Qwen3.8 27B
Qwen · released Aug 5, 2026 · Qwen/Qwen3.8-27B
- Type
- Open weightsApache License 2.0
- Params
- 27.8B
- Context
- 262K
about 197K words of context
Our take
Written Sep 30, 2026Qwen3.8 27B is a downloadable model you can run yourself, and its measured strength is maths and reasoning. Agentic recovery and tool use sit near the bottom of the field, so it suits a solver rather than an agent that has to get itself back on track.
Reach for it on maths and reasoning work, where its LiveBench maths and reasoning averages are the deciding evidence, and for structured-data tasks such as tables and event ordering. You can download it and run it yourself, and the licence allows commercial use, changes and redistribution (Apache License 2.0). Skip it if you need an agent that recovers from failed commands or picks the right tool without inventing one, or if you want a standout for general conversation or creative writing.
The case for it
- 86.21% on LiveBench Mathematics, which covers competition and olympiad problems, so problem-solving work has evidence behind it.
- 80.03% on LiveBench Reasoning, which covers the reasoning tasks of the monthly-refreshed set, so it is a candidate for multi-step questions rather than a proven one.
- 76.59% on LiveBench Data Analysis, which covers table and event-ordering tasks, so structured-data work is a reasonable place to start.
- The licence allows commercial use, changes and redistribution (Apache License 2.0), so the terms are readable before you build on it.
The case against it
- 44th of 55 on Arena Agent · Recovery as of 25 Sep 2026 and 42nd of 55 on Arena Agent · Tool use as of 25 Sep 2026, so an agent that has to get back on track after a failed command is a poor fit.
- 93rd of 168 on Arena Creative Writing as of 25 Sep 2026, a board that records which answer people preferred rather than a prose-craft rubric.
- 63rd of 168 on Arena Text (overall) as of 25 Sep 2026, so it is not a standout for general conversation.
How good is it?
An open text model for chat and writing, though it struggles with calling tools and recovering from failed steps.
- calling tools to carry out requestsArena Agent · Tool use · 42nd of 55
- getting back on track after a step failsArena Agent · Recovery · 44th of 55
EverydayGeneral questions and everyday reasoning
Arena Text (overall)63rd of 168 · 1438
CodingWriting and fixing code on its own
Arena Coding45th of 168 · 1504
AgenticPlanning, calling tools, staying on task
Arena Agent33rd of 55 · −0.014
WritingDrafting and rewriting prose
Arena Creative Writing93rd of 168 · 1364
Placings on Arena's agent boards, from live sessions people ran themselves. A model can lead on one of these and sit mid-field on the others.
Boards this model appears on that none of the ratings above are built on.
Every published score for this model20 scoresEvery figure we hold, from 20 boards, with who ran it and a link to the source — including the boards no rating above is built on.
Can you run it yourself?
Comfortable fit
GeForce RTX 4090 · 24 GB
Room to spare. 3 GB spare means a 10% error in the size would not change the answer.
GeForce RTX 5090 · 32 GB
Room to spare. 11 GB spare means a 10% error in the size would not change the answer.
Apple M1 Pro (16-core GPU) · 32 GB
Room to spare. 4.2 GB spare means a 10% error in the size would not change the answer.
Memory use by level
Against a 24 GB card.
What is quantisation? →This model on every device we track71 devicesThe Q4 build most people download, on each device: what the weights come to, how much context the memory leaves, and whether it runs. Smallest device that runs it first.
Check against your own machine → · Where to rent it hosted →
Or rent it from someone else
Prices checked between 55 min and 13 hours ago — each listing carries its own date.
Some hosts sell this model at two prices: on their own price list (“direct”) and on their OpenRouter listing (“through OpenRouter”). Where the two differ, the row shows both, each with the date we last read it.
Phala, through OpenRouter
Cheapest of the 4 listings we can compare like for like — at 1M of context, out of 18 in the table below. 2 cheaper rows there are outside that comparison: a different context length.
