tiyuvta inference
Menu
model · live

Qwen3.8 27B

A compact, deployment-friendly dense model: a native vision-language model that understands images and videos, with flexible thinking control, designed to carry complex, multi-step tasks through to completion with greater reliability.

Qwen’s own description of the model, from Qwen3.8-27B model card . Served here as qwen/qwen3.8-27b.

What Qwen built it for

in the vendor’s words

Quoted from Qwen3.8-27B model card, retrieved 23 Aug 2026. Thinking on by default, with a reasoning-depth dial. The card does not label its benchmark tables thinking or non-thinking, so no row here claims a mode the vendor did not state.

Published benchmarks

Qwen’s figures, not ours
Qwen3.8 27B — scores as published by Qwen, retrieved 23 Aug 2026
Benchmark What it measures Score How Qwen ran it
Terminal Bench 2.1
Terminus harness
 Agentic terminal coding 73.0 The card names the Terminus harness for this row and states no further sampling detail. Mode: not stated for this row; the card states the model operates in thinking mode by default.
SWE-bench Pro Agentic coding, real-world software engineering 61.7 Claude Code harness, temperature=1.0, top_p=0.95, 256K context window; "Problematic tasks were corrected, and all baseline models were re-evaluated on the refined benchmark". Mode: not stated for this row; the card states the model operates in thinking mode by default.
LiveCodeBench v6 Competitive coding 90.3 The card states no separate harness or sampling for this row. Mode: not stated for this row; the card states the model operates in thinking mode by default.
Agents' Last Exam Frontier agentic tasks 20.4 pass@1 / 42.9 score The card reports this row as two figures, pass@1 and score. Mode: not stated for this row; the card states the model operates in thinking mode by default.
GPQA Diamond Graduate-level science reasoning 89.2 The card states no separate harness or sampling for this row. Mode: not stated for this row; the card states the model operates in thinking mode by default.
HLE Humanity's Last Exam, multidisciplinary reasoning 30.8 Judged by GPT-4o. Mode: not stated for this row; the card states the model operates in thinking mode by default.
MathVision Visual maths problem solving 90.0 without code interpreter / 94.6 with Evaluated with the fixed prompt "Please reason step by step, and put your final answer within \boxed{}"; a small number of incorrect ground-truth annotations were corrected before scoring. Mode: not stated for this row; the card states the model operates in thinking mode by default.
OSWorld-Verified Computer use, from screenshots 84.3 The card states no separate harness or sampling for this row. Mode: not stated for this row; the card states the model operates in thinking mode by default.

Every score above is Qwen’s own published result for this model, copied from Qwen3.8-27B model card on 23 Aug 2026 — we did not run these benchmarks and we do not restate them as our own. Each vendor ran its own harness with its own judge, so these tables are not a league table: scores from two different cards are not comparable, even where the benchmark name matches. Each model here shows only its own maker’s table, for that reason.

Recommended usage

what Qwen recommends

Sampling

For thinking mode, Qwen recommends temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0. That is what this endpoint sends when your request omits them, so the default request shape is the vendor’s recommendation rather than ours. An explicit value in your request always wins.

Qwen publishes a second set for instruct (non-thinking) mode: temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0. This endpoint carries one set of sampling defaults per model, not one per thinking mode, so a request that turns thinking off still gets the thinking arm's values unless it sets its own. If you turn thinking off, send these six parameters yourself to match Qwen's recommendation for that mode.

On repetition, Qwen notes: “you can adjust the presence_penalty parameter between 0 and 2 to reduce endless repetition. However, using a higher value may occasionally result in language mixing and a slight decrease in model performance.”

Thinking

Thinking is on by default and its depth is a dial: reasoning_effort takes xhigh (the vendor default, "for complex tasks demanding thorough analysis"), medium ("balancing accuracy and speed") and low ("efficient reasoning optimizing for speed and cost").

Qwen’s own caution: “In multi-turn agentic tasks, lower reasoning effort does not always reduce overall task completion time. Although it may produce faster per-turn responses, it can also lead to insufficient analysis, more failures, and repeated retries, which may increase total latency and token consumption.”

On this endpoint: Reasoning depth is a dial: reasoning_effort accepts low, medium, and xhigh (the vendor default), and "none" turns reasoning off entirely for latency-critical turns.

