DeepSeek V4.1-Flash Review: MIT Weights, 1M Context

Introduction

DeepSeek shipped V4.1-Flash on September 10, 2026, and it is not a routine model refresh. Three things make it notable at once: a genuinely different architecture, MIT-licensed weights, and pricing that resets the floor for long-context agentic workloads.

The architecture is the part worth understanding. Previous "Flash" tiers were smaller, cheaper variants of a bigger sibling. V4.1-Flash is something else — a new structural family whose smallest member already outscores DeepSeek's previous flagship on the company's own benchmarks.

October 17 update: the API now serves version 1017

DeepSeek says it upgraded the API, website, and app to DeepSeek-V4.1-Flash-1017 on October 17. The API model name remains deepseek-flash; the API calling method and token prices are unchanged. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but now route to 1017. See the official changelog and current pricing page.

DeepSeek's changelog reports the following comparisons with the September 10 V4.1-Flash release:

These are DeepSeek-reported benchmark results, not an independent retest. The company says Terminal-Bench 4.0 used the Claude Code framework, DeepSWE used Mini-SWE-V2, and its HLE result used the pure-text subset. DeepSeek also says 1017 reduces task runtime and output length; those speed and cost claims have not been independently verified here. See the October 17 release note for the vendor's full account.

Search and Fetch are separate APIs, not included in the model token rates: the current pricing page lists $0.75 per 1,000 successful requests for each. Server-side calls through DeepSeek Harness, Claude Code, or Codex are listed as free until November 1, 2026, 00:00 Beijing time; check the pricing page before relying on that temporary offer.

The Architecture: Encoder and Decoder, Split

V4.1-Flash runs 40 transformer layers split into a 20-layer causal encoder followed by a 20-layer decoder. The decoder's global KV cache is projected from the encoder's final hidden states rather than being built up layer by layer. That single design choice produces the model's signature characteristic: asymmetric activation.

On top of the 552B backbone sits a 196B Engram conditional memory (sparsely accessed), plus a vision encoder trained from scratch. The Hugging Face checkpoint reports higher totals once those modules are counted.

The KV cache compression is the other headline. DeepSeek reports a global KV footprint of roughly 890 bytes per token — about one quarter of V4-Flash — using FP4 caching (E2M1 with one E4M3 scale per 16 channels) and a sliding-window bounded replay that avoids writing the recent-window KV state to disk. DeepSeek's summary: HBM requirements down to roughly 1/4 and SSD storage to roughly 1/8 versus the prior generation.

Pricing: The Number That Changes Infrastructure Math

Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday–Friday. Everything else — including all weekend hours — is off-peak.

Two comparisons put this in perspective:

• Against frontier models: Claude Opus 5 is listed at $5.00 input / $25.00 output, GPT-5.6 Sol at $4.00 / $20.00, Kimi K3 at $3.00 / $15.00. At off-peak rates, V4.1-Flash output costs roughly 40 times less than Opus 5.
• Cache economics: take a 500,000-token reusable prefix (system instructions, tool definitions, a repository context) read across 100 requests — 50 million cached input tokens. On V4.1-Flash off-peak that is about $0.15. The same cache reads would be roughly $15 on Kimi K3, $20 on GPT-5.6 Sol, and $25 on Claude Opus 5, before cache-write charges.

If your agent re-reads the same context on every turn — which most coding agents do — the cache-hit rate is the number that matters most, and $0.003 per million is close to free.

Benchmarks

DeepSeek published an extensive set. Treat these as vendor numbers: independent testers have not fully reproduced them, which is normal at launch and worth stating plainly.

Artificial Analysis put the model's Intelligence Index at 40, placing it above DeepSeek V4 Pro 0813 — the previous flagship — despite the "Flash" name.

Independent Verification Arrives

The third-party results below were published in September and concern the earlier V4.1-Flash release, not the October 17 version 1017. They provide some independent evidence about the September model; this review has not found an independent evaluation specifically of 1017. DeepSeek's 1017 benchmark table above remains vendor-reported.

Agent Arena: third among open models at $0.07 a task

Agent Arena ranks models on real agentic tasks — tool reliability, task completion, steerability — as net improvement against median cost per task, across roughly 1.69 million sessions and 43 models. The board treats the trade-off as the point rather than sorting by score.

