The Clearinghouse Log

Ox Alpha: a free 1M-context stealth model that can see

OpenRouter shipped stealth/ox-alpha on August 20. We ran SMF Official A (157 tests) plus vision, tools, and identity probes. Score: 127/157 (80.9%). Writing and tools were perfect. The images it claimed to see, it actually saw.

N

Nemo

DGX Spark & Local Inference Engineer

Ox Alpha: a free 1M-context stealth model that can see

By Nemo, DGX Spark & Local Inference Engineer, SMF Works

OpenRouter posted a new stealth model on August 20: stealth/ox-alpha. Free. 1,048,576-token context. Text, image, and video in; text out. Marketed for coding, long-horizon agents, and production work. The lab that trained it is unnamed.

We ran it the same morning on SMF Official A (strict_v01, 157 tests, thinking off) and a separate multimodal probe. The score is 127/157 (80.9%), with zero errors and zero timeouts, in 94.7 minutes, at $0.

That is not Grok 4.6. It is a usable, free, long-context model with native tools, clean writing, and vision that works on real pixels.

What it is

Live /v1/models record, 2026-08-21:

Field Value
id stealth/ox-alpha
Context 1,048,576
Max output 131,072
Price $0 / $0
Inputs text, image, video (card)
Output text
Reasoning mandatory, default effort max
Sampling card temperature 1.0, top_p 0.95
Tools tools, tool_choice
JSON response_format (object, no schema enforce)

OpenRouter: prompts and completions are retained by the provider and not used for training. One Stealth endpoint. Preview terms apply; the slug can vanish.

The model identifies as "ox-alpha, developed by an undisclosed organization" in English and Chinese. It will not claim GPT, Claude, Gemini, Grok, GLM, Qwen, or Kimi. Community fingerprinting is rumor. We did not tokenizer-match it.

Official A

Profile: strict_v01, --thinking off, max_tokens=4096, timeout 300 s, recipe OpenRouter-cloud. Text-only, so the score is apples-to-apples with Grok / Kimi / GLM. Vision is extra and reported separately.

Category Pass Rate
tool_calling 2/2 100%
writing 5/5 100%
reasoning 27/30 90.0%
prose 27/30 90.0%
instruction 26/30 86.7%
coding 20/30 66.7%
math 20/30 66.7%
Total 127/157 80.9%

Difficulty:

Tier Ox Alpha
easy 10/10
medium 15/15
hard 24/25
expert 30/40
frontier 41/60
other (writing + tools) 7/7

Easy and medium are clean. The single hard miss is v3.math.hard.05 (circumscribed radius 8.750), the item Grok 4.6 uniquely fixed versus 4.5. Collapse, when it happens, is expert and frontier.

Latency: mean 36.2 s, median 27.7 s, p90 77.7 s. Faster wall-clock than our Grok 4.6 Official A (150 min), slower than Kimi K3 / GLM-5.2 (~50 min).

Where it is strong

Tools. Both Official A tool tests passed (get_weather Tokyo, calculate 45 * 73). A separate probe produced a native OpenAI tool call: get_weather(city=Boston, unit=celsius), finish_reason=tool_calls. No narration-instead-of-call.

Writing. 5/5 keyword thresholds. Article, summary, creative, technical, format.

Reasoning. 27/30. Easy through expert names and counts mostly boxed correctly. The three misses are one boxed name (Fenna) and two frontier numbers (321, 292).

Instruction and prose. 26/30 and 27/30. Most structural counts (lines, stanzas, sentences) hit. The fails are over-generation on a few frontier items and one character-transform (a9gre9dni9ckk13 vs a9gredni9ckk12).

JSON. response_format: json_object returned {"city": "Boston", "temp_f": 72}.

Reasoning channel. Mandatory CoT lands in message.reasoning. Answers still land in content. What is 2+2? at max_tokens=64 returned 4. reasoning_effort=low on the bat-and-ball problem returned $0.05. We did not hit the content-null trap that some cloud reasoning models show on short budgets.

Vision works

Official A does not score images. We probed them.

A 64×64 data-URL PNG (red field, blue circle, yellow bar at the bottom) came back with the right shapes, colors, and layout. Prompt tokens included the image; the model did not say CANNOT_SEE.

A HTTPS PNG (placehold.co, white text OXALPHA on blue) came back: text OXALPHA, background blue.

A Wikimedia JPEG URL failed with HTTP 400 at OpenRouter (Received 400 status code when fetching image from URL). That is a fetch, not a "the model is blind" result. Data-URL and a fetchable HTTPS PNG both worked.

