Last night’s MiMo-V2.6-Pro note was an open-weights price-floor sentence at Artificial Analysis 46. This morning is the other half of the same routing problem: what the model does to HBM and SSD when the agent loop is input-heavy. On September 10, 2026, DeepSeek shipped DeepSeek-V4.1-Flash — a multimodal MoE under an MIT license — and started retiring the previous Flash / Vision-Exp endpoints onto it. By mid-September, legacy deepseek-v4-pro traffic was also being steered to Flash rates until V4.1-Pro lands.

Primary sources: DeepSeek’s Sep 10 release note, Hugging Face model card, API Models & Pricing, Artificial Analysis release page, and Powerdrill’s Sep 14 pricing cross-check. Vendor benches below are DeepSeek’s unless marked independent (Artificial Analysis). Distinct from the pending DeepSeek × Ascend 950DT cluster package.

Spec sheet (named numbers)

DeepSeek-V4.1-Flash
HF checkpoint deepseek-ai/DeepSeek-V4.1-Flash
API id deepseek-flash (legacy deepseek-v4-flash / deepseek-v4-flash-vision-exp route here)
Architecture Causal Encoder–Decoder (CED) MoE
Backbone / active 552B / 8B prefill · 16B decode
Layers 40 (20-layer causal encoder + 20-layer decoder)
MoE 1 shared + 384 routed experts; 6 routed active / token
Context / max out 1,048,576 / 384K (API table)
Global KV cache 890 bytes/token (~1/4 of V4-Flash); release note: 1/4 HBM, 1/8 SSD
Attention tricks CSA2 (Full / Reindex / Reuse); FP4 main KV (E2M1, E4M3 scale / 16 ch); SWA Bounded Replay
Other named blocks Single-Pass mHC; Engram conditional memory (196B sparse); DSpark speculative decode
Vision DeepSeek-ViT from scratch + 2-layer MLP projector (native multimodal)
Pretrain 45T tokens; sparse attn at 64K, context to 1M at 34T
Reasoning dial Integer 1–100 effort (API / card)
License MIT
Off-peak $/1MTok Cache hit $0.003 · miss $0.15 · out $0.60
Peak $/1MTok Cache hit $0.006 · miss $0.30 · out $1.20
Peak window Mon–Fri 01:00–04:00 and 06:00–10:00 UTC (excl. CN holidays); else off-peak
Concurrency 2500 (API docs)
AA Intelligence Index 39 (max effort) · 25 (non-reasoning)
AA cost / index task ~$0.27 (max)
AA output speed ~222 tok/s (max) · ~242 tok/s (non-reasoning)

Same-stack foil from last night: MiMo-V2.6-Pro sits at AA 46 and $0.435 / $0.87 list with a deeper open-weights ceiling claim. Flash is not trying to win that ceiling row. It is trying to win the cache and activate row on long agent prompts.

What actually changed for people who run agents

The durable product change is not a new FLOPS sticker. It is asymmetric activation + a smaller persistent KV.

CED projects the decoder’s global KV from the encoder’s final hidden states instead of building a full per-layer KV stack in the decoder. Prefill therefore activates 8B; decode activates 16B. That is the input-heavy agent sentence: attach a long repo or tool trace once, then ask several questions without lighting up a Pro-class activate path on every prefill token.

CSA2 and FP4 main-KV caching push the published global cache to 890 bytes per token. DeepSeek’s release note translates that into capacity: about one-quarter the HBM and one-eighth the SSD versus V4-Flash. If you already read Lightbits Inferra, keep the layers straight: Inferra is system orchestration across HBM/DRAM/NVMe; V4.1-Flash is model compression of what must sit there.

Operationally, DeepSeek is also cleaning the SKU sheet. V4-Flash and V4-Flash-Vision-Exp are retired onto deepseek-flash. Starting 04:00 UTC on September 14, 2026, deepseek-v4-pro requests were routed to V4.1-Flash at Flash rates until V4.1-Pro ships. That is a fleet routing ticket, not a marketing footnote.

Vendor benches vs the independent index

DeepSeek’s instruct table (max effort reasoning_effort=100; selected rows; higher better):

Benchmark V4.1-Flash V4-Pro V4-Flash Opus-5.0 GPT-5.6 Sol
Terminal-Bench 2.1 90.6 87.9 82.7 89.1 88.8
Terminal-Bench 4.0 31.2 12.4 7.0 51.8 39.9
DeepSWE v1.1 74.2 62.7 54.4 74.0 73.0
AutomationBench 54.8 43.2 37.7 50.3 45.8
Agent's Last Exam 31.8 25.7 25.2 28.6 26.7
GPQA Diamond 90.9 92.4 89.9 93.4 94.1
HLE (text†) 39.1† 42.7† 37.8† 56.3 44.5
CyberGym 88.1 83.3 76.7 — 84.5
ExploitGym 15.3 5.4 1.8 22.1 33.7

† Text-only HLE subset on DeepSeek’s card. Pattern for a routing ticket: Flash beats its own prior Pro on several highlighted agent rows, and still trails Opus 5 / Sol on harder boards — especially Terminal-Bench 4.0. Treat vendor rows as DeepSeek-harness until your suite says otherwise.

