Illustrative only — not run in this preflight
from openai import OpenAI
client = OpenAI(base_url="https://openrouter.ai/api/v1", api_key="...")
resp = client.chat.completions.create(
    model="tencent/hy4",
    max_tokens=2048,
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "Extract the trend and units from this chart."},
            {"type": "image_url", "image_url": {"url": "https://example.com/chart.png"}},
        ],
    }],
)

If your client still points at tencent/hy4-preview, you will not get the GA vision path. Swap the model string and re-test image payloads before you blame “Hy4 can’t see charts.”

Vendor benches (label as vendor)

From the 2026-11-04 Tencent press (and TechNode where noted). All scores below are vendor:

• Blind financial eval (Tencent Hy + Tencent Financial Technology): 213 professionals, 579 tasks — Hy4 3.04, DeepSeek-V4.1-Flash-1017 2.97, Kimi K3.1 Pro 2.68, GLM-5.3 2.68.
• iRaB +12.03 pp vs Hy4 Preview (vendor internal / Rabyte).
• Hallucination rate Hy4 Preview 12.95% → Hy4 7.2%; unsupported-assumption errors 23.1% → 5.6% (vendor internal).
• WorkBuddy: avg 83.18, pass rate 96.15%, first in 9/12 mainland CN scenarios (vendor).
• GameDev +19.1%, GameLogic +9.8% vs Preview (vendor).
• Cyber (high-level only): trained on defensive security; offensive exploit materials excluded; Xuanwu / Keen assessment found no significant offensive concerns in areas evaluated — no exploit detail here.
• Bio (vendor): strong foundational bio knowledge; open-ended experimental design more limited vs frontier.

Useful as a map of where they want you to look. Useless as a substitute for your charts, multi-app computer-use tickets, and finance workflows.

Checklist before you trust it

1. Treat Hy4 today as a multimodal API + computer-use product story; open weights are a promise, not a shard path.
2. Call tencent/hy4 (not preview) for vision; confirm text+image→text on the live models card.
3. Budget from promo $0.417 / $1.251 and especially $0.021 cache-read through Dec 31, 2026; design agent loops for cache hits.
4. Separate 1,048,576 context from 65,536 max completion — pin max_out.
5. Freeze a golden eval: charts, multi-app computer-use, finance tickets you actually run — before trusting vendor financial / hallucination / GameDev numbers.
6. Put weights / system card / vLLM / SGLang on a watchlist; re-evaluate self-host only when recipes exist.
7. Do not invent closed-model scores from “still trails frontier” press language.

What I actually ran (offline)

/workspace/reputation/runs/33-hy4/ — 2026-11-04 09:05 America/Toronto (UTC-5) (box stamp Wed Nov 4 14:05:30 UTC 2026):

NO API KEY / NO CHAT COMPLETION / NO WEIGHT DOWNLOAD

HTTP 200: tencent.com/.../tencent-launches-hy4-... (press)
HTTP 200: technode.com/2026/11/04/tencent-launches-hy4-with-vision-and-computer-use-capabilities/
HTTP 200: openrouter.ai/tencent/hy4 (+ /api/v1/models JSON for tencent/hy4)
HTTP 567: hy.tencent.ai/research/hy4 (EdgeOne Restricted Access)

OR models API: tencent/hy4 ctx 1048576; pricing prompt 0.000000417 /
  completion 0.0000012505 / input_cache_read 0.000000021;
  arch text+image->text
OR FAQ: max completion 65536; tools + structured outputs; reasoning supported
OR provider snapshot: ~3.30s P50 latency, ~53 tok/s (dashboard — not our bench)

NOT tested: chat completion; TokenHub auth; WorkBuddy/CodeBuddy;
Hermes/OpenCode; weight download; vLLM/SGLang serve; golden eval re-score

When Hy4 might become a default API string

Only after a callable path and a golden eval win on your problems:

1. You need vision + computer-use on a Hy-line model and can live on API until weights land.
2. You have pinned promo vs list pricing and instrumented cache hit rate for agent loops.
3. You treat 1M vs 65K max-out as two different budgets.
4. Vendor financial / WorkBuddy / GameDev numbers are maps, not certificates.
5. You can roll back the model string without rewriting the harness.

If you needed open weights yesterday, this launch is the wrong trigger — watch for the system card and community recipes. If you already route through OpenRouter and your failures are chart reading, multi-app GUI/CLI agents, or finance tickets, Hy4 is a same-day candidate for a side-by-side, because the GA modality jump and the promo/cache math are concrete enough to instrument.

Hy4’s builder-relevant release is the vision + computer-use API, the $0.417/$1.251 promo with $0.021 cache-read, the 1M context vs 65K max-out cliff, the preview→GA modality jump, and vendor benches labeled as such — with weights still pending. Treat those as inputs, run your own eval, and only then change the production model string.

---

— YongBo Yu · Toronto AI engineer · yongbo-yu.vercel.app · GitHub @YongBoYu1