| Model profile | ||
|---|---|---|
| Model | Size | Approximate size |
| DeepSeek V4 FlashModel source: provider1,048,576 tokens max native context | 284Bactive 13B | 170.4 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 |
| M5 Ultra | 256 GBUnified RAM | 1200 GB/s | no | 58 tok/s4-bit quantization |
| DGX H200 | 1128 GBGPU HBM3e | 4800 GB/s | yes | 186 tok/s4-bit quantization |
| DGX Station | 748 GBCoherent Memory | 7100 GB/s | yes | 242 tok/s4-bit quantization |
| ET900N G3 | 748 GBCoherent Memory | 7100 GB/s | yes | 242 tok/s4-bit quantization |
| DGX B200 | 1440 GBGPU HBM3e | 8000 GB/s | yes | 261 tok/s4-bit quantization |
| GB200 NVL72 | 13400 GBGPU HBM3e | 8000 GB/s | yes | 261 tok/s4-bit quantization |
| GB300 NVL72 | 20000 GBGPU HBM3e | 8000 GB/s | yes | 261 tok/s4-bit quantization |
Demonstrates Colibri v2.1 installation, system configuration, concept diagrams, chat UI, monitoring, profiling, running Flint GGUF, and testing Qwen3.6-35B-A3B and DeepSeek V4 Flash on one GPU.
Tests DeepSeek-V4-Flash-Vision-Exp-MLX on Mac Studio M3 Ultra using Inferencer v2.2.7 for image-to-game generation, thinking modes, AAA graphics, face recreation, OCR, procedural assets, logic, coding
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.
Tests DeepSeek V4 Flash Vision using DeepSeek Harness on browser workflows, 3D wrestling games, Blender scenes, Godot, C++ skate games, cinematic games, and frontend design.
Ranks frontier models based on real use across front-end design, one-shot capabilities, cost, speed, and subscriptions, placing Fable 5 at the top.
Thorough testing of DeepSeek-V4-Flash-Vision-Exp.
Benchmarking DeepSeek V4 Flash 284B with DSpark on RTX PRO 6000 via llama.cpp measures 31 tok/s using multi-turn coding sessions and tests drafter placement in RAM versus VRAM.
Linking two DGX Sparks via high-speed cable pools their memory into approximately 220 GB unified VRAM, enabling local execution of quantized Kimi K3 and GLM models.
Running DeepSeek V4 Flash locally on two NVIDIA DGX Spark units while developing BridgeMind One and updating BridgeBench using Claude Opus 5, Fable 5, GPT 5.6 SOL, Kimi K3, Claude Code, and Codex.
Compares three open stalwart models: Qwen 3.8 Max, DeepSeek V4 Flash, and Kimi K3 through tests evaluating their performance capabilities.
Speed test of DeepSeek-V4-Flash-0731-MLX using Inferencer App v2.2.3 and DSpark on Mac Studio M3 Ultra 512 GiB.
Evaluates DeepSeek-V4-Flash-0731-MLX using Inferencer App v2.2.3 on M3 Ultra 512 GiB across local and cloud setups against GLM 5.2 benchmarks.
Tests DeepSeek-V4-Flash-0731 checkpoint for speed, coding quality, tool use, hallucinations and reliability, then compares results against Tencent HY3 and Qwen3.6-27B.
Benchmarks DeepSeek V4 Flash against GPT-5.6 Luna, Kimi K3, GLM 5.2, and DeepSeek V4 Pro via WoAI Benchmark in coding, AI agents, frontend development, Three.js, game generation, and UI design.
Tests DeepSeek V4 Flash on browser workflows, C++ games, FPS dev, frontend design, 3D CAD modeling, and the Steve’s PC Repair game test.
Tests DeepSeek-V4-Flash-0731, the official GA build.
Tests how many AI agents the ASUS ExpertCenter Pro ET900N G3 can run using its 748GB unified memory, 400Gb networking, and 1400 watt superchip.
Runs the 284B parameter DeepSeek V4 Flash model locally on an AMD Ryzen AI MAX+ 395 with 128GB unified memory using the Lucebox inference engine with DSpark speculative decoding.
Demonstrates DeepSeek V4 Flash inference on AMD Strix Halo via ds4 and ROCm/HIP, covering Q2/hybrid 4-bit configs under 128GB RAM, multi-node 4-bit clusters, and benchmarks against SWE Bench Mini.