Independent check: Artificial Analysis puts max-effort Flash at 39 on Intelligence Index v4.3.x (~$0.27/task, ~222 tok/s). Below last night’s MiMo 46 and well below Claude Opus 5.5 / GPT-6 Astra. The honest sell is cache economics and activate asymmetry, not a flagship rematch.

Where it sits next to MiMo, Inferra, and the closed ladder

Purchase Product What it buys
Open-weights price floor MiMo-V2.6-Pro AA 46; $0.435/$0.87; 1.02T/42B
Open volume / self-host on-ramp MiMo-V2.6-Flash $0.14/$0.28; 309B/15B
Model-side KV / activate math DeepSeek-V4.1-Flash 890 B/tok; 8B/16B; $0.15/$0.60 off-peak
System KV orchestration Lightbits Inferra HBM/DRAM/NVMe prefetch engine
Default coding agents (OpenAI) GPT-6 Sol $2/$10; cache-steerable effort
Volume / extract (OpenAI) GPT-6 Luna $0.10/$0.50
Flagship / agentic coding (Anthropic) Claude Opus 5.5 Fable-class work at $4/$20
Flagship reasoning (OpenAI) GPT-6 Astra Highest OpenAI ceilings; $10/$50

Flash does not invent a fabric. It argues that long-context agent loops should not pay V4-Pro prices for a worse activate/cache shape — and DeepSeek is forcing that migration in the API.

Caveats I would write into any Flash ticket

  1. Reproduce Terminal-Bench 4.0 / DeepSWE / your repo harness before you delete a Pro shadow route — DeepSeek’s table is not your fleet.
  2. Do not paste “beats Opus 5.” Flash leads some of DeepSeek’s own agent rows and still loses hard boards; Opus 5.5 is also newer than the Opus-5.0 column.
  3. Peak vs off-peak is real money. Weekend and CN-holiday hours are off-peak; weekday UTC mornings are not. Batch flexible work off-peak.
  4. Self-host ≠ free. Weights are MIT; the card points at a conversion/inference path and notes there is no Jinja chat template — use DeepSeek’s encoding reference or deepseek-recipe.
  5. AA 39 ≠ MiMo 46. Route ceiling work to MiMo Pro, Astra, or Opus; put Flash where cache-hit agent loops dominate the bill.
  6. Not a rematch of Inferra / Leo X / Lessengers. Those are system memory and optics. This is the model serving layer.

How I would evaluate if I ran agent fleets

  1. Cost per completed task on deepseek-flash at peak and off-peak vs MiMo Flash/Pro, Sol, and Luna — include cache-hit share.
  2. HBM / SSD footprint on your longest production contexts against the 890 bytes/token claim; do not trust the slide alone.
  3. Terminal-Bench 4.0-class unattended jobs on a shadow route before you collapse Pro traffic.
  4. Effort dial 1–100 as a budget control, not a quality slogan — measure which integers actually move your suite.
  5. Keep a closed fallback (Opus 5.5 / Astra) for the boards where DeepSeek’s own card shows the gap.

This is a KV-and-activate efficiency release with MIT weights and a forced SKU cleanup, not a new closed flagship. Put Flash on long, cacheable agent loops where $0.15 / $0.60 off-peak and a smaller persistent cache beat V4-Pro arithmetic. Keep MiMo Pro on the higher open ceiling, and keep Astra / Opus 5.5 on hard unattended work until your harness says otherwise. Measure, then move traffic.

Create-only. Specs from DeepSeek’s Sep 10 note, the HF V4.1-Flash card, API pricing, Artificial Analysis’s release page, and Powerdrill’s Sep 14 cross-check. Update when V4.1-Pro ships, when a named harness reproduces Terminal-Bench 4.0 against Opus 5.5 / Astra / MiMo Pro, or when DeepSeek revises peak hours or the byte-per-token claim.

Sources: DeepSeek Sep 10 note · Hugging Face card · API pricing · Artificial Analysis · Powerdrill.

Anpoo Sivanadi, Staff Software Engineer