by Alibaba Cloud · Qwen3.8 family · best for lowest-price 1M-context coding and office agents
Qwen3.8-Flash is Alibaba Cloud's low-cost Qwen model, released August 26, 2026 on Model Studio and Qwen Cloud. It is the production version of the open-weight Qwen3.8-Flash-Next, a 125B-parameter mixture-of-experts model with 6B parameters active per token, which Alibaba describes as an early preview of the Qwen4 architecture. It reads text, images and video, holds 1M tokens of context, thinks by default at an adjustable effort level, and lists at $0.15 input / $0.47 output per 1M tokens, flat across the whole window. Artificial Analysis scores it 40 on its Intelligence Index, level with DeepSeek V4.1-Flash and ahead of GPT-5.6 Luna even at Luna's max effort, at the lowest list price in that group. For a buyer: it is the cheapest per-token route to near-flagship coding and office-agent quality. The costs are slow generation (about 57 tokens/sec), heavy token use, and a weights license that requires a separate deal for anyone running a model-hosting or coding/office-assistant business.
low, medium, xhigh; xhigh is the default) and preserve_thinking, which keeps earlier reasoning in multi-turn agent sessions.| Benchmark | Score | Source |
|---|---|---|
| Humanity's Last Exam | 38% | artificialanalysis.ai 2026-09-21T00:00:00.000Z |
| GPQA Diamond | 92.3% | artificialanalysis.ai 2026-09-21T00:00:00.000Z |
| LiveCodeBench | 91.9% | qwen.ai 2026-08-26T00:00:00.000Z |
| Terminal-Bench | 86.1% | artificialanalysis.ai 2026-09-21T00:00:00.000Z |
| Artificial Analysis Index | 40 | artificialanalysis.ai 2026-09-21T00:00:00.000Z |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“Near-flagship agent quality at the lowest list price in its tier, but it ties you to Alibaba Cloud and its weights license has sharp edges.”
Qwen3.8-Flash gives a strong capability-per-price case: Artificial Analysis places it level with DeepSeek V4.1-Flash and above GPT-5.6 Luna, and Alibaba now makes it the default low-cost recommendation. The strategic questions are dependency and license. The hosted model is only on Alibaba Cloud, without AWS, Azure or Google distribution. The open weights are a real exit route, but a company that sells API access or builds a coding or office assistant needs a separate license, whatever its size. Roadmap confidence is decent: this architecture is the declared basis for Qwen4, so investment will continue, though the Next branding signals it is still experimental.
“Alibaba is selling its Qwen4 preview at Flash prices. The architecture is the moat, and the license fences off its direct competitors.”
The positioning is sharp. Alibaba ships a new architecture as its cheapest serious model, prices it below GPT-5.6 Luna and DeepSeek V4.1-Flash per token, and publishes the weights for credibility. It then writes a license that lets almost anyone use them except the businesses that compete with Alibaba's Model Studio, Qoder and QwenWork. That is a more guarded stance than Apache-era Qwen and than MIT-licensed GLM-5.3-Flash and DeepSeek V4.1-Flash. On capability it sits in a four-way tie at the top of the Flash tier with GLM-5.3-Flash, DeepSeek V4.1-Flash and GPT-5.6 Luna. Its distinct edges are native video input and computer-use tuning.
“$0.15/$0.47 is the lowest rate card in the tier, but its verbose agents cost more per finished task than rivals charging higher list prices.”
The rate card is excellent: $0.15 in, $0.47 out, cache reads at $0.016, flat to 1M tokens, and 1M free tokens for 90 days in Singapore. The Global-scope regions list an even lower $0.113 / $0.382. The catch is consumption. Artificial Analysis spent about $0.37 per index task on it, against $0.25 for GLM-5.3-Flash, $0.27 for DeepSeek V4.1-Flash and $0.18 for GPT-5.6 Luna at max effort, because it produces long reasoning traces and re-reads large contexts. There is no batch discount, and prompts under 32K cost several times more than on Qwen3.7-Flash. Model it per task, not per token.