Context length

Qwen gives the native window as 262,144 tokens, which is what we serve — the full checkpoint window, with input and output sharing it. Beyond the native window the vendor points at YaRN RoPE scaling, to a ceiling of 1,000,000 tokens, and warns against leaving it on for ordinary traffic.

Qwen’s caution on extending it: “All the notable open-source frameworks implement static YaRN, which means the scaling factor remains constant regardless of input length, potentially impacting performance on shorter texts.” We serve the native window and do not enable static RoPE scaling, so ordinary-length requests are not paying that cost.

What to point it at

Qwen names these workloads for this model:

  • “coding, professional work, research, and long-horizon agentic tasks”
  • “from STEM diagrams and documents to hour-scale videos”

Output length

“To optimize performance on agentic tasks, we recommend allocating sufficient output length to allow the model to generate detailed and comprehensive responses.” For agentic work the vendor asks for room to think: a 262,144-token cap on reasoning content and a 131,072-token cap on the final response.

Output here is not capped below the context window, so there is room to follow that advice.

Recommendations on this page are Qwen’s, from Qwen3.8-27B model card, retrieved 23 Aug 2026. We default to them rather than to a house preference; where our default would differ, this page says so.

What it does on this endpoint

our measurements · not vendor figures
output speed up to 180 tok/s single stream, measured live
first token 185 ms median, same-region US, reasoning off
context 262K tokens, the native window

These three are ours, measured on the live endpoint — observations, not service guarantees, and not to be confused with the vendor scores above. The protocol, the vantages, the build each figure is pinned to and how to check it against the system_fingerprint in your own response are all on the measurements page. Run your own and use those instead.

Price

per million tokens · no request fee · no subscription
Token rates — Qwen3.8 27B
Input Cached input Output
$0.40$0.10$2.03

One price, no tiers. Cache hits are applied automatically and reported per response as cached_tokens. Full price sheet and worked examples.

Capabilities

qwen/qwen3.8-27b

Three API formats, one base URL

OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages — all served from https://api.tiyuvta.ai/v1. Point an existing client at it and set the model id; there is no product-specific SDK.

Tools and structured output

tools and tool_choice, JSON mode, and JSON schema via response_format. The tool-call round trip is part of the release gate on every deploy.

Vision and video

Images go in as image_url content parts and bill as ordinary input tokens — vision costs nothing extra. Video goes in as video_url content parts; sampled frames bill as input tokens.

Reasoning: an effort dial

Reasoning depth is a dial: reasoning_effort accepts low, medium, and xhigh (the vendor default), and "none" turns reasoning off entirely for latency-critical turns.

Prompt caching

Continuing a conversation reuses its prefix and bills the cached rate for what actually hit — reported per response, priced per token, no hit-rate promises.

Exactness-gated serving

Speculative decoding is verified byte-identical to plain decode on the serving configuration before a build takes traffic — speed never changes the output. The correctness gates.

First request

qwen/qwen3.8-27b
shell — default https://api.tiyuvta.ai/v1 
curl https://api.tiyuvta.ai/v1/chat/completions \
  -H "Authorization: Bearer $TIYUVTA_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen/qwen3.8-27b",
       "messages": [{"role": "user", "content": "Hello"}]}'

The default request reasons before it answers, at Qwen’s recommended sampling — best quality for hard problems.

shell — low latency reasoning off 
curl https://api.tiyuvta.ai/v1/chat/completions \
  -H "Authorization: Bearer $TIYUVTA_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen/qwen3.8-27b",
       "reasoning_effort": "none",
       "messages": [{"role": "user", "content": "Hello"}]}'

reasoning_effort: "none" makes the first streamed token the answer itself — the measured 185 ms first-token figure above is this request shape.

Create account Agent client configurations Read the docs
Weights and vendor sources
huggingface.co/Qwen/Qwen3.8-27B, Apache-2.0. Qwen trained this model; we serve it at its full native context in NVFP4 4-bit weights. Benchmark scores and usage recommendations on this page come from Qwen3.8-27B model card , retrieved 23 Aug 2026.
Machine-readable metadata
GET https://api.tiyuvta.ai/v1/models for supported parameters, and the provider feed at GET https://api.tiyuvta.ai/models?schema=openrouter.
Rate limits
Per account, for abuse control, raised on request — capacity scales with demand rather than sitting behind a published ceiling.