| Model profile | ||
|---|---|---|
| Model | Size | Approximate size |
| Qwen 3.8 Flash NextModel source: provider262,144 tokens max native context | 125Bactive 6B | 75 GB4-bit quantization |
| Compatible hardware Estimated speeds, not benchmark results: calculated from memory bandwidth and model size. Real results can differ significantly because there is no precise formula for deriving LLM generation speed from hardware specifications alone. | ||||
|---|---|---|---|---|
| Hardware | Memory | Bandwidth | vLLM | Estimated generation |
| Radeon 8060S 96GB | 96 GBUnified RAM | 256 GB/s | yes | 28 tok/s4-bit quantization |
| DGX Spark | 128 GBUnified RAM | 273 GB/s | yes | 30 tok/s4-bit quantization |
| M5 Max | 128 GBUnified RAM | 614 GB/s | no | 64 tok/s4-bit quantization |
| M3 Ultra | 96 GBUnified RAM | 819 GB/s | no | 83 tok/s4-bit quantization |
| M5 Ultra | 256 GBUnified RAM | 1200 GB/s | no | 115 tok/s4-bit quantization |
| RTX PRO 6000 | 96 GBGPU VRAM | 1792 GB/s | yes | 159 tok/s4-bit quantization |
| DGX H200 | 1128 GBGPU HBM3e | 4800 GB/s | yes | 304 tok/s4-bit quantization |
| DGX Station | 748 GBCoherent Memory | 7100 GB/s | yes | 369 tok/s4-bit quantization |
| ET900N G3 | 748 GBCoherent Memory | 7100 GB/s | yes | 369 tok/s4-bit quantization |
| DGX B200 | 1440 GBGPU HBM3e | 8000 GB/s | yes | 388 tok/s4-bit quantization |
| GB200 NVL72 | 13400 GBGPU HBM3e | 8000 GB/s | yes | 388 tok/s4-bit quantization |
| GB300 NVL72 | 20000 GBGPU HBM3e | 8000 GB/s | yes | 388 tok/s4-bit quantization |
| Benchmark reports for this model | ||||
|---|---|---|---|---|
| Model | Date | Status | Tokens | Score |
SC# Coding & Agentic Benchmark by Ai4Users A custom Ai4Users benchmark suite for SC#, covering code generation, reasoning, and stateful agentic workflows where models use tools, observe results, adapt their actions, recover from failures, and verify the final state. | ||||
| 2026-09-26 | Finished | IN4 398 019 OUT122 411 |
97.6% | |
Qwen 3.8 Flash NextNVFP4Think: medium857 AI queries
372 tool calls
331 total
152 passed176 failed3 broken |
2026-09-26 | Finished | IN8 520 134 OUT1 758 976 |
45.9% |
HumanEval for SC# HumanEval adapted to SC#, adding an extra challenge: models must generate correct code in a custom sandboxed language while following its documentation and constraints. | ||||
| 2026-09-26 | Finished | IN1 412 675 OUT28 074 |
95.7% | |
Qwen 3.8 Flash NextNVFP4Think: medium162 AI queries
0 tool calls
162 total
35 passed126 failed1 broken |
2026-09-26 | Finished | IN1 417 211 OUT186 262 |
21.6% |
LiveCodeBench for SC# LiveCodeBench adapted to SC#, adding an extra challenge: models must solve newer coding problems while generating correct code in a custom sandboxed language from its provided specification. | ||||
Qwen 3.8 Flash NextNVFP4Instruct1022 AI queries
0 tool calls
1054 total
550 passed472 failed32 skipped |
2026-09-26 | Finished | IN9 457 973 OUT728 528 |
53.8% |
Instruction Following Benchmark based on IFBench IFBench-based benchmark for precise instruction following, ported to the SC# evaluation runtime. It checks whether models follow verifiable formatting, structure, wording, length, and other response constraints. | ||||
Qwen 3.8 Flash NextNVFP4Instruct271 AI queries
0 tool calls
300 total
123 passed148 failed29 skipped |
2026-09-26 | Finished | IN256 375 OUT59 835 |
45.4% |
Tool Calling Benchmark based on BFCL A BFCL-based tool-calling benchmark ported to the SC# evaluation runtime. It tests whether the model selects the correct function, uses the right arguments, and avoids calling tools that do not match the request. Each test can see all tools declared in its script, so the model often has more tools to choose from than in the original case. | ||||
Qwen 3.8 Flash NextNVFP4Instruct1248 AI queries
1421 tool calls
1265 total
1156 passed92 failed17 skipped |
2026-09-26 | Finished | IN8 952 270 OUT198 020 |
92.6% |
Reasoning Benchmark based on BBH BBH-based reasoning benchmark ported to the SC# evaluation runtime. It tests multi-step reasoning across challenging tasks involving logic, deduction, symbolic manipulation, planning, and problem solving. | ||||
Qwen 3.8 Flash NextNVFP4Instruct5388 AI queries
0 tool calls
5390 total
5172 passed216 failed2 skipped |
2026-09-26 | Finished | IN5 646 466 OUT2 801 620 |
96.0% |
Benchmarks Qwen3.8 27B and Flash Next on one/two Intel Arc Pro B70 GPUs via llama.cpp, vLLM, Strata, comparing prompt processing and generation against Radeon AI PRO R9700 and Strix Halo.