“Point Claude Code or Codex at it and it just works. The xhigh default is slow, so tune reasoning effort per call before you judge it.”
Integration is easy. The same qwen3.8-flash ID works on OpenAI Chat Completions, the Responses API and an Anthropic-compatible endpoint, with function calling, parallel tool calls, structured outputs, partial mode, explicit caching and built-in web search, code interpreter and PDF parsing. Thinking defaults to xhigh effort, so latency is long until you set reasoning_effort or disable thinking for simple turns. Built-in web search is missing in the Global-scope regions, which matters for EU or US routing. Self-hosting is a different project: a new qwen4_exp architecture, a checkpoint of roughly 185 GB at FP8, and recipes for SGLang, vLLM and KTransformers, but little field experience yet.
“Smart for the money, with good screenshot and video skills, but you wait: roughly 57 tokens a second, after half a minute of thinking.”
Answer quality is where a Flash model rarely goes. GPQA and HLE results sit with GPT-5.6 Luna at max effort, coding help is strong, and it reads screenshots, charts and long videos natively. Its hallucination rate on Artificial Analysis's knowledge test is far lower than Luna's or DeepSeek V4.1-Flash's. The daily-use problem is pace: the default xhigh effort thinks for tens of seconds and then streams at under 60 tokens/sec, so chat feels sluggish next to Gemini 3.5 Flash-Lite or Luna. There is no consumer app for it; you reach it through API tools, Qwen Cloud's playground or QwenWork.
“Brand new, mostly self-graded, and slow. One independent lab agrees it is smart, but in-house suites and a corrected task set carry the rest.”
Independent evidence is thin, as it is for any model a few weeks old. Artificial Analysis corroborates the headline: an index of 40, GPQA 92.3%, HLE 38.0% and Terminal-Bench scores ahead of its vendor's own claims. Everything else comes from Alibaba. CoWorkBench and JobBench are in-house. SWE-bench Pro ran on a task set Alibaba corrected itself, and the DeepSeek baseline is the superseded V4-Flash-0731. The weights carry an "experimental preview" label, and there is no LMArena rating yet. Low list price is also offset by verbosity. The model is probably as capable as claimed on reasoning; treat the agentic claims as unproven until third parties reproduce them.
The full research notes behind this review — verified against primary sources.
Qwen3.8-Flash-Next is a sparse mixture-of-experts decoder with a vision encoder. The main model has 125B parameters, 6B active per token, with 512 experts of which 10 routed plus 1 shared fire per token (expert width 640). On top of that sit 51B N-gram Embedding parameters and a 4B multi-token-prediction module, so the checkpoint holds about 180B parameters in total. The 48 layers follow a repeating block of three Gated DeltaNet layers and one Qwen Sparse Attention layer, each feeding a MoE layer. Gated DeltaNet compresses history into a fixed-size state. QSA handles precise retrieval: a lightweight indexer (4 query heads, 1 shared key head) scores the context in micro-blocks and attends to a budget of 512 blocks or 2,048 tokens, which cuts both attention and indexing cost at long range. Gated Residual widens the residual stream into four gated branches. The N-gram Embedding looks up bigrams and trigrams from a 20M-entry table at layer 2; because the lookups are known in advance, Alibaba says the table can live in host memory and be prefetched asynchronously. Hidden size is 2,560 and the vocabulary is 248,320 tokens (padded). The weights natively handle 262,144 tokens and extend to 1M with YaRN; the hosted Qwen3.8-Flash runs at 1M by default and adds Alibaba's built-in tools. Training-token counts, data mix and knowledge cutoff are not disclosed; the model card links a technical report on the architecture and training recipe.
