KVARK Flash 1.0

KVARK Flash 1.0 vs DeepSeek on 4x H100, raw vs KVARK Flash: GPQA Diamond 91.2 vs 89.6, difference βˆ’1.5 [βˆ’3.5, +0.4], βˆ’49% reasoning tokens, 1.5Γ— faster; AIME 2026 95.8 vs 98.3, +2.5 [0.0, +6.7], βˆ’29%, 1.6Γ—; HMMT 97.2 vs 94.8, βˆ’2.4 [βˆ’4.8, βˆ’0.4], βˆ’30%, 1.4Γ—; IFEval 94.8 vs 95.0, +0.2 [βˆ’1.7, +2.0], βˆ’35%, 1.4Γ—; LiveCodeBench 90.7 vs 92.0, +1.3 [βˆ’1.3, +4.0], βˆ’41%, 1.7Γ—; MMLU-Pro 88.3 vs 88.3, 0.0 [βˆ’1.2, +1.1], βˆ’34%, 1.5Γ—; FRAMES 69.5 vs 69.1, βˆ’0.5 [βˆ’1.9, +1.0], βˆ’38%, 1.4Γ—.

A serving recipe for DeepSeek V4.1-Flash-1017. This repository contains no weights.

KVARK Flash is a small OpenAI-compatible proxy in front of an unchanged DeepSeek V4.1-Flash-1017. Per request, it sets how long the model may think and what it is told before it starts. On our public benchmark set it uses 30–50% fewer reasoning tokens than DeepSeek's default setting. Accuracy stays within 2.5 points on 4x H100 and within 3.5 points on 2x DGX Spark. The largest drops are HMMT on H100 (βˆ’2.4, borderline significant) and AIME on Spark (βˆ’3.3, not significant).

Built by Egzakta AI Lab for local, sovereign deployment.

What the recipe does

Reasoning effort per task typelower effort for chat, math, science and code; higher for long planning
Routerpicks a profile from the request's tools and text; rules only, no model call
Budget forcing for codepast 28,672 thinking tokens, the thinking is closed and the model is made to answer
Stop rule for document workafter three searches with no supporting page, the model answers NOT IN DOCUMENTS
Tool-cap guardwhen the app allows no more tool calls, removes stray tool-call text and, if no answer is left, asks once for a final answer

Profiles

Call the model kvark-flash-1.0 and the router picks a profile per request. To choose one yourself, call its name directly.

Model nameEffort, H100 (SGLang)Effort, Spark (TensorFold)Thinking capRouter picks it when
kvark-flash-1.0-general0.3minimalnoneeverything else: chat, math, science
kvark-flash-1.0-code0.3minimal28,672a code tool (python, bash, edit...) or the request asks for code
kvark-flash-1.0-docsdefaultdefaultnoneretrieval tools (search, read_page...) and no code tool; stop rule added
kvark-flash-1.0-plan0.5lownonenever; ask for it by name

On Spark the names carry a -spark suffix (kvark-flash-1.0-spark, kvark-flash-1.0-spark-plan, ...). A reasoning_effort you send yourself always wins over the profile. Any other model name passes through to DeepSeek untouched.

The router reads the tool list the app sends. An app that sends code tools always gets the code profile, even for a question that needs none. Apps that mix code, search and chat tools in one agent should call a profile by name.

Results

Both setups run on the same engine and machine at the same time, with thinking on, a 65,536 token cap, the same prompts and the same sampling (temperature 1.0, top_p 0.95, top_k 20, min_p 0). "DeepSeek" is the same engine with no reasoning_effort set, so it runs at the engine's default (high on TensorFold, Spark). Differences are KVARK Flash minus DeepSeek, computed before rounding, with 95% intervals from a paired bootstrap over questions (all seeds of a question kept together). Faster is DeepSeek's mean time per answer divided by KVARK Flash's, with both running at the same time. FRAMES answers are judged by two LLMs, GLM-5.3-Flash and MiMo-V2.6-Flash; when they disagree, a rule-based check decides.

What these benchmarks exercise. None of the benchmark requests carry tools, so the router routes on text alone. LiveCodeBench goes to the code profile (the 28,672 cap); the other six run the general profile (no cap). Both profiles use effort 0.3 on H100 and minimal on Spark. The stop rule and the tool-cap guard are not exercised here: FRAMES pastes the documents into the prompt and gives no search tool.

