FULL STORY

DeepSeek V4-Flash: From Leak to Local Deployment Frenzy

Following an unexpected leak and API beta, DeepSeek V4-Flash sparked a local deployment frenzy. It delivers top-tier performance across benchmarks at a fraction of competitors' costs.

2026-07-31 ~ 2026-08-11 · 15 episodes · 88 posts

Episode 1 · DeepSeek-V4-Flash Architecture Leaked with Million-Token Context (2026-07-31, 3 posts)

Accidental leaks on Hugging Face and vLLM Recipes revealed the DeepSeek-V4-Flash model, featuring 284B total parameters, 13B active parameters, FP4 mixed precision, and native support for a one-million-token context window.

Episode 2 · Unsloth Releases Quantized DeepSeek V4 Flash 0731 for Local Deployment (2026-07-31, 6 posts)

Unsloth released GGUF dynamic quantizations of DeepSeek V4 Flash 0731 from July 31 to August 2, enabling local deployment. The model is based on the deepseekv4 architecture and is MIT-licensed. This release significantly lowers the barrier for local deployment, sparking community discussion on the trade-off between quantization precision and performance.

Confirmed

  • Unsloth published quantized versions of DeepSeek-V4-Flash-0731 on Hugging Face, with multiple dynamic quantization (UD) options ranging from 83GB to 162GB.
  • The 162GB UD-Q8KXL version is fully lossless, while the smallest 83GB IQ1S version retains 73% performance; the 97GB version is also a viable option.
  • Users can run the model via Unsloth or llama.cpp. Hardware requirements: 168GB RAM for lossless 4-bit quantization, 110GB RAM for 3-bit quantization.

Unconfirmed

  • Developer danielhanchen addressed community questions about quantization naming and precision differences, but the specific details and clarifications are not fully presented in the materials.

Why it matters

  • This release enables local deployment of DeepSeek V4 Flash 0731 on consumer hardware, offering flexible options for developers with different VRAM environments, which is significant for AI developers and researchers.

Episode 3 · DeepSeek on Huawei Ascend Beats OpenAI in Inference Profitability (2026-08-01, 2 posts)

DeepSeek achieves higher inference profitability on V4 Flash using Huawei Ascend chips than OpenAI does with Nvidia Blackwell. This is driven by their industry-leading KV cache offload system and top-tier low-level operator engineering, despite OpenAI's larger model size.

Episode 4 · DeepSeek-V4-Flash Excels in Frontend Coding with Unmatched Cost-Performance (2026-08-01, 3 posts)

DeepSeek-V4-Flash-High ranked 7th overall and 3rd among open-source models in frontend coding, achieving exceptional performance at just 3% of the cost of comparable models.

Episode 5 · DeepSeek V4 Preview: Flash to Introduce Four-Level Reasoning Effort (2026-08-01, 2 posts)

Developer discussions indicate that the current DeepSeek beta is actually a preview of V4-Flash. The upcoming series will introduce four distinct reasoning effort modes and is expected to integrate visual capabilities in future iterations.

Episode 6 · DeepSeek Drastically Reduces Training Compute Costs Across Models (2026-08-01, 2 posts)

DeepSeek has drastically reduced its training compute costs across model generations. The latest V4-Flash model requires only 66,000 GPU hours to train 1T tokens, a significant drop from the 300,000 hours needed for V1, showcasing extreme efficiency in MoE architecture.

Episode 7 · DeepSeek V4-Flash API Public Beta Launches with Major Agent Upgrades (2026-08-01, 5 posts)

DeepSeek officially announced the public beta of V4-Flash API, with significant improvements in agent capabilities and benchmark scores, some surpassing the previous V4-Pro-Preview, drawing widespread community attention.

Confirmed

  • Model architecture and specs: DeepSeek V4-Flash uses a 284B-parameter MoE architecture with 13B active parameters, supports 1M token context, and natively supports Codex.
  • Agent capability upgrade: Official statements indicate major upgrades in agentic coding and tool use.
  • Terminal-Bench score: The model scored 82.7 on Terminal-Bench, a 25.8-point jump from the April preview's 56.9.
  • Frontend Code Arena: DeepSeek V4 Flash High reached 7th overall (score 1586) and 3rd among open-source models. In subcategories, it ranked 4th in consumer products, and also high in reference design, data analysis, and games.

Unconfirmed

  • According to leaked data cited by @teortaxesTex, DeepSeek V4 Flash scores more than doubled compared to the previous preview, and V4 Pro Preview also saw a large leap, but full benchmark details await official verification.

Why it matters

  • V4-Flash, while remaining open-source, significantly enhances agent execution and tool-use capabilities, marking a key step for DeepSeek in practical coding AI competition and providing developers with a stronger open-source alternative.