Coding (8.6): on Alibaba's table the weights score SWE-bench Pro 62.5, SWE-bench Multilingual 81.0, DeepSWE 1.1 58.7 and LiveCodeBench v6 91.9. Artificial Analysis independently measures 86.1% on Terminal-Bench 2.1 and 25.3% on the harder Terminal-Bench 4.0, second only to GLM-5.3-Flash (32.8%) among the Flash-tier peers. Reasoning (8.6): GPQA Diamond 92.3% and Humanity's Last Exam 38.0% in Artificial Analysis testing, slightly above Alibaba's own 91.7 and 35.9. Math (8.2): no competition-math result is published; MathVision 90.6 (95.7 with code interpreter) and the base model's MATH 72.8 are the only signals. Agentic (8.6): vendor-reported CoWorkBench 73.9, JobBench 55.7, Toolathlon Verified 73.5 and AndroidWorld 84.5. Artificial Analysis puts GDPval-AA at an Elo of 1612 and its agentic banking tool-use test at 45%. Long context (8.0): 1M tokens hosted, 79.7% on Artificial Analysis's long-context reasoning test (AA-LCR), a little behind DeepSeek V4.1-Flash (84%) and GPT-5.6 Luna at max effort (84%). Multilingual (8.5): SWE-bench Multilingual 81.0, and the base model scores MMMLU 84.9, MGSM 89.3 and INCLUDE 78.4. Vision (8.4): MMMU-Pro 79.8% (Artificial Analysis), RealWorldQA 88.5, LVBench 76.6 for long video, and ClawEval-MM 64.4 pass@3 for multimodal tool use (vendor). Document and OCR (7.8): CharXiv reasoning 84.6 and a hosted PDF-parsing tool, with no document-parsing benchmark published for this model. Instruction following (8.3): IFBench 81.3 (vendor). Function calling (8.5): OpenAI- and Anthropic-compatible tool calling, parallel calls, structured outputs and a Toolathlon lead over DeepSeek-V4-Flash-0731 (73.5 vs 70.3 on Alibaba's table). Creative writing (7.0): no evidence either way. Safety calibration (6.8): Artificial Analysis measures a 45% hallucination rate on AA-Omniscience, and Alibaba publishes no safety evaluation. Realtime data (6.0): built-in web search, web extraction and image search on the Beijing and Singapore endpoints, not in the Global-scope regions.
| Benchmark | Score | vs Predecessor | vs Top Competitor | Source |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index v4.3 | 40 (39.8) | Qwen3.7-Plus 25 | GLM-5.3-Flash 42, DeepSeek V4.1-Flash 39, GPT-5.6 Luna 37 at max effort (32 at high), Gemini 3.5 Flash-Lite 22 | Artificial Analysis |
| GPQA Diamond | 92.3% (Artificial Analysis); 91.7 (vendor) | Qwen3.7-Plus 90.3 (vendor) | GLM-5.3-Flash 91.2%, GPT-5.6 Luna max 91.1% | Artificial Analysis |
| Humanity's Last Exam | 38.0% (Artificial Analysis); 35.9 (vendor) | Qwen3.7-Plus 34.7 (vendor) | GLM-5.3-Flash 39.9%, GPT-5.6 Luna max 39.5%, DeepSeek V4.1-Flash 39.2% | Artificial Analysis |
| Terminal-Bench 2.1 | 86.1% | no published predecessor figure | GLM-5.3-Flash 84.3%, GPT-5.6 Luna max 80.9% | Artificial Analysis |
| Terminal-Bench 4.0 | 25.3% | no published predecessor figure | GLM-5.3-Flash 32.8%, DeepSeek V4.1-Flash 26.8%, GPT-5.6 Luna max 11.6% | Artificial Analysis |
| MMMU-Pro | 79.8% | Qwen3.7-Plus 80.5% | Gemini 3.5 Flash-Lite 79.0%, GPT-5.6 Luna max 78.6% | Artificial Analysis |
| AA-LCR (long context) | 79.7% | Qwen3.7-Plus 73.0% | DeepSeek V4.1-Flash 84.0%, GPT-5.6 Luna max 83.7% | Artificial Analysis |
| LiveCodeBench v6 | 91.9 (vendor) | Qwen3.7-Plus 89.6 | DeepSeek-V4-Flash-0731 90.6 on Alibaba's table | Qwen blog |
| SWE-bench Pro | 62.5 (vendor) | Qwen3.7-Plus 55.8 | DeepSeek-V4-Flash-0731 56.0 on the same harness | Qwen blog |