Evaluates whether a locally-hosted model can compete with Kimi K3.1 Pro, described as the largest open-weights model at 2.8T parameters, through challenges and code reviews.
Demonstrates Qwen3.8-Flash-Next controlling a Seeed Studio reBot Arm B601 via plain-English instructions for throwing, shooting, writing, drawing and pouring, using Meta SAM 3.1 for vision.
Compares updated Strata Coder and Turbo Brilliance via 9 tests: speed, strategy game, tool calling, CAPTCHA, 262K context, SVG, voxels, motion design and physics puzzles.
Runs Inferencer v2.5.0 with DeepSeek V4.1 SE and Qwen 3.8 Flash SE on M1 Max, M5 Max and Mac Studio M3 Ultra hardware.
Tests Strata Coder IQ1_M against Qwen 3.8 Flash Next NVFP4 using 9 benchmarks: speed, battle strategy, tool use, visual CAPTCHAs, long-context retrieval, SVG, voxels, motion and physics.
Tests Strata 0.2 with Qwen3.8-Flash-Next via Apple Metal, covering setup, model loading, memory needs, and prompt processing plus generation speed.
Tests DeepSeek-V4.1-Flash-1017 against Qwen3.8-Flash-Next on a robotic arm 3D model and compares its World War I voxel art output with the original V4.1 Flash.
Tests NVIDIA Nemotron 3.5 Bolt on coding and agentic work, comparing performance against Qwen3.8-27B, Qwen3.8-Flash-Next, and Strata Coder.
Integrates Strata Metal Inference Engine with Pi Coding Agent to perform .NET migrations, code fixes, UI modernization, and application execution using Qwen3.8-Flash-Next on an Apple M5 Max.
Tests Qwen3.8 27B Q4 on Blender and Godot, evaluates Qwen3.8 Flash Next via Strata, and compares local performance against Claude Opus 4.6 using an ASUS ROG system with RTX 5090.
Benchmarks Qwen3.8 27B and Flash Next on single/dual AMD Radeon AI PRO R9700 using GGZ14, Radiance, R9V, tcclaviger vLLM, llama.cpp and vLLM via AI Toolbox Cockpit.
Tests whether Qwen 3.8 Flash Next Coder uses less memory than Qwen 3.8 Flash Next through head-to-head coding challenges using Strata.
Tests Strata, an open-source AI inference engine with native Apple Metal backend, against LM Studio using Qwen3.8-Flash-Next on Apple M5 Max with 128GB unified memory.
Seeed Studio reBot Arm B601 executes tasks via Qwen3.8-Flash-Next reasoning, Meta SAM 3.1 vision, and a Jev-like decision model using plain-English voice instructions.
Develops and tests a UI framework for Dust to Dominion using Qwen 3.8 Flash Next via Strata within a spec-driven vibe coding experiment.
Tests running the 125B parameter Qwen3.8-Flash-Next using Strata across various VRAM configurations to evaluate its performance on low VRAM setups.
Evaluates Qwen3.8-Flash-Next with Strata on an RTX 4060 Ti 16GB via Pi agent across three coding tasks while measuring generation speed.
Demonstrates Strata running Qwen 3.8 Flash on RTX 3060, achieving 900+ tok/s prompt processing via a voice agent placing a grocery order in under two minutes.
Tests Qwen3.8 variants on RTX PRO 6000 via SGLang/vLLM across 10 tasks including BFCL, long context, SVG, video editing, voxels, motion design, Rube Goldberg physics and CAPTCHAs.
ISTA-DASLab demonstrates running QWEN3.8-FLASH-NEXT-GSQ-RCO in LM Studio and compares its performance against QWEN3.8-Q4 for .NET migration and SVG image generation tasks.
Compares Space Bunny Alpha via OpenRouter against local Qwen Q9 using Open Code, Inferencer, and Mac Studio M3 Ultra 512 for destruction animations, procedural faces, and live coding tasks.
Tests dual RTX 3060 performance using llama.cpp tensor split and Raylight for Qwen 3.6 35B MoE, Qwen 3.8 27B, and MiniMax H3 video generation within a Proxmox/ZFS environment.