Episode 8 · DeepSeek V4-Flash Costs 105x Less, But Stability and Benchmark Overfitting Questioned (2026-08-01, 17 posts)

DeepSeek V4-Flash-0731 has recently demonstrated exceptional cost-effectiveness on multiple benchmarks, with Artificial Analysis reporting that its overall cost to complete the same tasks is 105x lower than competitor Fable 5. This has prompted calls to shift focus from traditional per-token pricing to per-task cost. Despite impressive scores, the model's API stability, overall capability, and billing fairness remain disputed, and it is highly sensitive to test frameworks.

Confirmed

  • In Terminal-Bench 2.1, DeepSeek V4-Flash scored 82.7, nearly matching Fable 5's 80.5.
  • Artificial Analysis reports DeepSeek's overall cost for the same benchmark tasks is 105x lower than Fable 5, confirmed by @zainhas and @mustafamhus.
  • Output pricing is $0.28/M, exactly in the same range as OpenAI's GPT-5.6 Luna after an 80% price cut (plus OpenRouter promo to $0.60/M).
  • LM Studio has listed the model, supporting local or cloud use, and it shows cost-effective advantages in codebase learning and modification tasks.
  • The model offers Low, High, and Max effort levels (no Medium). @teortaxesTex notes DeepSeek needs Max to match Grok 4.5's M-level performance.

Unconfirmed

  • API stability and actual total cost: Developer @kmsdev tested 34 prompts and found unstable provider performance on OpenRouter, multiple generation failures, and a final cost of $1.29, losing to Kimi K3 in both performance and cost.
  • Misleading per-token pricing: @cephaloform and investor Chamath point out that despite low per-token prices, if the model needs more rounds to complete tasks, actual total cost may be higher.
  • Extreme sensitivity to test frameworks: @PMinervini and @teortaxesTex found DeepSeek V4 Flash is very sensitive to test frameworks. With mismatched toolchains (e.g., Pi, Claude Code), performance degrades or even burns millions of tokens; with the right framework, it shows near-GPT-level cost-effectiveness.
  • Benchmark overfitting controversy: @rickasaurus argues current LLM benchmarks are severely overfitted, noting that 5 months ago GPT-4o scored 51 on Artificial Analysis Intelligence Index, and this week DeepSeek V4-Flash also scored 51, rendering the charts meaningless.

Why it matters

  • Shift in evaluation metrics: @mustafamhus and @AravSrinivas note that a two-order-of-magnitude cost reduction is extremely rare, calling it DeepSeek's "2.0 moment." This requires decision-makers and developers to move beyond per-token pricing and focus on "per-task cost." @intellectronica calls it a moment when AI is too cheap to meter. @svenai adds that DeepSeek achieves 50 intelligence points at a cost of 3 cents, highlighting extreme value.
  • Open-source model comparison: In specific game generation tests (@mustafamhus), DeepSeek V4-Flash, despite low cost, still trails Kimi K3 (9.5/10, $0.0740) and GLM 5.2 (9/10, $0.0480) in scores, showing that cost reduction alone cannot replace overall capability competition.

Episode 9 · DeepSeek V4 Flash Sparks a Wave of Local Deployment Tests (2026-08-01, 26 posts)

DeepSeek-V4-Flash-0731 was rapidly tested across NVIDIA, AMD, Apple, and DGX Spark hardware after release. The broad takeaway is that the model can run on personal hardware, not just datacenter systems, but the gap between “it runs” and “it feels usable” is still large. In many cases, quantization choices, software backends, speculative decoding, and memory bandwidth matter as much as the underlying GPU.

Confirmed

  • On NVIDIA, community results now span single consumer cards, pro cards, and multi-GPU setups. An unoptimized RTX 3090 setup reported 4.02 tok/s, while OkNinja7526 pushed a 24GB RTX 3090 with 128GB DDR5 to 12.5 tok/s on UD-IQ3S. A Reddit test relayed by USBhost showed about 17.2 tok/s on an RTX A6000 with 256GB DDR4 running Q8. Another benchmark showed UD-Q3KXL occupying 119GB on 3x RTX 3090. Higher-end setups went much further: backslashHH measured 44-59.5 tok/s on an RTX PRO 6000 Max-Q eGPU, and a separate developer reported a 243 tok/s median on 2x RTX PRO 6000 Blackwell using DSpark speculative decoding.
  • Optimization clearly changes the picture. HockeyDadNinja said that re-quantizing only 129 router-expert tensors to IQ3XXS improved decode speed by 40% on mixed multi-GPU systems that spill into RAM. But raw hardware scale does not automatically solve usability: fragmentme loaded IQ3XXS fully into VRAM across five consumer GPUs and still saw only about 600 t/s prompt processing, concluding it was not yet ideal for everyday use.
  • AMD, Apple, and DGX Spark all posted workable results as well. Kamal965 reached 15 tok/s with 3x MI50, and Hyungsun reached 11 tok/s with a 7900 XTX plus 3x MI60. On Apple hardware, milphickelson improved an M1 Ultra 128GB setup from 5-6 tok/s to 15-16 tok/s with a community engine patch, while antirez’s DwarfStar on a 512GB M3 Ultra reached about 37 tok/s on mxfp4. For DGX Spark, one test reported 16.5 tok/s for a 3-bit run, and another confirmed that a 2-bit build can fit the full 155GB model.