Benchmarks. In the images, "Raw" is DeepSeek and n counts answers (questions Γ— seeds). GPQA Diamond: 198 questions Γ— 4 seeds. AIME 2026: 30 problems Γ— 4 seeds. HMMT February 2025 and 2026: 63 problems Γ— 4 seeds. IFEval: 541 prompts, prompt-level strict. LiveCodeBench: release v6 (January to April 2025), newest 150 problems. MMLU-Pro: 980 questions, 70 per category, fixed seed. FRAMES: 824 questions, documents in the prompt ("oracle docs"), LLM judge.

4x H100 NVL, SGLang, FP8

Three groups of benchmarks ran in parallel with 8 requests each, so up to 24 requests in flight per setup.

The results are in the image at the top of this card.

HMMT is borderline: the interval ends at βˆ’0.4 here but reaches 0.0 with a different bootstrap seed, and a McNemar test gives p = 0.11. Read it as effort 0.3 costing up to about 2 points on hard competition math. For that load, call kvark-flash-1.0-plan or send your own reasoning_effort. GPQA Diamond and FRAMES lean slightly negative on both machines, inside the interval.

The 28,672 cap fired on 9 of 150 LiveCodeBench answers (6%). Without the cap, 16 DeepSeek answers thought longer than that.

2x DGX Spark, TensorFold, EXL3 2.9 bpw

2 requests in flight per setup. Same prompts, cap, sampling and judges as on H100.

KVARK Flash 1.0 vs DeepSeek on 2x DGX Spark, raw vs KVARK Flash: GPQA Diamond 90.9 vs 89.9, difference βˆ’1.0 [βˆ’4.0, +2.0], βˆ’51% reasoning tokens, 2.0Γ— faster; AIME 2026 96.7 vs 93.3, βˆ’3.3 [βˆ’10.0, 0.0], βˆ’49%, 1.6Γ—; HMMT 93.7 vs 95.2, +1.6 [βˆ’4.8, +7.9], βˆ’40%, 1.3Γ—; IFEval 93.5 vs 94.6, +1.1 [βˆ’0.9, +3.1], βˆ’51%, 1.1Γ—; LiveCodeBench 90.0 vs 92.0, +2.0 [βˆ’0.7, +5.3], βˆ’30%, 1.6Γ—; MMLU-Pro 88.7 vs 88.0, βˆ’0.7 [βˆ’2.0, +0.6], βˆ’40%, 1.4Γ—; FRAMES 68.9 vs 67.7, βˆ’1.2 [βˆ’2.8, +0.4], βˆ’41%, 1.2Γ—.

No difference on Spark is significant. HMMT goes the other way from H100 (+1.6 here, βˆ’2.4 there) and AIME touches zero at the upper end, so read both as noise around a small effect on hard competition math. The cap fired on 5 of 150 LiveCodeBench answers (3%).

Speed

Seconds per answer on 4x H100, median and 95th percentile, from the benchmark run above. Both setups shared the machine with up to 24 requests in flight each, so these are times under load, not single-user latency.

BenchmarkFlash p50DeepSeek p50Flash p95DeepSeek p95
GPQA Diamond3349423768
AIME 202670894401,230
HMMT1041227641,066
IFEval2836114165
LiveCodeBench48797991,485
MMLU-Pro78137203
FRAMES1515101179

On average an answer arrives 1.4 to 1.7 times faster with KVARK Flash on H100, and 1.1 to 2.0 times faster on Spark. All Spark times are in the dataset.

Capacity on a mixed load (H100, earlier run on a held-out split). One batch of 263 tasks: LiveCodeBench Γ—1, GPQA Diamond Γ—1, AIME Γ—2 and HMMT Γ—2, run through each setup in turn. This run's answers are not in the dataset.