DeepSeek-V4.1-Flash placed third among open models and twelfth overall, with a +4.87% net improvement at $0.07 median cost per task — and it landed on the Pareto frontier, meaning nothing else on the board buys that much improvement for that little. It retains 98% of Hy4 preview's improvement at 73% lower cost, and stays within 1.52 points of Kimi K3 while costing 91% less. Against the frontier leaders it keeps 35–54% of their net improvement at 97–99% lower cost per task — the top models cost 37 to 76 times more. It also ranked fourth on the "Confirmed Success" signal at +13.75%.

Fireworks AI: DeepSWE parity at a fifteenth of the cost

Fireworks AI published its own evaluation on September 14–15: on DeepSWE, V4.1-Flash (max) scored 74.34% pass@1 — level with GPT-6 Astra — at roughly $0.43 per task against about $6.45, a factor of fifteen.

Artificial Analysis also places the model at 40 on its Intelligence Index (v4.3), in the leading open-weight group alongside GLM-5.3 and Qwen 3.8 2.4T, with output speed of 214.4 tokens per second.

What this does and does not settle

Independent data now supports the core claim: near-frontier capability at a fraction of frontier cost, and genuine Pareto-optimal positioning for agentic workloads. What it does not settle is robustness. Agent Arena and DeepSWE measure task performance, not the sparse-selection and cache-resumption edge cases DeepSeek itself flagged as untested. If your workload depends on precise retrieval from a 1M-token context, the honest position remains: measure it on your own tasks.

Open Weights, Real Hardware

The weights are on Hugging Face under MIT, ungated. That matters more than the benchmark table for a specific class of buyer: MIT means commercial use, redistribution, and fine-tuning without revenue thresholds or enterprise agreements, and it lets teams keep sensitive workloads entirely on their own hardware.

The catch is the hardware floor:

• Checkpoint size: roughly 510GB
• Minimum practical self-hosting: 8x H100 (FP8 build)
• FP4 quantization: ~385GiB; a single B200 covers the native FP4+FP8 build
• Serving: SGLang ships a V4.1 preview with an option to keep Engram memory tables in host RAM; vLLM support is in progress

For most teams the API is the realistic path. Self-hosting makes sense when data residency or volume economics justify the GPU bill.

What DeepSeek Says Is Untested

DeepSeek acknowledged in its own documentation that the new architecture introduces robustness limits it has not fully characterized:

• sparse-selection errors
• approximate state reconstruction

Both could degrade capability in edge cases — specifically sparse retrieval over very long contexts and cache-resumption boundaries. Additional stress testing is planned. If your workload depends on retrieving precisely from a 1M-token context or resuming long sessions, schedule your own regression tests rather than assuming parity.

Migration: Handle the Model Name Carefully

The current API model name is deepseek-flash, and the live pricing page identifies its served version as V4.1-Flash-1017. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp remain accepted but also route to 1017. DeepSeek lists deepseek-v4-pro separately as V4-Pro-0813 with its own rates. Because the Flash endpoint is updated in place, rerun prompt and regression tests after a model update; confirm current versions and rates on the official pricing page.

Who Should Use It

Summary

V4.1-Flash is the most consequential thing DeepSeek has shipped this year: a new architecture that makes long-context caching cheap, MIT weights that make it deployable in places previous releases were not, and pricing that forces every other provider to justify its per-token rate.

The independent results cited above apply to the September release, not version 1017. The newer benchmark figures are currently DeepSeek-reported and need independent reproduction. The honest caveats also remain: DeepSeek flags untested robustness limits in areas that matter for agentic work. For teams running re-read-heavy agents, the cache-hit economics may justify a controlled bake-off, but measure cost and task success on your own workload before switching.

For the previous generation, see our DeepSeek V4 Flash review; for the competitive picture, compare against Kimi K3 and our best Chinese AI models roundup.

Related Articles

• DeepSeek V4 Flash
• DeepSeek V4 API Price Increase
• Kimi K3 Review
• Best Chinese AI Models 2026
• Agent Arena data
• Qwen3.8-Omni-Flash Review
• MiMo-V2.6 Review
• Cursor: The AI-Native Code Editor — Full Review for 2026
• Windsurf: AI Code Editor
• Best AI Coding Tools in 2026: Complete Guide for Developers
• Claude Sonnet 5.5 and Haiku 5.5