staccs

lecore-deepseek-v4-flash-hrr

Live API (2026-08-13)

Public door: https://lecore-front.fly.dev/v1 · model id deepseek-v4-flash · OpenAI-compatible.

Table with columns: Surface, Number, Meaning
Surface	Number	Meaning
`GET /v1/models` `context_length`	128000000	Client-usable: what an SDK may send or persist via spill / `POST /v1/hrr/bind`. Not attention.
One HTTP body	~9.8M tokens	RunPod edge ~32 MiB. Bigger corpora are multiple bind calls.
vLLM `max_model_len`	131072	Tokens the transformer attends in one forward pass.
Wikipedia BIND + ASK	99M words, 33 items	Planted fact returned, `usage.gpu_tokens=938`, 9.6s. Those words were never in the model.
1B-word BIND	not done	Died around 12M words. Do not cite.

Regular-user chat test (OpenRouter-shaped, no attach): hf-user-test.md. No live pass files until recall returns a non-HTML hit.

Cold start: the GPU worker is scale-to-zero on 2×H200 (or 2×RTX PRO 6000 Blackwell), volume locked to CA-MTL-3. GET /v1/models is instant (CPU front door). First POST /v1/chat/completions after idle loads ~156GB — hold it open for minutes. If the pool has no cards, chat waits or fails; that is placement, not an auth error.

bash
# Cursor / any OpenAI client
# Override OpenAI Base URL = https://lecore-front.fly.dev/v1
# Model = deepseek-v4-flash
# Authorization: Bearer <issued key>

# Power path (optional): bind a book, then ask
curl -s https://lecore-front.fly.dev/v1/hrr/bind \
  -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d '{"text":"...corpus chunk..."}'
# → context_id
curl -s https://lecore-front.fly.dev/v1/chat/completions \
  -H "Authorization: Bearer $KEY" -H "X-HRR-Context: $CID" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"..."}]}'

Or dump a huge messages body. The gateway auto-spills. Cursor does not send X-HRR-Context; spill still runs.

Retrieval at serve time is leCore's Okapi BM25 (lexical) — corrected 2026-08-15. An earlier version of this card said "Retrieval is HRR / VSA, not BM25." That was wrong. rank.py ranks with holographic.semantic_router.holographic_bm25 over ~600-char chunks. The VSA/HRR encoder below is the FALLBACK arm, reached only when BM25 is unavailable or the corpus is over-cap; the server source annotates it "the dead dense-VSA arm". SEMANTIC_STAGE (the quantized-bge second stage) defaults to off and is off in production, so there is no dense stage either. What the VSA arm still is: Per-leaf IDF, k ≤ 0.1·D (DIM 1024), bind(NAME) / bind(CODE), fail-closed on a zero query. Shared encode_record({kind:passage}) tied every score and is gone.

For degens & non-ML folks (read this first)

Stock Flash is a whiteboard. Empty board → it forgot. You paste the book again, or you pay for the book again.

leCore puts the book in a holographic sidecar. Ask a question; the gateway finds the page, stuffs a short slice into Flash, and you pay for that slice. Scratch the crystal, the picture is still there. That’s HRR.

The 64 embed rows in this repo are a small in-weight address book from an earlier install. They do not hold Wikipedia. The 99M-word ASK used the sidecar.

Table with columns: What you care about, Stock Flash vibe, This overlay (Flash + in-weight HRR)
What you care about	Stock Flash vibe	This overlay (Flash + in-weight HRR)
“Will it remember a nonce I said last turn if I don’t paste the chat?”	No. Empty board → UNKNOWN / “give me context.”	Yes, if sticky is on. Lab: 0/5 vs 5/5.
“Three facts across a thread, then quiz with no history in `messages`?”	0/3	3/3
“98k-token prompt with a needle past 32k?”	Vanilla Flash 400s / refuses over 32k.	Recalled. ~98k auto-spills to HRR; vLLM saw 2090 tokens; needle `NEEDLE_KV_SPILL_9f3c` recalled. `max_model_len` still 32768.
“Do I have to send ?”

Side-by-side vs vanilla Flash

Memory is measured both arms on this box. Sticky-off is vanilla Flash behavior (empty messages → forget). Sticky-on is ours. Same overlay, same night, SIGD-6EF1FF5EAF.

Capability is ours greedy (temp 0, no reasoning_effort=max) vs published Flash-0731 effort splits (non-think / high / max). Official 0731 card does not list these rows; published numbers are third-party effort tables, not a same-day OpenRouter dual-serve. PR #4.

Table with columns: Meter, Vanilla Flash, Ours, Beat?
Meter	Vanilla Flash	Ours	Beat?
T2 nonce cite (empty next `messages`)	0/5 (sticky off = stock)	5/5	yes
Multi-turn 3-cite	0/3	3/3	yes
Re-prompts	paste the secrets back	1 ask, no paste	yes
98k-token needle (over 32k)	400s / refuses

SWE / Terminal-Bench / DeepSWE: not run, not claimed. Headline is the memory rows. Capability rows are no-regress. LiveCodeBench is a lite split — do not vs published. MMLU-Pro is first-n, not the full 12032.

