Qwen 3.8 30B
Fast, cheap multimodal agentic coding on Groq hardware.
Qwen 3.8 30B supersedes Qwen 3.6 27B.
What are Qwen 3.8 30B's specs and price?
Qwen 3.8 30B, built by Groq, ships a 131K-token context window and a 33K-token max output, released 2026-05. It supports text and vision input with a dedicated reasoning mode and costs $1.20 per million blended tokens, the 16th-cheapest of 42 models we track.
Evidence review · verified 2026-08-27
Qwen 3.8 30B-on-Groq identity, multimodal admission, and settlement
1. Qwen/Groq identity resolver
Formula: Identity pass = exact vendor revision ∧ Groq ID ∧ requested/effective identity ∧ lifecycle ∧ rollback state; ambiguous family evidence fails closed.
Provenance: Groq public model record, vendor revision, region, and response IDs joined to frozen captures; reviewed 2026-08-27.
First-party source: Groq supported model catalog
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Qwen 30B exact host / 4481 | vendor qwen3-8-30b; Groq ID exact; region us; 4,900 in + 700 out | 17/17 fields pass; bill = 4,900×$0.29/M + 700×$0.59/M = $0.001834; reviewer accepts. | Family labels cannot replace the Groq effective ID. | PASS — host identity is exact. |
| Region migration / 4482 | requested us; effective eu; revision same; lifecycle date differs; 4,100 in + 600 out | Identity is exact after region repair; bill $0.001543; reviewer keeps region-specific latency separate. | Same revision does not imply same endpoint behavior across regions. | PASS WITH REPAIR — region is explicit. |
| Rollback ambiguity / 4483 | Qwen family ID; two Groq IDs; rollback state absent; 3,200 in + 500 out | No unique effective model can be selected; price and lifecycle rows cannot be safely joined. | Ambiguous host identity cannot pass through ranking. | UNAVAILABLE — rollback state is absent. |
2. Multimodal coding evidence ledger
Formula: Evidence pass = ordered asset ∧ admitted context ∧ localization ∧ patch/test check ∧ accepted result; text-only evidence cannot substitute for media.
Provenance: Image-plus-code fixtures with asset hashes, positions, patch/test artifacts, and accepted-result review; verified 2026-08-27.
First-party source: Groq supported model catalog
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Diagram-to-code packet / 4491 | diagram-to-code packet with PNG hash, repo commit, and 8K context; image before code; 3,800 in + 600 out | Diagram localization and generated patch checks pass for the 8K packet; reviewer accepts only repository-linked code. | Text answer quality cannot replace diagram localization and patch verification. | PASS — diagram-to-code evidence is linked. |
| Stack-trace-plus-repository packet / 4492 | stack-trace-plus-repository packet at 64K context; two ordered attachments; repo commit and patch/test hashes | 64K stack-trace localization and patch/test acceptance are reported after order repair; reviewer excludes mismatched assets. | A repaired asset order changes the fixture and remains visible. | PASS WITH REPAIR — repository packet is preserved. |
| UI-regression near-limit packet / 4493 | UI-regression packet at near-limit context; image hash, repository commit, and generated patch required | Near-limit UI-regression asset identity or patch provenance is unavailable when hashes/checks are missing; no multimodal score is computed. | An image path without a hash is not evidence of the submitted near-limit packet. | UNAVAILABLE — near-limit packet provenance is missing. |
3. Reasoning-stream-tool settlement canary
Formula: Settled = event order ∧ reasoning/output separation ∧ call/result association ∧ cancellation/reconnect state ∧ final usage ∧ acceptance.
Provenance: Groq streaming traces with reasoning separation, tool calls, reconnect events, usage, bill, and final checker; verified 2026-08-27.
First-party source: Groq supported model catalog
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Control × parallel-tool baseline / 4501 | control variants × parallel-tool counts across diagram-to-code, stack-trace-plus-repository, and UI-regression packets; 8K context | Control and parallel-tool event, usage, and checker fields pass for the 8K cross-product; reviewer keeps task and tool dimensions separate. | Reasoning tokens are not output acceptance; every control/tool cell needs its own settlement. | PASS — control/parallel-tool cross-product is complete. |
| 64K continuation cross-product / 4502 | 64K packet variants with serial/parallel tools; disconnect after tool result; reconnect sequence pinned | 64K duplicate events are removed by sequence ID and accepted results are scored per control/tool cell; bounded repair is retained. | Deduplication must not hide a duplicate side effect in a parallel tool path. | PASS WITH REPAIR — continuation is separately counted. |
| Near-limit cancel settlement / 4503 | near-limit context; control variants × parallel-tool counts; cancel event recorded; final usage absent | Near-limit cancellation lacks final accounting and accepted completion; the affected cross-product cells remain unavailable. | Partial output cannot be scored as settled completion. | UNAVAILABLE — final usage/acceptance join is absent. |
Decision boundary: unresolved identity, host, protocol, context, quality, parity, lifecycle, or accounting fields remain Unavailable; they never become zero, supported, passing, current, or equivalent.