| SWE-bench Multilingual | 81.0 (vendor) | Qwen3.7-Plus 75.8 | Claude Opus 4.6 (Max) 77.5 | Qwen blog |
| Toolathlon Verified | 73.5 (vendor) | Qwen3.7-Plus 50.6 | DeepSeek-V4-Flash-0731 70.3 | Qwen blog |
| IFBench | 81.3 (vendor) | Qwen3.7-Plus 79.1 | DeepSeek-V4-Flash-0731 79.2 | Qwen blog |
| AndroidWorld | 84.5 (vendor) | Qwen3.7-Plus 81.0 | Claude Opus 4.6 (Max) 62.0 | Qwen blog |
Alibaba's figures are for the Qwen3.8-Flash-Next weights, run by Alibaba: SWE-bench Pro on the Claude Code harness over a task set Alibaba corrected and re-ran for every baseline, and HLE judged by GPT-4o. Its DeepSeek comparison is the older DeepSeek-V4-Flash-0731, not the current V4.1-Flash. Artificial Analysis lists the model as Qwen3.8-Flash-Next at the Qwen3.8-Flash API price. CoWorkBench and JobBench are Alibaba in-house benchmarks. The base-model table (MMLU 90.4, MMLU-Pro 73.2, BBH 90.9) describes the pre-trained checkpoint, not the model you call, so those fields stay null. No LMArena, SWE-bench Verified, AIME or MMMU figure exists for this model.
Slow for its tier. Artificial Analysis measures a median of 56.6 output tokens/sec (5th to 95th percentile: 28 to 72) and 2.72 seconds to the first token. Because thinking runs at xhigh effort by default, the first answer token arrives after a median 38.1 seconds, and a 500-token answer takes about 47 seconds end to end. Same-tier rivals generate far faster in that testing: Gemini 3.5 Flash-Lite at about 389 tokens/sec, DeepSeek V4.1-Flash at 226, GPT-5.6 Luna at about 150 to 165 depending on effort, and GLM-5.3-Flash at 90. It is also verbose: running the Intelligence Index produced 245M output tokens. For interactive use, drop reasoning_effort to low or send enable_thinking: false. Alibaba warns that lower effort can raise total task time in multi-turn agent work through retries. The architecture's advantage is long-context prefill rather than decode speed: Alibaba reports 8.6x Qwen3.7-Plus's prefill throughput at 1M tokens with heavy prefix caching.
| Surface | Cost | Notes |
|---|---|---|
| API input | $0.15 / 1M tok | Singapore (International), flat from 0 to 1M tokens |
| API output | $0.47 / 1M tok | Chain-of-thought tokens bill as output |
| Cached input | $0.016 / 1M tok | Same rate for implicit cache hits and explicit cache reads (about 89% off) |
| Cache write | $0.20 / 1M tok | Explicit cache creation |
| Batch (in/out) | not offered | Batch Inference is listed as unsupported in all regions |
| Other regions | $0.113 / $0.382 | China (Beijing), and the Global-scope deployments in Hong Kong, Frankfurt, Tokyo and Virginia |
| Direct UI | none | QwenWork's Standard mode runs on Qwen3.8-Flash-Next; no consumer plan is sold for the model itself |
| Free tier | 1M tokens | Singapore region only, valid 90 days from activation or model release |
| Rate limits | 15K RPM / 2M TPM | Qwen Cloud listing. Singapore and Beijing use dynamic limits tiered by monthly spend; Global-scope regions list 30,000 RPM / 5M TPM |
List price is not task price. Artificial Analysis's agentic evaluations are input-heavy, and there Qwen3.8-Flash cost about $0.37 per index task. That is more than GLM-5.3-Flash ($0.25), DeepSeek V4.1-Flash ($0.27) or GPT-5.6 Luna at max effort ($0.18), despite the lowest per-token list price of the four. Cache hits made up most of its bill, so prompt structure matters more than the headline rate.
