Throughput (tokens / sec)
avg. 105tokens/s
Dense multimodal coding with thinking control
Run instantly. Pay only for what you use.
Qwen 3.8 27B is Alibaba's dense open-weight vision-language model under Apache 2.0: all 27B parameters active on every token, a 262K-token native context, and native multimodal input — text, images, and video. Thinking is on by default, with per-request enable_thinking, preserve_thinking, and reasoning_effort (xhigh / medium / low).
The 3.8 generation sits on Qwen3.5 hybrid attention — Gated DeltaNet mixed with gated attention, plus multi-token prediction — so a compact dense model can hold long agent loops without MoE routing variance.
Qwen-reported scores with thinking enabled: 61.7 on SWE-bench Pro, 73.0 on Terminal-Bench 2.1, and 42.2 on DeepSWE 1.1 (vs 13.3 for Qwen 3.6 27B), with strict function calling and JSON output.
27Bdense parameters
Every parameter is active on every token — consistent quality with no routing variance.
262Ktoken native context
Native window, extendable to 1M tokens via YaRN for whole-repo prompts.
84.3on OSWorld-Verified
Qwen-reported autonomous desktop computer use — screenshots in, actions out.
Qwen 3.8 27B is strongest on agentic coding and long-horizon computer-use work — repository patches, visual front-end tickets, and multi-turn agents that keep earlier reasoning.
Qwen reports 61.7 on SWE-bench Pro and 73.0 on Terminal-Bench 2.1 with thinking enabled — plans, edits, and verifies real repositories.
Qwen reports DeepSWE 1.1 jumping 13.3 → 42.2 over the previous generation — multi-step fixes hold to the end of the run.
On by default and switchable per request — reasoning_effort (xhigh, medium, low) and preserve_thinking tune depth and carry traces across turns.
Screenshots, diagrams, and hour-scale video are first-class inputs — front-end bugs and visual tickets stay in one model.
Repository-level coding agents and refactors
Qwen reports 61.7 on SWE-bench Pro and 42.2 on DeepSWE 1.1 with thinking on — dense weights, no expert-routing variance.
Front-end work from screenshots and recordings
Screenshots, UI recordings, and hour-scale video are native inputs, so visual bugs and design-to-code stay in one model.
Multi-turn agent loops
preserve_thinking keeps chain-of-thought across turns; reasoning_effort trades depth for speed and token cost per request.
Long-horizon terminal and computer-use tasks
73.0 on Terminal-Bench 2.1 and 84.3 on OSWorld-Verified (Qwen-reported) for autonomous terminal and desktop work.
Cost-first, high-volume agent fleets
Every token runs all 27B parameters. When per-token cost dominates, use Qwen 3.6 35B-A3B — near-peer quality at 3B-active pricing.
Tight output-token budgets with thinking on
Thinking is on by default and verbose. Disable it or drop reasoning_effort to medium/low for cost-sensitive replies.
Audio input or speech understanding
Inputs are text, image, and video only. Transcribe audio upstream before sending it to the model.
Compare Entrim’s pricing, powered by an optimized inference runtime, against other providers using the same token counts per request.
avg. 105tokens/s
avg. 780ms
61.7SWE-bench Pro
Harder, contamination-resistant software engineering tasks.
73.0Terminal-Bench 2.1
Autonomous work in a real terminal — builds, debugging, ops tasks.
42.2DeepSWE 1.1
Long-horizon agentic coding with environment feedback.
Code examples
curl -X POST "https://api.entrim.ai/v1/chat/completions" \ -H "Content-Type: application/json" \ -H "Authorization: Bearer $ENTRIM_API_KEY" \ -d '{ "model": "Qwen/Qwen3.8-27B", "messages": [ { "role": "user", "content": "Why do they call it a building if it is already built?" } ] }'SDKs & docs
Use familiar SDK patterns with quickstart examples for common production setups.
OpenAI-compatible
Keep the request format your team already knows.
No cold starts
Production requests are served without model spin-up delays.
Tool-heavy coding at 3B-active cost
High-throughput agentic work with 1M context
Multimodal long-context work with thinking control