Qwen 3.8 Flash-Next ran on Halogen Flash Server via 72 coding jobs across OpenCode, dsh, and Pi harnesses on a 128 GB Strix Halo system.
Evaluates Claude Code, OpenCode and native Inferencer using Qwen4 Exp on Mac Studio M3 Ultra via Tic-Tac-Toe generation, modifications, image inferencing and token usage comparisons.
Tests Qwen3.8-Flash-Next on M5 Max using MTPLX v2.11.3, which fixes 8 defects in decoding, caching, tokenization, sampling, state restoration, and tool handling.
Tests Qwen 3.8 Flash-Next generation speed across OpenCode, Pi and DeepSeek Harness using 18 benchmark jobs at xhigh thinking level on a GMKtec EVO-X2 with Strix Halo hardware.
Tests Qwen3.8-Flash-Next on RTX PRO 6000 with SGLang, measuring speed, BFCL, tau^2-bench, long context, strategic reasoning, SVG drawing, video editing and motion design capabilities.
Tests oMLX v0.7.0.dev2 with Qwen3.8-Flash-Next 125B MoE on an M5 Max with 128GB unified memory, measuring inference speeds up to ~70 tokens/sec.
Tests performance, memory, reasoning, OpenAI Human Eval, Kanban, Sand Physics, Dungeon Crawler, Blender and Godot for unsloth/Qwen3.8-Flash-Next-GGUF on a 16GB local setup.
Compares Halogen Flash Server, EngramHalo and ROCm/Vulkan llama.cpp for Qwen3.8-Flash-Next on AMD Strix Halo via Terminal Bench Mini agentic tasks and speed benchmarks.
Compares Qwen3.8-Flash-Next inference via llama.cpp, SGLang, and FreeToken on RTX PRO 6000 Blackwell using AIPerf, measuring first-token latency, decode speed, accuracy on GSM8K/MATH-500, and startup
Demonstrates Qwen3.8-Flash-Next on 4x RTX 3090s using DeepSeek Harness with MiniMax H3 for autonomous music video creation, comparing throughput against Qwen3.8-27B.
Benchmarks unsloth/Qwen3.8-Flash-Next-GGUF on RTX 3060 using llama.cpp and thecodacus/spec-wins coding lab with contradictory specs and decoys.
Tests MTPLX v2.10.0 optimizations for local Qwen3.8-Flash-Next coding agents via .NET migration and SVG image generation comparisons against oMLX.
Benchmarked Qwen3.8-Flash-Next in llama.cpp across 0–96GB VRAM using AMD Ryzen 9 9950X and NVIDIA RTX PRO 6000 Blackwell, measuring decode speeds, prefill gains, and configuration impacts.
Compares DeepSeek V4 Flash and Qwen 3.8 Flash Next using Unsloth UD-IQ1_S 1-bit builds on 19 hard problems via CPU-only and RTX 5090 expert layer offload benchmarks.
Hands-on comparison of GLM-5.3-Flash vs Qwen3.8-Flash evaluating bug fixing, vision coding capabilities, and AI flirting performance through three distinct tests.
Compares MLX/oMLX and GGUF with llama.cpp for local inference of Qwen3.8-Flash-Next 125B on Apple M5 Max, achieving ~60 tokens/sec with oMLX 0.6.3 and Lightning MTP.
Tests Qwen 3.8-Flash-Next 1-bit to 4-bit IQ4_XS via llama.cpp on GMKtec EVO-X2 against Qwen 3.8-27B references using four one-shot tests, reporting ~20 tok/s and peak memory usage.
Benchmarks Qwen 3.8 Flash Next on Intel Core Ultra 9 285K CPU versus RTX 5090 offloading using llama.cpp, measuring throughput and perplexity for 1-bit and 2-bit quantizations.
Evaluates Qwen3.8-Flash-Next-MLX-Q9 and Q4 variants using Inferencer App v2.3.4 on an M3 Ultra 512 GiB system via local testing.
Tests Qwen3.8-Flash-Next coding and Splunk Cyber CTF performance, investigates hallucinations caused by Qwen Sparse Attention, and demonstrates fixing issues via sparse attention settings.
Reviews Qwen 3.8-Flash-Next specs and Unsloth GGUF sizes.
Tests Qwen3.8 Flash Next on browser workflows, C++ games, Blender, Godot, local Q4, FPS generation, and 3D scenes.
Locally installs and tests Qwen3.8-Flash-Next.
Explores Qwen3.8-Flash-Next architecture and runs local inference via Unsloth GGUF and llama.cpp on Apple Silicon, covering 1-bit and 4-bit quantization setups.
Compares Qwen 3.8 Flash Next and Qwen 3.8 27B architectures, detailing MoE, sparse attention, gated residuals, 51B n-gram lookup tables, and RTX 5090 MTP measurements.