Tenzy Reports 2026 Independent reading DeepSeek V4.1 Flash
The model got bigger. The active work got smaller.
552B backbone. 8B active during prefill. 16B during decode.
DeepSeek rebuilt Flash around a Causal Encoder-Decoder architecture designed to reduce the cost of long, input-heavy agent workloads.
01 / Architecture
Twenty layers in. Twenty layers out. One global state between them.
V4.1-Flash splits its 40 Transformer layers into two halves. A 20-layer causal encoder reads the input. A 20-layer decoder writes the output. The decoder does not keep its own copy of the conversation. Its global KV state, the stored attention memory it answers from, is projected from the encoder’s final state.
-
Input
Prompt and image tokens enter the encoder together.
Text + image, one sequence -
Stage 01 · Causal encoder
20 layers read the whole input once, in a single pass.
8B active / prefill -
Projected global state
One compressed global KV state is projected from the encoder’s final hidden states. It is not rebuilt for each decoder layer.
890 bytes / token -
Stage 02 · Decoder
20 layers generate output tokens against that single projected state.
16B active / decode -
Output
Text, generated autoregressively.
One global cache, held once
The asymmetry is the point. Prefill is the pass that reads your prompt. It runs once, and it activates about 8B parameters per token. Decode is the pass that writes the answer, one token at a time, and it activates about 16B.
Neither number is the size of the model. The full model holds 552B parameters across its experts. Only six of the 384 experts in a layer are active for any given token, which is why the working set is so much smaller than the model.
DeepSeek reports the resulting global KV cache at 890 bytes per token. That is the architectural consequence the rest of this page keeps returning to.
02 / KV cache
The cache is the story.
890 bytes / token
DeepSeek reports the global KV cache per token at 890 bytes. That is roughly one quarter of V4 Flash, and roughly one four-hundred-and-thirty-seventh of V1.
Log scale: equal distances are equal ratios
Two mechanisms do the work. CSA2 sparse attention lets layers share key and value projections and reuse each layer’s Top-K indices. FP4 main KV caching stores the cache in a narrower number format. The smaller cache is a consequence of the architecture, not of a smaller model.
Bar length is logarithmic, so equal distances are equal ratios, not equal byte counts. The 890-byte figure is DeepSeek’s. The two predecessor values are editorial arithmetic. This page derived them by applying DeepSeek’s published ratios, about 4× and 437×, to its reported 890 bytes. The byte totals themselves are not given in the text of the model card, so they are marked derived wherever they appear. They are not measurements.
The absolute scale is what makes the ratios legible. The exact predecessor totals are the weakest numbers on this page.
A separate claim, deliberately not merged with the one above. DeepSeek also reports SWA Bounded Replay. Sliding-window KV states are rebuilt by replaying only the most recent window, instead of being persisted. DeepSeek says this reduces the persistent KV footprint to roughly 1/8 of V4 Flash. That is a storage-side figure about what gets written out. It is not a second measurement of the 890-byte global cache.
03 / Three readings
Three numbers that argue with each other.
Each one is arithmetic on DeepSeek’s published figures. This is an editorial reading, not a new measurement.
-
Terminal-Bench 2.1
+1.5
V4.1 Flash vs Claude Opus 5
V4.1 Flash leads on this published benchmark. 90.6 against 89.1.
-
Terminal-Bench 4.0
−20.6
V4.1 Flash vs Claude Opus 5
Within the same benchmark family, harder variants reverse the comparison. 31.2 against 51.8.
-
DeepSWE v1.1
8.7
Spread across published scaffolds
Same weights, same benchmark, different harness. 65.5 to 74.2.
The three sit together because they point in different directions. Difficulty moves the comparison one way; the scaffold around the model moves it another. Neither is a verdict on the model in isolation.
04 / Scaffold effect
Change the harness. Watch the number move.
Same model. Different harness.
A harness is the agent scaffolding around the model: the loop that plans, calls tools and retries. DeepSeek tested V4.1 Flash with several of them. The weights never change, but the reported score does.
That matters because an agent model does not work alone. The harness can move the result as much as the model does.
Model comparison for Terminal-Bench 2.1
Values are DeepSeek’s published figures. Cells with no published value are marked as not reported.
One benchmark, eight scaffolds
The table the harness view reads from. DeepSeek published both series across eight scaffolds. Every row is the same model, at maximum reasoning effort.
| Scaffold | DeepSWE v1.1Resolved | Terminal-Bench 2.1Pass@1 |
|---|---|---|
| Claude Code | 69.8 | 88.0 |
| Codex | 65.6 | 84.1 |
| OpenCode | 65.5 | 85.0 |
| Pi | 66.2 | 86.1 |
| mini-SWE | 74.2 | 90.3 |
| DSH Minimal | 72.6 | 90.6 |
| DSH Standard | 70.5 | 85.8 |
| DSH PTC | 67.6 | 85.8 |
Methodology. DeepSWE v1.1: N=8 per task · Terminal-Bench 2.1: N=3 per task · 1M-token context · max_steps=500 · Linux containers · temperature=1.0, top_p=0.95. Terminal-Bench 2.1 was evaluated without network access.