Unconfirmed

  • A MiaAI result reposted by pbaylies claimed that a single DGX Spark reached 58.5 tok/s aggregate throughput at 12-way concurrency in an agentic workflow and outperformed a dual-GPU vLLM setup. Lucebox also claimed 51.1 tok/s on the full 284B model using heterogeneous consumer hardware with AMD. Both are notable, but this cluster does not include independent reproduction details.

Why it matters

These posts show that the local deployment barrier for a frontier MoE model is falling quickly across consumer PCs, Apple silicon, AMD systems, and compact AI machines like DGX Spark. They also underline that fitting the model is only step one: prompt throughput, unified or system memory bandwidth, and software-stack tuning remain decisive for real-world usability.

6 more related posts →

Episode 10 · DeepSeek V4 Flash Quantization Tests Show Major Speed Gains Without Quality Loss (2026-08-04, 2 posts)

Benchmark tests on DeepSeek V4 Flash reveal that the Q3 xxs quantized version runs several times faster than the Q8 version while maintaining lossless model quality.

Episode 11 · Single RTX 5090 Runs DeepSeek 1M Context (2026-08-04, 2 posts)

A developer successfully deployed DeepSeek-V4-Flash with a 1 million token context on a single RTX 5090. By utilizing dual CUDA graphs, they further optimized and accelerated inference speed on consumer hardware.

Episode 12 · DeepSeek-V4-Flash Tops Cost-Efficiency with Ultra-Low Running Costs (2026-08-05, 4 posts)

The new DeepSeek-V4-Flash model achieves a breakthrough in cost-efficiency, with an average benchmark cost of just 3 cents, far below competitors like Kimi K3. Ranking 21st on the Agent Arena leaderboard, it has become a favorite among developers for handling complex tasks at extremely low costs.

Episode 13 · DeepSeek V4 Flash Benchmarks Leak, Outperforming Pro (2026-08-05, 3 posts)

DeepSeek V4 Flash benchmarks leak, scoring 82.7 on Terminal-Bench 2.1 with a fifth of the parameters of the Pro version, while Unsloth AI's DSpark boosts local inference to 120 tokens/s.

Episode 14 · DeepSeek V4 Flash Tops ARC-AGI Cost-Performance, Costing a Quarter of GPT-5.6 (2026-08-08, 9 posts)

DeepSeek V4 Flash impresses on ARC-AGI, redefining the performance-cost Pareto frontier with top scores at minimal inference cost. ARC Prize official validation shows leading semi-private task scores, with cost only a quarter of GPT-5.6, breaking the convention that higher performance implies higher cost, prompting industry rethink on agent deployment costs and inference mechanisms.

Confirmed

  • ARC Prize official validation: DeepSeek V4 Flash scores 61.4% on ARC-AGI-2 semi-private tasks at $0.04/task, and 89.0% on ARC-AGI-1 at $0.02/task (cited by @petrusenkomax and @GregKamradt).
  • Cost advantage: @GregKamradt notes cost is a quarter of GPT-5.6; @togethercompute finds in DeepSWE benchmark that at equal budget, DeepSeek V4 Flash costs about a third of GPT-5.6 and solves more tasks.
  • Counterintuitive cost curve: Developer @DeArgonaut's tests on DeepSeek V4 (0731) confirm that higher scores come with significantly lower per-task inference cost. @GregKamradt cites official technical analysis explaining that models in highest inference mode are actually cheaper than high mode because more reasoning tokens lead to more derailment.
  • Solving agent cost barrier: @Teknium and @MaziyarPanahi believe this disruptive cost-performance will break the funding bottleneck for running AI agents. @MaziyarPanahi adds that in real clinical handoff workflow tests, the model runs fast and is cost-controllable at scale.

Unconfirmed

  • Specific test environment, hardware configuration, and detailed calculation of inference costs have not been fully disclosed.

Why it matters

  • DeepSeek V4 Flash achieves top reasoning performance at extremely low cost, potentially driving down AI inference costs significantly, impacting developer model selection and industry pricing strategies. Its counterintuitive 'higher performance, lower cost' phenomenon offers a new perspective on LLM inference cost curves.

Episode 15 · DeepSeek V4 Flash Passes 22 Coding Tests on Dual DGX Spark Cluster (2026-08-10, 2 posts)

Hardware evaluations reveal that running DeepSeek V4 Flash on a dual NVIDIA DGX Spark cluster makes it the only local model to pass all 22 certified coding benchmarks. With NVFP4 support, it delivers top-tier performance and cost-effectiveness, acting as a killer app for GB10 systems.