CorrectWall clockCorrect per hourTokens
DeepSeek241 / 2632,033 s4272.52M
KVARK Flash242 / 2631,270 s686 Γ—1.61.53M (βˆ’39%)

Other measurements (H100, held-out split)

  • Agentic coding, LiveCodeBench Γ—2 with a python sandbox tool (up to 16 calls per task): 90.8 vs 87.5 for DeepSeek, 5.9k vs 6.9k mean thinking tokens, 2.1 vs 2.5 tool calls per task.
  • AA-Omniscience (public 600): index 21.2 for KVARK Flash vs 16.7 for DeepSeek, with fewer wrong answers (27.7% vs 32.2%).
  • AA-LCR: 82.5 for KVARK Flash vs 86.0 for DeepSeek, not significant, with 58% fewer tokens.

Run it

1. Get the files. The hf command comes with pip install -U huggingface_hub.

hf download Egzakta/Kvark-Flash-1.0 --local-dir kvark-flash
cd kvark-flash

2. Start the proxy in front of your running DeepSeek V4.1-Flash-1017 server. Standard library only, no install.

The proxy sends requests under the model name in the profile file: serve DeepSeek as deepseek-v4.1-flash on SGLang (for example --served-model-name deepseek-v4.1-flash) and as DeepSeek-V4.1-Flash-1017 on TensorFold, or change upstream_model in the profile file. With any other name, requests can fail.

H100 (SGLang):

FLASH_UPSTREAM=http://localhost:30000 FLASH_PORT=8000 \
FLASH_ENCODING_DIR=/path/to/DeepSeek-V4.1-Flash-1017/encoding \
python3 proxy/flash_proxy.py

2x DGX Spark (TensorFold, with the 2x DGX Spark recipe):

FLASH_UPSTREAM=http://localhost:8890 FLASH_PORT=8000 \
FLASH_PROFILES=proxy/profiles_spark.json \
python3 proxy/flash_proxy.py
Setting Meaning
FLASH_UPSTREAM address of your DeepSeek server
FLASH_PORT port KVARK Flash listens on
FLASH_ENCODING_DIR the encoding folder from the DeepSeek checkpoint, used for budget forcing on SGLang
FLASH_PROFILES profile file; proxy/profiles.json by default
FLASH_LOG request log, flash.jsonl in the folder you start from: one JSON line per request with the profile, effort and cap used
FLASH_DEFAULT_MAX max_tokens when the request sets none; 65,536 by default

3. Call it with any OpenAI client. Turn thinking on in the request, as in all our tests: the code cap only works when thinking is on.

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="local")
response = client.chat.completions.create(
    model="kvark-flash-1.0",  # on Spark: "kvark-flash-1.0-spark"; or a profile, e.g. "kvark-flash-1.0-plan"
    messages=[{"role": "user", "content": "How many primes are there between 100 and 150?"}],
    max_tokens=65536,
    extra_body={"chat_template_kwargs": {"thinking": True}},
)
print(response.choices[0].message.content)

Endpoints: POST /v1/chat/completions (streaming and not), GET /v1/models, GET /flash/health. The API key is passed to the engine and never stored.

Tests: cd proxy && python -m pytest tests runs against a fake engine, no GPU needed.

Limitations

  • Tuned and measured on DeepSeek V4.1-Flash-1017 only. A new checkpoint needs the benchmarks re-run.
  • Lower effort can cost a few points on hard competition math: HMMT βˆ’2.4 on H100 (borderline) and AIME βˆ’3.3 on Spark (not significant). For that load, call kvark-flash-1.0-plan (kvark-flash-1.0-spark-plan on Spark); the router never picks it on its own.
  • Long multi-step planning is not shortened on purpose: cutting its thinking costs accuracy.
  • The router is rule-based. A wrong guess means the model thinks longer or shorter than needed; the request is still answered.
  • The public benchmarks run no tools, so the stop rule, the tool-cap guard and tool-based routing are not measured on them. The agentic coding run above is the only measurement here with a tool.
  • The fixed-effort baseline (effort 0.3 on every request, without the router) was not run. On six of the seven benchmarks it would be the same request as KVARK Flash.

License

Apache 2.0 for the recipe and the code in this repository. The DeepSeek V4.1-Flash weights are under DeepSeek's MIT license.

Contact us

hello@kvark.ai

Citation

@misc{kvarkflash2026,
  title  = {KVARK Flash 1.0},
  author = {Egzakta AI Lab},
  year   = {2026},
  url    = {https://huggingface.co/Egzakta/Kvark-Flash-1.0}
}
Downloads last month
28
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Collection including Egzakta/Kvark-Flash-1.0