AuditTested AIBy DanPublished Tested Updated
Qwen 3.8 Flash Next on my RTX 5080: 81 tokens a second with 202K tokens read
The report
Verdict
Targeted use: fast and correct, slow to start
At 202,164 input tokens it wrote 80.8 tokens a second and its code passed all 18 grader checks. The first visible answer took about 5 minutes, so it suits batch long-context jobs, not quick back-and-forth.
On the tier list: These runs measure speed, context and one coding task on my PC. Its S tier seat is my own verdict, not a score from this report.
| Run on the RTX 5080 | Input tokens | Writing speed | Grader |
|---|---|---|---|
| Short control | 375 | 88.2 tok/s | 9/9 |
| Long repair task | 202,764 | 83.1 tok/s | 9/9 |
| Long repair task | 217,894 | speed target met | 8/9 |
| Long repair task, clarified | 217,499 | 82.0 tok/s | 9/9 |
| Dependency task | 202,164 | 80.8 tok/s | 18/18 |
What I ran
Qwen 3.8 Flash Next is a 176.94B mixture-of-experts model that uses about 6.67B parameters per token. On October 20 and 21 I ran a 2-bit-class IQ2_XS build on the Strata local server, on my RTX 5080 alone. The target: at least 50 tokens a second while really using 200,000 input tokens at the highest effort setting, with correct code.
The hard task
The main test spread one function's rules and seven dependency files across a 202,164-token prompt, with an out-of-date decoy rule near the start. The task line named only the function. The answer passed all 18 grader checks with no repair, and matched the reference on 200,000 random inputs.
The catch
Speed held from 4K to 202K input, but the model thinks for a long time. With an 18,432-token thinking budget, the first visible answer took 305.7 seconds. It also holds about 41 GB of my 47 GB WSL memory, so heavy work beside it competes for RAM.
My verdict
Qwen 3.8 Flash Next is my CHAMPION seat in S tier. This report backs the speed and long-context side of that verdict; the 5-minute wait is the price.
Limits of this report
- One sample per condition at temperature 0.2, so run-to-run variance is unmeasured.
- One function of about 100 lines with 7 dependency files. Imports pointed to the files and the filler was unrelated code, so this is not repository-scale refactoring.
- This run used the original IQ2_XS checkpoint, not an abliterated (uncensored) build, so it says nothing about uncensored behaviour.
- The runtime cut thinking off at its budget; whether more thinking changes quality is untested.
- RTX 5080 only: the 5060 was not used, so there is no dual-GPU result.
Follow the lab
New test reports and tier-list changes get announced in my Discord's #news channel. The full videos are on YouTube.
Receipts
DanManREAL · AI lab
Test receipt
Qwen3.8-Flash-Next GSQ-RCO IQ2_XS (176.94B total, about 6.67B active per token)
- Writing speed at 202K input
- 80.8 tok/s19,424 accepted tokens in 240.4 s
- Input tokens actually read
- 202,164none truncated or reused; 229,376 context configured
- Code correctness
- 18/18frozen grader, no repair; 0 mismatches against the reference on 200,000 random inputs
- Reading speed
- 3,013 tok/s202,164 tokens in 67.1 s
- First visible answer
- 305.7 scomplete answer at 315.4 s; most of the wait is 18,432 thinking tokens
- Short control, 4,148 input tokens
- 80.6 tok/s · 18/18speed did not drop from 4K to 202K input
- Scored requests
- 7, of which 6 passedone 217,894-token run scored 8/9
- Memory
- about 41 GB of 47 GB WSL RAM3,202 MiB stayed free on the 5080
Verdict
Targeted use: fast and correct, slow to start
Tier list These runs measure speed, context and one coding task on my PC. Its S tier seat is my own verdict, not a score from this report.
danmanreal.com/tests/qwen-3-8-flash-next-200k-local-run
What was tested
Related
- AI Coding Agents Tier List
- Jarvis 27B on my own PC: 30 tokens a second, and it still recalls at 251K
- How I test AI tools
Independent testing. Sponsorships are always disclosed, and verdicts are never for sale.
Disagree? Bring your receiptsAll reports
- One inbox for my AI coding agents: triage took 34% less time, resuming work did not improveHead-to-headNov 20
- Mythos 5.2: research and security benchmark source reviewPublic benchmarkNov 17
- Fable 5.2: agentic coding benchmark source reviewPublic benchmarkNov 17
- Claude Opus 5.5: public coding benchmark receiptsPublic benchmarkNov 17
- Astra 6.2: new model and public task-evaluation evidencePublic benchmarkNov 17
- Sol 6.2: new model and public coding benchmark evidencePublic benchmarkNov 17
- Gemini 4 Carbon: benchmark source and usage profilePublic benchmarkNov 17
- Qwen 4: Qwen4-Max launch and benchmark sourcesPublic benchmarkNov 17
- Kimi K3.1 Pro: public coding benchmark receiptsPublic benchmarkNov 17
- GLM-5.3: public coding benchmark receiptsPublic benchmarkNov 17
- Gemini 4 Argon: public coding benchmark receiptsPublic benchmarkNov 17
- Claude Sonnet 5.5: public coding benchmark receiptsPublic benchmarkNov 17
- DeepSeek V4.1 Flash: public coding benchmark receiptsPublic benchmarkNov 17
- Muse Wonder Full 1.0: public coding benchmark receiptsPublic benchmarkNov 17
- Grok 4.7: separate release and public coding resultsPublic benchmarkNov 17
- ai-design-check on two of my FABLE screens: 10 warnings, 4 of them false alarmsField testNov 2
- Vercel Agent Desktop used a sample Trip app in 6 actions without moving the mouseField testNov 1
- Cerberus Guard on sample data: one crawler name, two verdicts, and a block earned by behaviorField testOct 31
- GPT-6.1 Terra vs Opus 5.5: three real projects, and the designs I picked blindHead-to-headOct 29
- Steps → Reduced: opening a terminal agent went from 98 seconds to 11AuditOct 28
- One ChatGPT chat directed Opus and Codex at once through DMR_REMOTEField testOct 28
- Claude Duet ran my channel for a week and caught 269 fake AI crawlersField testOct 27
- Kimi K3.1 Pro built my animation family, after a quota stop and 21 minutes of failed requestsField testOct 24
- ChatGPT + Codex built 33 new DMR_REMOTE toolsField testOct 23
- Gemini 4 Argon reviewed a finished episode: 14 clips, one flag, one inconsistencyField testOct 21
- Qwen 3.8 Flash Next on my RTX 5080: 81 tokens a second with 202K tokens readAuditOct 21
- Does an extra AI review step pay off? A small DeepSeek trial said noHead-to-headOct 19
- My real DeepSeek bill: 12.31 billion tokens for $122.28Cost receiptOct 17
- Gemini 4 Argon kept hanging in Gemini CLI. The cause was a 60-second timeoutIncidentOct 14
- I asked 10 AI coding agents for their biggest weaknessSelf-reportOct 14
- Jarvis 27B on my own PC: 30 tokens a second, and it still recalls at 251KAuditOct 10
- My agents' first game benchmark was wrong. Here's how it was caughtIncidentOct 10
- Pixal3D + Blender turned my 2D art into animated 3D charactersField testSep 22
- DeepSeek Flash answered 200 OK and sent nothing for two hoursIncidentSep 14
- Grok, Codex and DeepSeek played WorldOS for Rank 1. Grok won with 98Head-to-headAug 20
- DeepSeek V4 Flash audited 18 of my real coding sessions. Two said 'done' too earlyAuditAug 4