Managed: Alibaba Cloud Model Studio and Qwen Cloud serve qwen3.8-flash through OpenAI-compatible Chat Completions and Responses APIs, the DashScope API, and an Anthropic-compatible endpoint. That makes it a drop-in backend for Claude Code, Codex, Qwen Code, Qoder CLI and OpenClaw. Regions include China (Beijing), Singapore (International), and Global-scope deployments in Hong Kong, Frankfurt, Tokyo and Virginia. Built-in web search is only available in Beijing and Singapore. Alibaba's pages list no AWS, Azure or Google Cloud availability.
Self-hosted: the Qwen3.8-Flash-Next weights are on Hugging Face and ModelScope in BF16 (about 360 GB) and an official FP8 build (about 185 GB). SGLang, vLLM, KTransformers and TokenSpeed have recipes. At FP8 the whole checkpoint needs a multi-GPU node, for example four 80 GB GPUs. Alibaba says the 51B N-gram table can sit in host memory, which would shrink the GPU footprint where the serving engine supports it. Featherless AI serves the weights through Hugging Face's inference-provider routing. The hosted version adds 1M context by default and Alibaba's tools. Running the weights yourself starts from a 262,144-token native window and needs YaRN to reach 1M.
License (Qwen Community License 1.0, verified on the LICENSE file): use, modification, fine-tuning, hosting and resale are permitted free of charge, subject to two conditions. First, if the model or a derivative powers a commercial product or service with more than 100M monthly active users or more than US$20M in monthly revenue, the model name must be displayed prominently in its user interface. Second, if you or any affiliate runs a "Model as a Service" business (giving third parties API or hosted access to language-model inference or fine-tuning with meaningful control over inputs, parameters or training data) or an "AI Work Assistant" business (an independent product mainly for AI-assisted coding or office productivity, such as Qoder or QwenWork), you need a separate license from Qwen before any commercial use. This second clause has no revenue threshold. Purely internal use is exempt as long as neither the software, its outputs nor its capabilities reach any third party. Merely relaying requests to models hosted by others does not count as Model as a Service. Single-purpose tools such as translation, assistants built for other domains, and assistant features inside products whose main purpose is not coding or office work fall outside the "AI Work Assistant" definition. The copyright and permission notice must travel with every copy. Licensing questions go to model-business@notice.qwencloud.com.
Alibaba publishes no safety framework, system card, refusal statistics or compliance certifications for this model. The model pages do not state whether API inputs are used for training, so that field is null. Data handling follows Alibaba Cloud Model Studio's terms. Region choice sets where requests are processed, from mainland China (Beijing) to Singapore or the Global-scope sites in Frankfurt and Virginia. Independent evidence is limited to Artificial Analysis's AA-Omniscience test: a 45% hallucination rate and an index of -9.7. That is far better than GPT-5.6 Luna (about 93%) and DeepSeek V4.1-Flash (about 96%), but worse than GLM-5.3-Flash (about 28%). Self-hosting keeps data on your own infrastructure but brings the license obligations above.
Served through Alibaba Cloud Model Studio and Qwen Cloud with Python and Node.js examples for the OpenAI SDK, the DashScope SDK and cURL, plus an Anthropic-compatible endpoint. Alibaba documents set-ups for Claude Code, Codex, Qwen Code, Qoder CLI and OpenClaw, and QwenWork's Standard mode runs on the model. The weights run on SGLang, vLLM, KTransformers, TokenSpeed and Hugging Face Transformers, and Featherless AI hosts them through Hugging Face's inference-provider routing. Adoption is growing quickly: about 775K downloads of the BF16 weights and 518K of the FP8 build in the past month on Hugging Face, with 5.5K likes. That is well behind the Apache-licensed Qwen3.8-27B.