- per 1M tokens
- $0.15 in / $1.88 out
- Context served
- 1M
- Throughput
- ~86 tok/s
| Provider | In / out per 1M tokens | Context | Throughput | Trains on prompts | Logs prompts | Zero retention |
|---|---|---|---|---|---|---|
| Cerebrasfp16Through OpenRouter | $0.99 / $1.49checked 55 min ago | 66K33K max reply | 272 tok/s | No | No | Confirmed |
| AkashMLfp8Through OpenRouter | $0.20 / $1.78checked 55 min ago | 262K131K max reply | 40 tok/s | No | No | Confirmed |
| DeepInfrabf16Direct and through OpenRouter | $0.20 / $2.50directchecked 56 min ago$0.15 / $1.88through OpenRouterchecked 55 min ago | 262K236K max reply through OpenRouter | 26 tok/sthrough OpenRouter | DirectUnknownThrough OpenRouterNo | DirectUnknownThrough OpenRouterNo | DirectUnknownThrough OpenRouterConfirmed |
| PhalaThrough OpenRouter | $0.15 / $1.88checked 55 min ago | 1M262K max reply | 86 tok/s | No | No | Confirmed |
| Parasailfp8Through OpenRouter | $0.24 / $2.20checked 55 min ago | 262K236K max reply | 64 tok/s | No | No | Confirmed |
| Darkbloomfp4Through OpenRouter | $0.050 / $2.20checked 55 min ago | 262K33K max reply | 2 tok/s | No | Yesunknown period | Unknown |
| Chutesfp8Through OpenRouter | $0.24 / $2.20checked 55 min ago | 262K66K max reply | 30 tok/s | No | Yesunknown period | Unknown |
| Mancer 2fp8Through OpenRouter | $0.20 / $2.50checked 55 min ago | 262K236K max reply | 3 tok/s | No | No | Confirmed |
| Ionstreamfp8Through OpenRouter | $0.089 / $2.50checked 55 min ago | 262K66K max reply | 11 tok/s | No | No | Confirmed |
| Alibaba CloudThrough OpenRouter | $0.42 / $2.55checked 55 min ago | 1M131K max reply | 53 tok/s | No | Yesunknown period | Unknown |
| DekaLLMThrough OpenRouter | $0.049 / $3.00checked 55 min ago | 262K236K max reply | 62 tok/s | No | No | Confirmed |
| Novita AIDirect and through OpenRouter | $0.42 / $3.00checked 56 min ago directchecked 55 min ago through OpenRouter | 1M131K max reply through OpenRouter | 32 tok/sthrough OpenRouter | DirectUnknownThrough OpenRouterNo | DirectUnknownThrough OpenRouterNo | DirectUnknownThrough OpenRouterConfirmed |
| CoreWeavefp8Through OpenRouter | $0.40 / $3.00checked 55 min ago | 262K236K max reply | 51 tok/s | No | No | Confirmed |
| OpenRouterOpenRouter's own listing | $0.42 / $3.00checked 57 min ago | 1M | not measured | Unknown | Unknown | Unknown |
| Cloudflare Workers AIThrough OpenRouter | $0.45 / $3.20checked 13 hours ago | 262K236K max reply | 25 tok/s | No | Yesunknown period | Unknown |
| Venice AIfp8Through OpenRouter | $0.45 / $3.20checked 55 min ago | 262K66K max reply | 28 tok/s | No | No | Confirmed |
| RekaThrough OpenRouter | $0.025 / $4.35checked 55 min ago | 262K236K max reply | 6 tok/s | No | No | Confirmed |
| WaferThrough OpenRouter | $0.025 / $4.35checked 55 min ago | 262K236K max reply | 54 tok/s | No | No | Confirmed |
Cerebras
DeepInfra
Novita AI
OpenRouter
Cloudflare Workers AIAcross the 18 listings we hold: 17 say they do not train on prompts (2 of them only through OpenRouter), 0 say they do and 1 does not say. 13 appear in the zero-retention registry we check (2 of them only through OpenRouter); the rest are unknown to us.
What each host's API supports
From the parameter list each endpoint publishes. Streaming is omitted: nothing we hold reports it, for any model.
| Provider | Tool calling | JSON output | Strict schema |
|---|---|---|---|
| Cerebrasfp16Through OpenRouter | ✗ | ✗ | ✗ |
| AkashMLfp8Through OpenRouter | ✓ | ✓ | ✓ |
| DeepInfrabf16Direct and through OpenRouter | ✓ | ✓ | ✓ |
| PhalaThrough OpenRouter | ✓ | ✓ | ✓ |
| Parasailfp8Through OpenRouter | ✓ | ✓ | ✓ |
| Darkbloomfp4Through OpenRouter | ✓ | ✓ | ✓ |
| Chutesfp8Through OpenRouter | ✓ | ✓ | ✓ |
| Mancer 2fp8Through OpenRouter | ✓ | ✓ | ✓ |
| Ionstreamfp8Through OpenRouter | ✓ | ✓ | ✓ |
| Alibaba CloudThrough OpenRouter | ✓ | ✓ | ✗ |
| DekaLLMThrough OpenRouter | ✓ | ✓ | ✓ |
| Novita AIDirect and through OpenRouter | ✓ | ✓ | ✗ |
| CoreWeavefp8Through OpenRouter | ✓ | ✓ | ✓ |
| OpenRouterOpenRouter's own listing | ✓ | ✓ | ✓ |
| Cloudflare Workers AIThrough OpenRouter | ✓ | ✓ | ✓ |
| Venice AIfp8Through OpenRouter | ✓ | ✓ | ✓ |
| RekaThrough OpenRouter | ✓ | ✓ | ✓ |
| WaferThrough OpenRouter | ✓ | ✓ | ✓ |
Tool calling: 17 of 18 listings say yes, 1 says no. JSON output: 17 of 18 listings say yes, 1 says no. Strict schema: 15 of 18 listings say yes, 3 say no.