The highest published value in each series is highlighted in the harness view. The ordering differs between the two benchmarks, so a scaffold that leads on one is not thereby better on the other.
Scroll sideways for both benchmark columns.
05 / Complete results
All nineteen. And what produced them.
Nineteen benchmarks, seven model columns, and the conditions behind each figure. Every value is DeepSeek’s published result.
| Benchmark | DeepSeek V4.1-Flash | DeepSeek V4-Pro 0813 | DeepSeek V4-Flash 0731 | GLM 5.3 | Kimi K3 | GPT 5.6-Sol | Claude Opus 5 |
|---|---|---|---|---|---|---|---|
| GPQA DiamondPass@1 | 90.9 | 92.4 | 89.9 | 88.1 | 92.9 | 94.1 | 93.4 |
| HLEPass@1 | 36.8 (39.1*) | 42.7* | 37.8* | 42.0* | 43.5 | 44.5 | 56.3 |
| CodeforcesRating | 3471 | 3348 | 3289 | Not reported | Not reported | Not reported | Not reported |
| MathArena ApexPass@1 | 65.6 | 65.3 | 58.6 | Not reported | 65.6 | Not reported | Not reported |
| Terminal-Bench 2.1Pass@1 · DSH Minimal · 1M ctx | 90.6 | 87.9 | 82.7 | 88.2 | 88.3 | 88.8 | 89.1 |
| Terminal-Bench 3.0Pass@1 · DSH Minimal · 1M ctx | 30.0 | 11.8 | 7.6 | 28.3 | 17.7 | 34.4 | 43.3 |
| Terminal-Bench 4.0Pass@1 · DSH Minimal · 1M ctx | 31.2 | 12.4 | 7.0 | 37.9 | 12.6 | 39.9 | 51.8 |
| DeepSWE v1.1Resolved · mini-SWE · 1M ctx | 74.2 | 62.7 | 54.4 | 66.9 | 67.5 | 73.0 | 74.0 |
| ProgramBenchAlmost@1 · DSH Minimal · 1M ctx | 20.3 | 15.5 | Not reported | 19.0 | 17.5 | 23.0 | 37.0 |
| NL2Repo-BenchScore · DSH Minimal · 1M ctx | 64.0 | 61.5 | 54.2 | 58.0 | 58.0 | 56.8 | 75.3 |
| CyberGymPass@1 | 88.1 | 83.3 | 76.7 | 84.5 | 80.0 | 84.5 | Not reported |
| SEC-Bench ProPass@1 · Claude Code | 62.8 | 56.4 | 30.9 | Not reported | Not reported | 74.3 | Not reported |
| ExploitGymPass@1 | 15.3 | 5.4 | 1.8 | 15.0 | Not reported | 33.7 | 22.1 |
| HLE (w/tools)Pass@1 | 63.9 | 60.0 | 51.5 | 62.5 | 59.8 | Not reported | 63.6 |
| Automation-BenchPass@1 · Official scaffold | 54.8 | 43.2 | 37.7 | 48.8 | 46.7 | 45.8 | 50.3 |
| Agents' Last ExamPass@1 · Official scaffold | 31.8 | 25.7 | 25.2 | 28.5 | 27.6 | 26.7 | 28.6 |
| Chartography (w/tools)Pass@1 · Claude Code · 512k ctx | 78.9 | Not applicable | Not applicable | Not reported | 68.1 | 79.9 | 84.0 |
| BabyVision (w/tools)Pass@1 · Claude Code · 512k ctx | 89.6 | Not applicable | Not applicable | Not reported | 85.7 | 88.9 | 94.1 |
| ZeroBench-main (w/tools)Pass@5 · Claude Code · 512k ctx | 49.0 | Not applicable | Not applicable | Not reported | 41.0 | 53.0 | 52.0 |
Seven model columns. Scroll the results sideways, or focus the region and use the arrow keys.
— Not reported. DeepSeek’s table gives no value for this cell.
n/a Not applicable. The benchmark needs a modality this column does not provide. DeepSeek publishes multimodal evaluations only for the V4.1-Flash base model and lists none for the V4-Pro or V4-Flash base models.
not run Not run. The evaluation was explicitly skipped. No cell in this table carries this state. DeepSeek marks every missing case with the same dash, so the first two states are this page’s reading of its own tables rather than DeepSeek’s wording.
* Tested only on the pure-text subset of the HLE benchmark set. DeepSeek’s model card marks the same subset with a dagger.
NL2Repo-Bench disagrees between two DeepSeek primary sources. The model card’s evaluation table reports 64.0 for V4.1-Flash. DeepSeek’s API release note of the same day lists 65.4. Both were published on 10 September 2026. This page displays the model card’s evaluation-table figure and records the disagreement rather than quietly picking the friendlier number.
Where a harness is named in a row header, DeepSeek’s model card states it. Where none is named, the card does not state one.
