This is not another incremental model drop.

Five months ago the best closed frontier model on the planet scored 51 on the standardized intelligence index. This week DeepSeek dropped an open weight model that scores 50. You can run it on hardware you already own. You can host it for less than a tenth the cost of GPT-4o.

That is the line we have all been waiting to cross. Nobody saw it coming this fast.

What actually got released

DeepSeek V4 Flash 0731 is a 720B total parameter Mixture of Experts model with 14B activated parameters per token. That is the critical number. Total parameter count is marketing. Activated parameters are what you actually pay compute for.

Unsloth published lossless GGUF quantizations within 12 hours of release. As of writing there are verified Q2_K, Q3_K_XL, Q4_K_M, Q8_0 and FP8 builds available. All work natively in llama.cpp, vLLM, SGLang and DS4 DwarfStar.

Context window is 512k native. No fine tuning required, no rope scaling hacks.

Benchmark standing

This is the part that matters. Nobody releases benchmark tables that don't flatter their model, so we are only using independent third party runs posted in the 72 hours after release.

Model	Intelligence Index	API Output Price / 1M tokens	Local minimum VRAM
DeepSeek V4 Flash 0731	50	$0.18	38 GB
GPT-5.6 Luna	52	$1.20	N/A
Laguna S 2.1	47	$0.72	48 GB
Kimi K3	61	$2.10	29 GB RAM
GPT-4o	41	$2.50	N/A

This is the first time an open model has ever been within 2 points of the peak frontier score. And it is cheaper than every single model below it.

The price war just became real

OpenAI cut Luna pricing 80% twelve hours after DeepSeek launched. OpenRouter immediately put Luna on 50% sale to bring it down to $0.60/M. That is still 3.3x more expensive than DeepSeek Flash.

Independent canvas testing shows DeepSeek beating Luna on physical simulation tasks. It correctly animates Rubik's cube rotations, gravity affected water jets and object buoyancy. Luna failed all three of those tests at the same prompt. Luna still holds an edge on Solidity coding. That is the only remaining gap.

Nobody has run a full SWE-bench run yet. Early informal tests put it roughly on par with Laguna S 2.1.

Observed local inference performance

This is what every engineer actually cares about. All numbers below are reported user runs, not vendor benchmarks.

All runs are at 128k context. Prefill speed sits consistently around 180-220 tokens per second on 4 card setups. That is fast enough for real time agent loops.

One user reported running the Q3_K_XL quant on a single RTX 5090 24GB with 4GB offloaded to system RAM, getting 7.8 t/s. That works. It is not nice, but it works.

Deployment options

You have four working production grade options right now:

1. llama.cpp Most widely tested. Works everywhere. Use the Unsloth GGUF builds. Add --flash-attn on --no-kvu for best performance.
2. DS4 DwarfStar Purpose built engine for this architecture. Consistently delivers ~2x the speed of llama.cpp on identical hardware. Has native persistent KV cache on SSD. Antirez released official quants 36 hours after launch.
3. vLLM 0.25.0+ For API serving. Auto detects quantization. Expect ~3500 concurrent users per 4x 5060 Ti node.
4. SGLang Best option for agent workloads and function calling. 15-20% faster than vLLM for batch generation.

Quantization tradeoffs

This model is unusually quantization tolerant. That is not an accident. DeepSeek trained with quantization aware calibration all the way down to Q3.

Quant level	File size	Quality loss	Recommended use
FP8	76 GB	0%	Production serving, benchmarking
Q8_0	76 GB	<0.1%	Default for all use
Q4_K_M	39 GB	~1.2%	Local desktop use
Q3_K_XL	31 GB	~3.8%	Minimum acceptable for general use
Q2_K	22 GB	~11%	Only for testing, do not use for work

Nobody has found any measurable difference between FP8 and Q8_0 on any benchmark. That is almost unheard of for MoE models.

What this changes

Until this week there was an unchallenged assumption: you paid a very large penalty to run open models. You gave up 15-20% capability, you paid more per token, you got worse speed.

That assumption is dead.

Right now today you can run DeepSeek V4 Flash on a $7000 server node, serve 3500 concurrent users, and undercut OpenAI's pricing by 7x while delivering near identical capability. You can run it on your laptop. You can run it air gapped. You never have to send user data to a third party ever again.

This is not the end of closed models. But it is the end of closed models having an overwhelming default advantage.

The fine print

This is not a perfect model. There are known issues:

• It will hallucinate citations at roughly the same rate as Luna
• Long context retrieval falls off sharply after 380k tokens
• MMLU score is 3 points lower than originally reported
• It is very bad at counting
• The default system prompt is aggressively censored. Ablation vectors already exist.

None of these are deal breakers. All are being fixed already.

What comes next

We just watched the inflection point. This release did not just move the line. It erased it.

Over the next 30 days every single LLM provider will cut their pricing. Every inference engine will add optimizations for this architecture. Someone will get it running on a 16GB card. Someone will make a 2 tok/s CPU only build.

WASTE engine already demonstrated you can run a 2.7T parameter model on consumer hardware. That was last week. This week we got one that is actually useful.

We are no longer waiting for the future where good models are open and run locally. That future arrived this Tuesday.

References

All source data and raw benchmarks are linked in the front matter. Raw user run logs are available in the linked LocalLLaMA threads. Unsloth GGUF builds are updated daily with improved quantization profiles.