DeepSeek V4.1 Flash: Benchmarks, Capabilities, and Architecture

DeepSeek replaced its flagship with a smaller model. On September 10, 2026, it released DeepSeek V4.1 Flash as deepseek-flash and will route deepseek-v4-pro traffic to it from 04:00 UTC on September 14.

The new model costs less, keeps a 1M context, and leads V4 Pro on most agent benchmarks DeepSeek reports. Gains come from a Causal Encoder-Decoder backbone and deep KV cache compression. Treat it as a new base model and re-test V4 workflows.

What shipped

• Launch: September 10, 2026. API name deepseek-flash. MIT weights at deepseek-ai/DeepSeek-V4.1-Flash.
• Retired: deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired. Both aliases now serve V4.1 Flash at Flash prices.
• Routing: From 04:00 UTC September 14, all deepseek-v4-pro calls serve V4.1 Flash at Flash rates until V4.1 Pro ships.
• Family: DeepSeek calls V4.1 Flash the smallest model in its new architecture family. V4.1 Pro has no published card or date.

The .1 label reads like a patch. The weights are a fresh 552B pretrain.

Architecture: what changed from V4

V4 Flash was a 284B MoE with 13B active per token. V4.1 Flash keeps 1M context and 384K max output and changes how context is processed.

Detail	V4 Flash (0731)	V4.1 Flash
Base	284B MoE, 13B active	552B MoE, 8B prefill / 16B decode
Extra memory	-	196B Engram conditional memory, sparse
Layers	Standard Transformer	40-layer CED: 20 encoder + 20 decoder
Experts	-	1 shared + 384 routed, 6 active per token
Vision	Separate Vision-Exp	Native DeepSeek-ViT
KV cache	Baseline	890 bytes per token, FP4 E2M1
Persistent cache	Baseline	~1/8 of V4 Flash via SWA Bounded Replay

Four mechanisms:

• CED. Decoder global KV is projected from encoder final states. Prefill uses 8B. Decode uses 16B. This favors input-heavy traces.
• CSA2. Layers use Full, Reindex, or Reuse modes. KV and indexer state is shared. Hierarchical indexer limits deeper layers to a candidate pool.
• FP4 cache. E2M1 FP4 with one E4M3 scale per 16 channels. DeepSeek reports ~1/4 HBM of V4 Flash and ~1/437 of V1.
• SWA Bounded Replay. Missing sliding-window KV is rebuilt from the last window. Persistent storage drops to ~1/8.

Other pieces include Single-Pass mHC, Engram, and DSpark decoding. Pretrained on 45T tokens, sparse attention at 64K extended to 1M at 34T. Recipe is SFT, RL, and on-policy distillation.

Pricing and API

Prices per 1M tokens. Off-peak is half of peak.

Meter (per 1M)	Off-peak	Peak
Input cache hit	$0.003	$0.006
Input cache miss	$0.15	$0.30
Output	$0.60	$1.20
Context	1M	1M
Max output	384K	384K
Concurrency	2,500	2,500

Peak: 01:00-04:00 and 06:00-10:00 UTC Mon-Fri. V4 Pro lists $0.022/$0.044 hit, $0.66/$1.32 miss, $1.98/$3.96 output. V4 Flash 0731 listed $0.007/$0.014, $0.22/$0.44, $0.66/$1.32.

A 1M miss plus 1M output job is $0.75 off-peak or $1.50 peak by rate card. Measured cost also varies with reasoning length, retries, and hit rate.

API:

• Bases https://api.deepseek.com and https://api.deepseek.com/anthropic
• Text plus image input, text output. JSON and function calling supported
• Thinking on by default. Continuous reasoning_effort 1 to 100. Published scores use 100
• Responses API native. Partners WorkBuddy/CodeBuddy and OpenCode listed day one

Benchmarks

All scores are vendor-reported at max reasoning effort. Use them to plan tests.