$0.15 per 1M input tokens and $0.47 per 1M output tokens in Singapore (International), flat up to 1M tokens, with cache reads at $0.016 and explicit cache writes at $0.20. Beijing and the Global-scope regions list $0.113 / $0.382. There is no batch discount, and new Singapore accounts get 1M free tokens for 90 days.
Qwen3.8-Flash-Next is the open-weight model; Qwen3.8-Flash is Alibaba's hosted production version of it, with 1M context by default and built-in tools. Benchmarks from Alibaba and Artificial Analysis are reported under the Flash-Next name.
Yes, under the Qwen Community License 1.0, with two conditions. Products above 100M monthly users or US$20M monthly revenue must display the model name. Any company that runs a Model-as-a-Service or coding/office AI-assistant business needs a separate license from Qwen before any commercial use, whatever its revenue; purely internal use is exempt.
Thinking defaults to xhigh effort, and Artificial Analysis measures about 57 tokens/sec with a 38-second median wait for the first answer token. Set reasoning_effort to low or medium, or disable thinking, for interactive steps.
Yes. Alibaba's Anthropic-compatible endpoint and OpenAI Responses API support both tools directly, as well as Qwen Code, Qoder CLI and OpenClaw.
The official FP8 checkpoint is about 185 GB and BF16 about 360 GB, so plan on a multi-GPU node, for example four 80 GB GPUs at FP8. SGLang, vLLM, KTransformers and TokenSpeed support it, and the 51B N-gram table is designed to be offloadable to host memory.
It depends on the endpoint: China (Beijing), Singapore (International), or the Global-scope deployments in Hong Kong, Frankfurt, Tokyo and Virginia. Built-in web search works only in Beijing and Singapore. Alibaba publishes no model-specific compliance certifications.
Alibaba positions 3.8-Flash as the stronger model, and its table compares it with the larger Qwen3.7-Plus rather than 3.7-Flash. On price, 3.7-Flash is cheaper for prompts under 256K tokens and keeps a 50% batch discount, while 3.8-Flash is cheaper above 256K.
Released the same day at almost the same price ($0.15 / $0.50) under an MIT license, with 320B total and 18B active parameters. It edges Qwen3.8-Flash on the Artificial Analysis Index (42 vs 40), Terminal-Bench 4.0 (32.8% vs 25.3%) and hallucination rate (about 28% vs 45%), and generates faster (about 90 vs 57 tokens/sec). Qwen wins on video input, cheaper cache reads and Alibaba's built-in tools.
MIT open weights (552B MoE) at $0.30 / $1.20 peak and half that off-peak, with 384K output tokens. It ties on the Artificial Analysis Index (39 vs 40), leads on long-context reasoning (84% vs 80% on AA-LCR) and streams about four times faster. It hallucinates far more (about 96% vs 45%). Qwen is cheaper at peak and handles video.
OpenAI's nano-tier model at $0.20 / $1.20 with a 1.05M context and a Batch API. It is far faster (about 150 to 165 tokens/sec) and cheaper per completed task in Artificial Analysis testing, but it scores lower (37 at max effort, 32 at the default high) and its AA-Omniscience hallucination rate is about 93% against Qwen's 45%. Pick Luna for speed and the OpenAI ecosystem, Qwen for capability per token.
$0.30 / $2.50 with audio and PDF input and roughly 389 tokens/sec. It is the faster, more multimodal utility model, but trails badly on the Artificial Analysis Index (22 vs 40) and on reasoning (HLE 18.8% vs 38.0%).
Primary references used to verify this review.
Last verified 2026-09-21