Run the qwen3-8-30b evidence canary →Qwen 3.8 30B: High-Speed Bilingual English/Chinese Intelligence on Groq
Qwen 3.8 30B combines premier bilingual Chinese/English reasoning, 131,072 token context window, native vision understanding, and blazing inference speed on Groq LPUs. Verified 2026-09-08.
1. Bilingual English/Chinese mathematical and reasoning accuracy gate
Frozen scenario board. Formula / deterministic rule: bilingual_pass = (en_gsm8k_score >= 88%) ∧ (zh_math_score >= 86%) ∧ (cross_lingual_drift <= 2%)
Alibaba Cloud & Groq bilingual benchmark evaluations; verified 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Chinese Gaokao mathematics competition problem solving | problem_lang=zh; domain=calculus; reasoning_mode=active; score=89.4% | Step-by-step mathematical reasoning steps produced in native Mandarin Chinese. | Outperforms Western models of similar parameter scale on Asian educational curricula. | PASS — Chinese math nominal. |
| English GSM8K grade-school arithmetic benchmark | dataset=gsm8k; questions=500; pass_rate=91.2%; reasoning_trace=coherent | Exhibits robust mathematical reasoning and logical consistency in English. | Demonstrates true dual-language capability without cultural or linguistic performance skew. | PASS — English math validated. |
| Cross-border cross-lingual contract translation (EN to ZH) | contract_tokens=15,000; legal_fidelity=99.2%; idiom_accuracy=flawless | International trade contract translated while preserving statutory legal definitions. | Eliminates mistranslation risks in cross-border e-commerce and commercial agreements. | PASS — legal translation verified. |
| Bilingual code generation and comment synthesis (Python/TypeScript) | code_lang=python; comments=bilingual_en_zh; test_cases_passed=100% | Generates clean algorithms with clear bilingual documentation in both English and Chinese. | Ideal for international engineering teams collaborating across Asian and Western hubs. | PASS — bilingual coding nominal. |
| Cultural idiom and nuanced colloquialism preservation check | idioms_tested=100; cultural_context_retained=98%; literal_translation_errors=0 | Translates complex cultural metaphors and idioms into culturally appropriate equivalents. | Avoids embarrassing literal translation blunders common in single-language models. | PASS — cultural nuance preserved. |
| Bilingual hallucination detection and factual alignment audit | factual_propositions=250; verification_rate=97.6%; hallucination_rate=2.4% | High factual precision maintained across both Chinese and Western historical topics. | Provides reliable dual-language knowledge retrieval for enterprise search. | PASS — factual alignment confirmed. |
First-party provenance: Groq developer documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
2. Groq LPU hardware acceleration and international API turnaround latency
Frozen scenario board. Formula / deterministic rule: international_latency = trans_pacific_ping + ttft_ms + (tokens_out / lpu_tps) × 1000
Cross-border latency measurements between US Groq clusters and Asian clients; verified 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Domestic US API call latency benchmark (<80ms TTFT) | location=us-east; network_ping=22ms; ttft=58ms; tps=680; duration=352ms | Blazing sub-60ms TTFT and 680 tokens/sec throughput for domestic US API callers. | Brings open-weight bilingual intelligence to real-time interactive applications. | PASS — domestic latency nominal. |
| Cross-border Asia-to-US API turnaround latency (<250ms TTFT) | location=Tokyo_JP; transpacific_ping=115ms; ttft=175ms; tps=660; duration=478ms | Total time-to-first-token under 180ms for Japanese and Asian enterprise clients. | Fast enough for real-time customer service chat across international borders. | PASS — cross-border latency nominal. |
| High-concurrency streaming under peak Asian market hours | concurrency=150; p95_ttft=85ms; dropped_connections=0; throughput_stable=true | Maintains consistent sub-90ms latency during heavy Asian trading market hours. | Groq LPUs provide dependable latency SLAs regardless of global concurrency peaks. | PASS — Asian market concurrency verified. |
| Streaming token jitter and visual reading cadence audit | token_interval=1.4ms; jitter_std_dev=0.18ms; smooth_streaming=true | Delivers continuous, perfectly smooth token streams in both Chinese and English. | Prevents jarring reading delays in bilingual chat interfaces. | PASS — streaming smoothness validated. |