Models people weigh against Qwen3.8 27B
When we formed this view
Recent changes
What moved
input −55% ($0.198 → $0.089 per 1M tokens), output −2% ($2.55 → $2.50 per 1M tokens)What moved
Qwen3.8 27B moved on 5 hosts: DekaLLM: input −39% ($0.080 → $0.049 per 1M tokens), output +20% ($2.50 → $3.00 per 1M tokens), cache read −50% ($0.040 → $0.020 per 1M tokens); Ionstream: input +37% ($0.145 → $0.198 per 1M tokens); AkashML: input −10% ($0.225 → $0.203 per 1M tokens), output −10% ($1.98 → $1.78 per 1M tokens); Wafer: output −1% ($4.40 → $4.35 per 1M tokens); Reka: output −1% ($4.40 → $4.35 per 1M tokens), cache read −22% ($0.0199 → $0.0155 per 1M tokens)What moved
Qwen3.8 27B moved on 4 hosts: Wafer: input −65% ($0.072 → $0.025 per 1M tokens), cache read −65% ($0.057 → $0.020 per 1M tokens); Reka: input −65% ($0.072 → $0.025 per 1M tokens), cache read −65% ($0.057 → $0.020 per 1M tokens); Ionstream: input +63% ($0.089 → $0.145 per 1M tokens); Phala: input −17% ($0.24 → $0.20 per 1M tokens), output −0.0% ($2.5 → $2.5 per 1M tokens), cache read −0.0% ($0.050 → $0.050 per 1M tokens)What moved
Qwen3.8 27B moved on 5 hosts: Wafer: input −13% ($0.083 → $0.072 per 1M tokens), output +80% ($2.45 → $4.40 per 1M tokens), cache read −28% ($0.0802 → $0.0574 per 1M tokens); DekaLLM: cache read −54% ($0.087 → $0.040 per 1M tokens); Darkbloom: input −28% ($0.069 → $0.050 per 1M tokens); Reka: input −20% ($0.090 → $0.072 per 1M tokens), cache read −32% ($0.0850 → $0.0574 per 1M tokens); Ionstream: input −15% ($0.105 → $0.089 per 1M tokens), cache read −15% ($0.100 → $0.085 per 1M tokens)What moved
Qwen3.8 27B moved on 4 hosts: Wafer: output −44% ($4.40 → $2.45 per 1M tokens); DekaLLM: input −16% ($0.0960 → $0.0802 per 1M tokens), output −43% ($4.40 → $2.50 per 1M tokens); Reka: input −1% ($0.081 → $0.080 per 1M tokens), output −43% ($4.40 → $2.50 per 1M tokens), cache read −1% ($0.081 → $0.080 per 1M tokens); Darkbloom: input −16% ($0.0825 → $0.0690 per 1M tokens); Wafer: input +3% ($0.0803 → $0.0828 per 1M tokens); Reka: input −1% ($0.080 → $0.079 per 1M tokens), cache read −1% ($0.080 → $0.079 per 1M tokens)What moved
Qwen3.8 27B moved on 2 hosts: Reka: input −2% ($0.092 → $0.090 per 1M tokens); Wafer: input −2% ($0.093 → $0.091 per 1M tokens)What moved
Qwen3.8 27B moved on 3 hosts: Ionstream: input −58% ($0.250 → $0.105 per 1M tokens); Reka: input −2% ($0.094 → $0.092 per 1M tokens); Wafer: input −2% ($0.095 → $0.093 per 1M tokens)Each date is the day we first saw the change, or the day the maker announced it.
What we do not know about this model yet
- We hold no measured file for it, so all 3 sizes on this page are calculated from the parameter count.
- Nothing we hold says whether an endpoint streams, so we do not show it either way.
- 1 of 18 listings does not say whether it trains on prompts, and 2 answer only through OpenRouter, not for their own listing.
- We hold no batch or off-peak rate for any of its listings.
- We hold a decode speed for it, but no prompt-processing (prefill) figure, so how long the input side of a job takes is unknown to us.
Licence and identifiers
What the licence allowsApache License 2.0, what it allows commercially, and the identifiers you need to pull this model — its Hugging Face repo, our slug and a machine-readable card.
Licence
Apache License 2.0
Fully permissive: commercial use, redistribution, and derivatives allowed. Requires attribution and a copy of the license. Includes an express patent grant.
Identifiers
- Hugging Face
- Qwen/Qwen3.8-27B
- Architecture
- Dense
- Takes in, gives back
- Text, images and video in, text out
- Catalogue slug
- qwen-qwen3-8-27b