Benchmark	V4 Flash	V4 Pro	V4.1 Flash	Reference
Terminal-Bench 2.1	82.7	87.9	90.6	89.1 Opus 5.0, 88.8 GPT-5.6 Sol
DeepSWE v1.1	54.4	62.7	74.2	74.0 Opus 5.0, 73.0 Sol
CyberGym	76.7	83.3	88.1	84.5 Sol
AutomationBench	37.7	43.2	54.8	50.3 Opus 5.0
HLE with tools	51.5	60.0	63.9	63.6 Opus 5.0
GPQA Diamond	89.9	92.4	90.9	94.1 Sol, 93.4 Opus 5.0
HLE no tools	37.8	42.7	36.8 (39.1 text)	56.3 Opus 5.0, 44.5 Sol
Terminal-Bench 3.0	7.6	11.8	30.0	43.3 Opus 5.0
Terminal-Bench 4.0	7.0	12.4	31.2	51.8 Opus 5.0
Codeforces	3289	3348	3471	-
SEC-Bench Pro	30.9	56.4	62.8	74.3 Sol

Takeaways:

• Agent coding leads. DeepSWE, Terminal-Bench 2.1, CyberGym, AutomationBench all rise. DeepSWE 74.2 matches Opus 5.0 in that setup.
• Reasoning mixed. GPQA sits between V4 Flash and V4 Pro. HLE without tools stays below frontier.
• Scaffold matters. DeepSWE 74.2 on mini-SWE is 72.6 on DeepSeek Harness Minimal and 65.6 on Codex. Terminal-Bench 2.1 is 90.6 Minimal, 90.3 mini-SWE, 84.1 Codex.

Base notes: MMLU-Pro 74.1 vs 73.5 V4 Pro vs 68.3 V4 Flash. HumanEval 79.4 vs 76.8 vs 69.5. Multimodal: 56.5 MMMU-Pro, 77.9 CVBench, 95.6 DocVQA, 86.0 RefCOCO. Artificial Analysis reports 190 tokens per second and Index 40 at max effort.

Capabilities

• Long context. 1M at 890 bytes per token is ~0.9 GB global KV for a full window. Helps repeated prompts and large repo traces.
• Native vision. One endpoint for text and images. No separate Vision-Exp model needed.
• Effort dial. 1 to 100 controls reasoning cost. Track output tokens with latency.
• Concurrency. 2,500 matches V4 Flash and tops V4 Pro at 500. Routed Pro traffic gets the higher limit.

Limits:

• Trails Opus 5.0 and Sol on Terminal-Bench 3.0 and 4.0.
• Bills by clock. Weekday peak windows cost double.
• New backbone changes behavior. Tuned prompts and tool schemas may need updates.

What to test before September 14

If deepseek-v4-pro is hard-coded, behavior changes without a deploy.

• Run identical workflows on deepseek-flash and deepseek-v4-pro now. Log versions.
• Keep context, tools, approvals, and success criteria fixed. Change only the model.
• Measure completion vs product state, retries and failures, hit vs miss ratio and output tokens, wall-clock time, reviewer edits.
• Include failure cases: incomplete data, denied access, ambiguous match, duplicate request, already completed action.
• Compare consolidated vision path against split model path for image work.

Use rate card for estimates and billing logs for totals. Savings quotes assume equal token counts. Totals change when token counts change.

Where Ginger Labs fits

At Ginger Labs, we provide an embedded agent that lives inside a SaaS product and works with its schemas, records, and permissions. It handles retrieval, tool execution, evaluations, and observability.

The product keeps authority. The model proposes a call and the product authorizes it, validates payload, and applies approval policy. V4.1 Flash can sit behind that boundary as V4 Flash or V4 Pro.

Teams keep their data model, domain rules, tenant scope, and definition of done. We supply the agent surface that makes cross-version checks measurable. Useful for a short pilot that compares September 14 behavior before it becomes default.

Sources

• DeepSeek-V4.1-Flash release and Models and pricing, DeepSeek API Docs. Accessed September 15, 2026.
• DeepSeek-V4.1-Flash model card, DeepSeek-AI. Accessed September 15, 2026.
• DeepSeek V4.1 Flash: Architecture, API Pricing, Benchmarks, and What Changed, Atoms. Accessed September 15, 2026.
• Artificial Analysis - DeepSeek V4.1 Flash (max), Artificial Analysis. Accessed September 15, 2026.
• Workspace comparison dashboard_1785080687994/set-2/deepseek-v4-flash-vs-gpt-56-luna-best-models-for-low-cost-agentic-workflows.md.

Material unknowns: No large independent benchmark yet. Production cache hit rates are unpublished. V4.1 Pro scale, price, and date are still unannounced.