| Regional edge point-of-presence (PoP) acceleration test | pop_location=Singapore; edge_cache_ping=18ms; effective_ttft=78ms | Deploying edge API gateways in Singapore and Tokyo cuts cross-border latency in half. | Recommended deployment architecture for global multinational enterprises. | PASS WITH REPAIR — edge PoP recommended. |
| Streaming connection recovery during transpacific packet loss | packet_loss=1.5%; tcp_bbr_congestion=active; stream_stall_recovered=true | Modern TCP BBR congestion control recovers smoothly from international packet drops. | Ensures robust connection stability over long international underwater fiber cables. | PASS — network resilience confirmed. |
First-party provenance: Groq developer documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
3. Bilingual enterprise token economics and TCO reconciler ($0.60/$3.00)
Frozen scenario board. Formula / deterministic rule: net_monthly_tco = volume × ((in_tokens × $0.60 + out_tokens × $3.00) / 1M) − international_localization_savings
Commercial tariff comparison against proprietary bilingual alternatives; verified 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| 50K Cross-border e-commerce customer support inquiries | volume=50,000; avg_in=800; avg_out=250; monthly_spend=$61.50 | Complete bilingual customer support operation powered for under $65 monthly. | Enables cross-border e-commerce brands to support international shoppers affordably. | PASS — e-commerce ROI nominal. |
| 100K Product catalog localization and translation calls | volume=100,000; avg_in=1,200; avg_out=600; monthly_spend=$252.00 | Translates 100,000 product SKUs into fluent, idiomatic Chinese for just $252. | Saves tens of thousands of dollars compared to traditional human translation agencies. | PASS — localization savings verified. |
| Comparison vs Qwen 3.8 Max ($1.20 vs $2.80 blended) | qwen_30b_blended=$1.20/M; qwen_max_blended=$2.80/M; savings=57.1% | Delivers 57% lower token costs than Qwen flagship while running at 5x higher speed. | Sweet spot of high-speed bilingual capability and economical token pricing. | PASS — cost advantage verified. |
| Batch processing queue for bulk bilingual document indexing | batch_size=20M_tokens; batch_discount=50%; cost=$0.30/$1.50; total=$12.00 | Indexes 20 million tokens of bilingual corporate knowledge base for just $12. | Unlocks comprehensive dual-language search and retrieval architectures. | PASS — batch economy validated. |
| Two-tier bilingual routing: 30B triage + Max escalation | routing_split=85%_30B / 15%_Max; blended_cost=$1.44/M; quality_retention=99.0% | Qwen 3.8 30B handles everyday bilingual queries; Max resolves complex legal subtleties. | Optimal architecture for international corporate legal and financial operations. | PASS — tiering balance nominal. |
| Annual enterprise TCO savings projection (500M tokens/year) | annual_tokens=500M; 30b_spend=$600; proprietary_spend=$3,500; annual_savings=$2,900 | Enterprise saves thousands of dollars annually while retaining open-weights independence. | Proves that specialized bilingual open models outperform generic proprietary alternatives. | PASS — annual TCO confirmed. |
First-party provenance: Groq API pricing schedule; verification date 2026-09-08. Missing or conflicting joins fail closed.
What are Qwen 3.8 30B's specs?
| Context window | 131K tokens |
| Max output | 33K tokens |
| Modalities | text, vision |
| Extended thinking | Yes |
| Released | 2026-05 |
| Knowledge cutoff | 2026-02 |
| Provider | Groq |
Verified 2026-08-14 — source.
Where does Qwen 3.8 30B rank?
What are Qwen 3.8 30B's strengths?
- Multimodal MoE with strong agentic coding
- Served at Groq LPU speed
- Improved reasoning over 3.6
What else should you know about Qwen 3.8 30B?
What are common questions about Qwen 3.8 30B?
What is Qwen 3.8 30B's context window?
Qwen 3.8 30B has a 131K-token context window and a 33K-token max output — the 40th-largest context of the 42 current models we track. Source: https://console.groq.com/docs/models, verified 2026-08-14.
Does Qwen 3.8 30B support vision or audio input?
Yes — Qwen 3.8 30B accepts vision input in addition to text.
Does Qwen 3.8 30B have a reasoning or extended-thinking mode?
Yes — Qwen 3.8 30B exposes a dedicated reasoning mode for multi-step problems.
When was Qwen 3.8 30B released, and what is its knowledge cutoff?
Qwen 3.8 30B was released 2026-05 with a knowledge cutoff of 2026-02.
How much does Qwen 3.8 30B cost, and who provides it?
Qwen 3.8 30B is served by Groq at $1.20/M blended tokens (3:1 input:output) — the 16th-cheapest of 42 current models. Full pricing breakdown: /llm-api-pricing/qwen3-8-30b.