DeepSeek V4 Flash Went Official: Checking the 'Flash Beats Pro' Claim Against the Model Card and config.json

The Claim

DeepSeek-V4-Flash-0731 is now the official V4 Flash release, superseding the preview. The model card states that on all nine agentic benchmarks it lists, Flash-0731 beats V4-Pro (Preview) — "despite its far smaller activated parameter count." That is a strong claim, so I checked it two ways: the table arithmetic, and the architecture in config.json.

9/9, Verified Row by Row

Flash-0731 wins 9/9 against V4-Pro (Preview), wins all 8 rows where GLM-5.2 has a score, and loses all 9 to Opus-4.8 — the closest margins being Agents' Last Exam (25.2 vs 25.7) and AutomationBench (25.1 vs 27.2). "Broadly competitive" is precise: close, not ahead. Caveats from the card itself: all numbers are DeepSeek's own evaluations (DeepSeek Harness minimal mode, max reasoning effort), and the two DSBench sets are marked † as internal DeepSeek test sets. None of this is independently reproduced.

What config.json Says

From the public config: model_type deepseek_v4, 43 layers, hidden_size 4096, 64 attention heads with a single KV head, vocab 129,280, and max_position_embeddings 1,048,576 — 2^20, so the "million-token context" in the paper title is literal. The MoE is 256 routed experts plus one shared expert per layer, with 6 routed experts active per token and an expert intermediate size of just 2048. Dense weights are FP8 (E4M3, dynamic activation); expert_dtype is fp4 — experts ship at 4 bits by design.

Reconciling 284B and 304B

One expert = up (4096×2048) + gate (4096×2048) + down (2048×4096) = 25,165,824 parameters. With 43 × 256 = 11,008 routed experts, that is ≈ 277.0B in routed experts alone. Adding shared experts, dense attention, and the (unshared) embeddings and head lands at ≈ 284B — matching unsloth's 284B figure. The card's "304B in safetensors" includes the in-checkpoint DSpark speculative-decoding module (11.3 GB at BF16) and quantization scale tensors. Different accounting, not a contradiction.

Activated per token: 7 experts per layer (6 routed + 1 shared) × 43 layers ≈ 7.6B, plus dense attention and embeddings ≈ 10B — roughly 3.5% of the model. The "far smaller activated parameter count" statement is consistent with the architecture.

The Engineering News: Speculation Is In the Checkpoint

DSpark ships inside the main checkpoint: 7 speculative tokens, greedy draft sampling, no separate draft model to deploy. vLLM: --speculative-config '{"method":"dspark","num_speculative_tokens":7}'; SGLang: --speculative-algorithm DSPARK; llama.cpp users can add it as a module (Q8_0, 10.9 GB). The release also adds three-level reasoning_effort (low/high/max, up to 384K output tokens at high/max) and replaces the Jinja chat template with an encoding/ folder of Python helpers — teams using transformers directly will want to migrate.

Local Deployment, Sized Honestly

unsloth's GGUF: UD-Q4_K_XL 155 GB, UD-Q8_K_XL 162 GB — the "lossless" Q8 is only 7 GB bigger than Q4 per unsloth's note. Ollama: ollama run hf.co/unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q4_K_XL. llama.cpp: llama serve -hf unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q4_K_XL. A 155 GB Q4 file means big-RAM workstations or multi-GPU servers; for most product teams, the hosted flash tier will be the economical path.

Takeaways

1. The flash tier's economics are now backed by the vendor's own benchmarks — but read them as the vendor's.
2. FP8 dense + FP4 experts + in-checkpoint speculation is a deployment recipe worth copying for latency- and cost-sensitive agent workloads.
3. A 1M-token context with a 384K output budget changes what fits in a single agent call.

Model card: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731 (MIT) · config.json: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731/resolve/main/config.json · GGUF: https://huggingface.co/unsloth/DeepSeek-V4-Flash-0731-GGUF · Paper: https://arxiv.org/abs/2606.19348

Not tested / not run — no inference or benchmark reproduction was performed. Benchmark numbers are the vendor's; the arithmetic and config reading are mine.