Status of the comparison columns. DeepSeek states that from 04:00 UTC on 14 September 2026, requests to deepseek-v4-pro will be routed to V4.1-Flash until V4.1 Pro ships. That transition is scheduled, not completed. At the time of this reading the V4-Pro 0813 column above is still its own model. The V4-Flash and V4-Flash-Vision-Exp API names, by contrast, have already been retired and currently route to V4.1-Flash.
06 / What the table supports
The win is on the easy one.
“Easy” is relative and local here. It means the least demanding published variant in the Terminal-Bench family, not an easy task in absolute terms.
Terminal-Bench ships in several versions, and each one is harder than the last. V4.1-Flash leads on the version DeepSeek calls 2.1. On the two harder versions it trails Claude Opus 5, and the gap widens.
V4.1-Flash minus Claude Opus 5By Terminal-Bench variant
- Terminal-Bench 2.1 · 90.6 vs 89.1+1.5
- Terminal-Bench 3.0 · 30.0 vs 43.3−13.3
- Terminal-Bench 4.0 · 31.2 vs 51.8−20.6
V4.1-Flash minus DeepSeek V4-Pro 0813The other side of the same variants
- Terminal-Bench 2.1 · 90.6 vs 87.9+2.7
- Terminal-Bench 3.0 · 30.0 vs 11.8+18.2
- Terminal-Bench 4.0 · 31.2 vs 12.4+18.8
Both statements can be true. V4.1-Flash makes its largest generational gains where the absolute scores are lowest. On those same harder variants, another frontier model is still well ahead. A big improvement over a predecessor and a remaining gap to a competitor are not in tension. Reading either one alone would misrepresent the table.
Highest among the seven published columnsWhere V4.1-Flash leads
- Codeforces (Rating)3471
- Terminal-Bench 2.190.6
- DeepSWE v1.174.2
- CyberGym88.1
- HLE (w/tools)63.9
- Automation-Bench54.8
- Agents’ Last Exam31.8
- MathArena Apex (tied with Kimi K3)65.6
Higher values published elsewhereWhere another column reports more
- GPQA Diamond · GPT 5.6-Sol94.1
- Terminal-Bench 3.0 · Claude Opus 543.3
- Terminal-Bench 4.0 · Claude Opus 551.8
- ProgramBench · Claude Opus 537.0
- NL2Repo-Bench · Claude Opus 575.3
- SEC-Bench Pro · GPT 5.6-Sol74.3
- ExploitGym · GPT 5.6-Sol33.7
- Chartography (w/tools) · Claude Opus 584.0
- BabyVision (w/tools) · Claude Opus 594.1
- ZeroBench-main (w/tools) · GPT 5.6-Sol53.0
Inside the DeepSeek columns, V4.1-Flash reports more than V4-Pro 0813 on fourteen of the sixteen benchmarks where both are reported. It reports less on two: GPQA Diamond (90.9 against 92.4) and HLE (36.8 against 42.7*). The three remaining rows are not reported for V4-Pro 0813.
The HLE row is not ranked anywhere on this page. It mixes full-set figures with text-only subset figures marked *, which are different measurements. Comparing them as one ordering would misstate what was measured.
These are single reported figures per cell. The comparison table carries no sample sizes, no uncertainty ranges and no methodology. So nothing here supports a claim about how often any model succeeds on real work. DeepSeek does publish sample counts for the scaffold series, N=8 and N=3. Even so, three or eight samples per task do not establish a success rate.
07 / Hands-on
What it felt like to use.
Fast enough to change how I iterate.
My experience with V4.1 Flash was very good. It is quick, capable at coding, and strong value for the price. The four ratings below are my own, from ordinary day-to-day use.
Hands-on experience, not a benchmark
SpeedSpeed was the first thing I noticed.
V4.1 Flash responds quickly enough that iteration feels lightweight. In my own testing it felt faster than Muse 1.3, which made repeated coding and design changes much easier to work through.
ValueThe price changes the equation.
Pricing is one of its strongest advantages for me. DeepSeek charges peak and off-peak rates, and off-peak costs half as much. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, so when I run heavier work matters. Scheduled outside those windows, V4.1 Flash gives me a lot of capability for the cost.
First-try qualityNot always right the first time.
For frontend and visual work I had to check the output rather than trust it. That is a workflow cost rather than a dealbreaker, and it is why the design rating sits below the coding rating.
A very good developer model.
V4.1 Flash is not perfect, especially when visual work needs strong first-try judgment. But its combination of coding ability, speed, long-horizon usefulness and price makes it a model I would genuinely use.
That’s my experience with DeepSeek V4.1 Flash so far.
Source note
Published benchmark results come from DeepSeek. Hands-on observations and ratings are mine.
The benchmark tables and architecture figures on this page are based on DeepSeek’s published material. My coding, design, long-horizon, speed, value and overall-experience ratings come from my own use. They are subjective observations, not formal benchmark measurements. I have not independently reproduced DeepSeek’s full benchmark suite.