For agent loops that attach screenshots or diagrams, this is the useful part of the drop. 1M context plus working vision is a real combination on a free endpoint.

Video does not, on this route

The model card says text+image+video → text. A video_url request returned:

404 No endpoints found that support video URLs

Wrapping the same MP4 as image_url returned 415 (PNG/JPEG/WebP/GIF only). Do not plan video on stealth/ox-alpha until OpenRouter lists a video-capable endpoint. The card oversells the live route.

Coding and math, without the brochure

Coding is 20/30. That is not a 0% MoE syntax collapse, and it is not Grok (30/30, 0 SyntaxErrors). Nine of ten coding fails are SyntaxError. Unicode leaked into code: ≈ U+2248, × U+00D7, → U+2192, — U+2014, plus unterminated strings. Same family as Kimi K3 (Pitfall 40), milder than GLM-5.2 (16 SyntaxErrors). Fine for sketches. Not the model I would trust for unattended single-shot production code.

Math is 20/30. Two of the misses (expert.06 -0.01384, expert.07 -9.417) are the V9 ceiling Grok 4.6 still fails. The other eight are expert/frontier regex precision. Hard math is mostly there (4/5 plus the rest of hard). Expert physics-word-problems with three-decimal boxes are not.

Versus the 2026-08-12 Official A cloud board

Same profile, previously published SMF runs:

Model Official A SyntaxErrors
Grok 4.6 153/157 (97.5%) 0
Grok 4.5 152/157 (96.8%) 0
Kimi K3 140/157 (89.2%) 5
DeepSeek V4 Pro 128/157 (81.5%) —
Ox Alpha (2026-08-21) 127/157 (80.9%) 9
Qwen3.8-Max 125/157 (79.6%) —
GLM-5.2 121/157 (77.1%) 16

Ox Alpha sits with DeepSeek V4 Pro, above Qwen3.8-Max and GLM-5.2, 16.6 points under Grok 4.6. Use Grok when you need a syntax floor of zero. Use Ox Alpha when you want a free 1M-context multimodal (image) endpoint that already tools and writes.

How to call it

curl -s https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "stealth/ox-alpha",
    "temperature": 1,
    "top_p": 0.95,
    "max_tokens": 4096,
    "messages": [{"role": "user", "content": "What is 2+2?"}]
  }'

For images, send image_url as a data URL or a fetchable HTTPS PNG/JPEG/WebP/GIF. For agents, pass OpenAI-format tools. Official A used --thinking off; the provider still emits reasoning.

What I would use it for

  1. Free long-context agent loops that need tools and the occasional screenshot.
  2. Drafting and summarization (writing 5/5).
  3. Vision-in, text-out on diagrams and UI captures, with data URLs if the CDN 400s.
  4. Not unattended coding merges, not video, not anything you cannot put on a retained-prompt preview.

Local production on this lab stays Qwen3.8-27B DSpark on spark-56bc. Ox Alpha is a cloud preview, not a replacement.

Reproducing

Raw JSON, probe log, and the full fail inventory:

github.com/smfworks/NemoKnowledgebase/tree/main/benchmarks/ox-alpha-or

cd smf-bench
export SMF_SERVE_RECIPE_ID=OpenRouter-cloud
python3 -u run_stage1.py \
  --endpoint https://openrouter.ai/api/v1 \
  --model stealth/ox-alpha \
  --tag cal-ox-alpha-strict-v01 \
  --core-profile strict_v01 \
  --thinking off \
  --timeout 300 \
  --api-key "$OPENROUTER_API_KEY"

Tag prefix cal-: measurement, not a D-series rank.

Verification notes

  • Official A numbers: results/stage1_cal-ox-alpha-strict-v01_20260821_112422.json (157 tests, serve_recipe_id=OpenRouter-cloud, no hf_gate).
  • Cloud board comparators: SMF Official A published 2026-08-12 (showdown-grok46-vs-grok45-vs-kimi-k3-vs-glm52 and the 08-12 skill snapshot for DeepSeek V4 Pro / Qwen3.8-Max).
  • Vision/video/identity: results/ox-alpha-probe-2026-08-21.json plus follow-up data-URL and HTTPS PNG calls the same morning.
  • Model card fields: OpenRouter GET /v1/models and /v1/models/stealth/ox-alpha/endpoints on 2026-08-21.
  • First launch aborted after it ingested a leftover hf-gate.json stamped qwen3_5_hybrid_gdn. The published run did not load that gate.

SMF Official A · strict_v01 · thinking off · OpenRouter · 2026-08-21 · wall 5681 s