DeepSeek Ships V4-Flash Update With Big Agent Benchmark Gains, Rivaling GPT-5.6 Terra

Terminal Bench jumped from 56.9 to 82.7 and Toolathlon from 51.8 to 70.3 versus the preview version. Compared to GPT-5.6 Terra: Flash beats it on Terminal Bench (82.7 vs 78.4) and Toolathlon (70.3 vs 53.1), but loses badly on DeepSWE (54.4 vs 69.6) and Agents Last Exam (25.2 vs 50.4). No clear winner, wildly differing scores, genuinely interesting.

They used their own harness for benchmarking so I'm skeptical this is an apples to apples comparison against Terra.

This is more exciting than k3 to me. V4 models are dirt cheap to serve, and DeepSeek has integrations with opencode and other providers so they're probably gathering tons of real developer usage data to post-train further. Flash can run at home for under $10k, unlike GLM or k3 sized models, and that capability doesn't disappear on a provider's whim.

It's DeepSeek-V4-Flash-284B-A13B, so it should barely fit on a single B300 or an M5 Max.

It runs really well on 2 DGX Sparks at 60 tokens per second.

Wait, weights for this new version aren't even out yet, that's still the old V4 preview people are running.

DeepSeek is always weirdly understated in their changelogs. This is actually a huge capability jump, not a minor tweak.

GPT-5.6 Luna having vision is a real edge for agentic tasks that Flash lacks, and DeepSeek has been quietly raising prices after their discount period ended.

They haven't actually raised prices, the peak-hour pricing plan for V4 Pro GA never happened. And Luna's cache read is $0.02 per million tokens versus V4 Flash's $0.0028, so 'a bit cheaper' undersells it hugely.

Per a leaked internal call, Liang Wenfeng doesn't seem worried about the Cost per Intelligence Index gap, and DeepSeek is apparently working on vision for V4 already.

This makes me optimistic about local inference. The CyberGym score especially stands out.

A ~200B model (actually only 160GiB) competing with GLM-5.2 and getting close to Opus 4.8 is genuinely impressive, and with DeepSeek's caching pricing this thing is going to see massive usage.

If these benchmarks reflect real use, it's insane that a 300B model beats the old 1.8T param V4 Pro preview and possibly GPT-5.6 Luna too, while staying cheaper. Though on DeepSWE specifically Luna still leads, 67% vs 54.4%.

I use DeepSeek Flash constantly for fun coding projects. My app has 400 users and hasn't even hit $50 a month in usage.

Finally a model with usable intelligence at a reasonable price. Can't imagine what the full Pro GA release will look like given the preview already has 1.6T parameters.

Why not call it V4.1 instead of reusing the V4-Flash name? OpenRouter's slug was already deepseek/deepseek-v4-flash for the old version, so this is going to cause confusion between versions.

DeepSeek did the same thing with v3 and R1, multiple versions under one name, classic DeepSeek naming chaos.

The flash variant matching Sonnet 5 on DeepSWE at 54% is a big deal if it holds up.

Curious how it performs on English prose rather than code, we're looking for cost effective ways to keep story wikis updated. Feels like these models are overly optimized for coding tasks specifically.