Hosted open models, specialized
The model that does your job, not the one in the headlines.
Pay for the task, not the brand. Every model here is on the roster to do a specific job, and you pick per request: one endpoint, one balance, the OpenAI-compatible or Anthropic-compatible API you already use, prepaid, so an LLM product or an always-on agent cannot produce an invoice.
- Output speed, measured on the serving build
- up to 179 tok/s
Fastest live model. How each figure was measured, and where it does not hold.
curl -N https://api.tiyuvta.ai/v1/chat/completions \
-H "Authorization: Bearer $KEY" \
-d '{"stream":true,
"model":"zai/glm-5.3-flash",
"messages":[{"role":"user",
"content":"hi"}]}' \
-o /dev/null \
-w 'ttft %{time_starttransfer}s\n'- GLM-5.3-Flashup to 179 tok/s · 1M · $0.15 / $0.50
- DeepSeek-V4.1-Flashup to 101 tok/s · 1M · $0.30 / $1.20
Models available now
Measurements and model details6 specialists on one endpoint, and every one of them is on the frontier: nothing here is beaten on every axis by another model on the roster. Pick the one built for the task; one key and one balance cover them all.
Whose numberBenchmark scores under each model name are the maker’s own published figures. They are not ours and are not comparable between makers. Output speed is measured by us on the live endpoint. Those are observations, not service guarantees. Every figure, with its conditions.
Strong, cheap, fastGLM-5.3-Flashzai/glm-5.3-flashStrong, cheap and the fastest here: the pick for anything that finishes in a few calls, with the full 1M window when the prompt is big.Z.ai: Terminal Bench 2.1 84.3, DeepSWE v1.1 63.4, GDPVal-AA v2 1773- context
- 1M
- output, up to
- 179 tok/s
- in / out per 1M
- $0.15 / $0.50
Smart, cheapest over long iterationDeepSeek-V4.1-Flashdeepseek-ai/deepseek-v4.1-flashSmart, with reasoning on by default, and the cheapest here for long iteration: once a loop replays its history, the cache carries most of the bill at the cached rate.DeepSeek: Terminal-Bench 2.1 90.6, DeepSWE v1.1 74.2, Agent's Last Exam 31.8- context
- 1M
- output, up to
- 101 tok/s
- in / out per 1M
- $0.30 / $1.20
Qwen3-Embedding-8Bqwen/qwen3-embedding-8bPick it for memory: turn your documents into vectors once and retrieve them forever, at the lowest rate on the roster.Embeddings: MTEB multilingual #1 open model- context
- 16K
- class
- retrieval · 429 beyond its advertised 8
- input per 1M
- $0.02
- Qwen3-Reranker-8B
qwen/qwen3-reranker-8bPick it for precision: score your top retrieval hits so the model answers from the right ones, not the nearest ones.Reranker: query + documents in, relevance scores out- context
- 16K
- class
- retrieval · 429 beyond its advertised 8
- input per 1M
- $0.05
Hebrew speech to textWhisper Large v3 Hebrewivrit-ai/whisper-large-v3-heHebrew transcription on the ivrit-ai Whisper Hebrew weights, streaming or whole file, priced by the audio hour rather than by the token.Audio in, text out, Hebrew:/v1/realtime(streaming) and/v1/audio/transcriptions(whole file)- streaming
- $0.18 / audio hour
- whole file
- $0.10 / audio hour
- billed on
- audio duration, not tokens
- English speech to textQwen3 ASR English
qwen/qwen3-asr-enEnglish transcription on a small, fast model, streaming or whole file, priced by the audio hour rather than by the token.Audio in, text out, English:/v1/realtime(streaming) and/v1/audio/transcriptions(whole file)- streaming
- $0.15 / audio hour
- whole file
- $0.04 / audio hour
- billed on
- audio duration, not tokens
- Billing
- No subscription, no minimum. Credit does not expire.
- Switching
- Standard APIs. No product-specific SDK to remove.
Make the first request
This request uses OpenAI Responses. Create a key, copy it, and replace the prompt. No product-specific SDK is required.
curl https://api.tiyuvta.ai/v1/responses \
-H "Authorization: Bearer $TIYUVTA_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"zai/glm-5.3-flash","input":"Hello."}'Your prompts stay yours
Purchased-credit prompts are processed in memory to answer the request and are not used to train, fine-tune, distil or evaluate a model. There is no sampling queue and no human review. Promotional sign-up credit is the exception, and its terms say so.
If you want to, you can turn training on for your own traffic and take 5% of your metered spend back as credit. It is one switch in the console, off unless you set it, and turning it off stops it for everything sent afterwards. The full data policy.
Running this at company scale?
Avi Fenesh researches and tunes the serving engine, and operates this service. Talk to us directly and we set up an arrangement for your workload: volume, limits, billing, terms. Include the model, expected concurrency and token volume, and the founder answers.
Need it tuned to your data, or on hardware you control? What the lab does.