KV spill prove (98k needle)

Gateway auto-spills a ~98k-token prompt into HRR so vLLM never sees the blob. Flash max_model_len stayed 32768. Vanilla Flash 400s / refuses over 32k.

Receipt: /workspace/lecore/gateway/kv_spill_prove.json

Table with columns: Field, Value
Field	Value
`http_status`	200
`needle_recalled`	true
`answer_excerpt`	`NEEDLE_KV_SPILL_9f3c`
`kv_spill`	true
`spill_tokens_est`	120035
(vLLM saw)

Notes from the prove file: vLLM prompt_tokens 1785 << 32768 (blob not forwarded; gateway log forward_tokens_est=597); HRR bound memory_units=37 attach_recalled=1 attach_tokens_est=1030; needle NEEDLE_KV_SPILL_9f3c fully recalled in assistant text; Flash max_model_len left at 32768.

How to use

Product path (what we serve): point any OpenAI client at https://lecore-front.fly.dev/v1, model deepseek-v4-flash, issued key. Cursor: Settings → Models → OpenAI API Key on, Override OpenAI Base URL = https://lecore-front.fly.dev/v1, add custom model deepseek-v4-flash (not a name that contains gpt-). First chat after idle can take minutes.

Weights path (this repo): overlay only. Do not serve these files as a full 48-shard model.

Get stock Flash (MIT, 48 shards) from deepseek-ai/DeepSeek-V4-Flash-0731.
Drop this overlay’s patched shard on top (see below).
Sit behind a leCore gateway if you want spill / bind / sticky. Overlay rows alone do not retrieve 99M words.
Optional X-HRR-Thread isolates lab runs. X-HRR-Sticky: off is the amnesia control.

Apply the overlay (do not download 48 shards twice)

bash
# stock Flash
huggingface-cli download deepseek-ai/DeepSeek-V4-Flash-0731 \
  --local-dir ./DeepSeek-V4-Flash-0731

# this overlay (~1.0 GiB)
huggingface-cli download staccs/lecore-deepseek-v4-flash-hrr \
  --local-dir ./lecore-deepseek-v4-flash-hrr

# replace ONLY shard 00001 (embed.weight). keep the other 47.
cp ./lecore-deepseek-v4-flash-hrr/model-00001-of-00048.safetensors \
   ./DeepSeek-V4-Flash-0731/model-00001-of-00048.safetensors
cp ./lecore-deepseek-v4-flash-hrr/lecore.json ./DeepSeek-V4-Flash-0731/lecore.json

# serve the stock dir (now with patched embed)
vllm serve ./DeepSeek-V4-Flash-0731 \
  --trust-remote-code \
  --tensor-parallel-size 2 \
  --enable-expert-parallel

Serve recipe that was live on Vast 47588973 (2×H200, Hopper-safe, max_model_len 32768): symlink the official tree, copy only the patched model-00001-of-00048.safetensors into the serve dir. Do not copy 156G.

Hopper note: official DeepGEMM MegaMoE / FP4 indexer flags are Blackwell. This overlay was served without --moe-backend deep_gemm_mega_moe and without use_fp4_indexer_cache.

What’s actually in the weights

Table with columns: Item, Fact
Item	Fact
Base	`deepseek-ai/DeepSeek-V4-Flash-0731` (MIT, 284B MoE / 13B active, 48 shards, ~156–167G)
This repo	Overlay only (~1022M): patched shard-00001 + `lecore.json` + embed-space router npz + tokenizer/config copies
`lecore.json`	`in_weight: 1`. Installed: `architecture`, `memory_index`, `router`.
`memory_index`	, , , rows (tokenizer / unused tail). 64 short passages written into those rows.

Engine: staccDOTsol/leCore.

Side-by-side (tech receipt)

Table with columns: Official Flash-0731, This overlay, Sidecar-only (in_weight=0)
	Official Flash-0731	This overlay	Sidecar-only (`in_weight=0`)
License	MIT	MIT	—
On-disk product	~156G / 48 shards	~1.0G overlay (+ stock 47 shards)	small npz
`memory_index`	none	in-weight rows 128000–128063	sidecar search, not vocab rows
Sticky SIG (headerless)

Provenance

Base: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731
Live door: https://lecore-front.fly.dev/v1 (Fly responder → RunPod 7kams7pp6wy6n9, 2×H200 / 2×RTX PRO 6000 Blackwell, volume 1kah2lfyvc CA-MTL-3)
Overlay built + served: Vast instance 47588973 (lecore-flash-serve, 2×H200), 2026-08-12
99M-word Wikipedia BIND+ASK: context ctx_01KZXPPMVRM2RAJJW2V9NX3C6X, gpu_tokens=938, 2026-08-13
SIG DIFF: /workspace/tooling-post-api/flash-memory-sigdiff-20260812.md (run SIGD-6EF1FF5EAF)

Not claims

Not a rewritten Flash attention kernel. The 99M-word ASK is spill + lexical (BM25) rank + inject on stock vLLM.
Not “128 million tokens in the transformer.” GET /v1/models advertises client-usable context. GPU attend is 131072.
Not a 1-billion-word prove. That bind died around 12M words.
Not “the 64 embed rows retrieved Wikipedia.” They did not.
Not SWE-bench / Terminal-Bench / DeepSWE / OSWorld. Those need sandboxes we did not run.
Not OpenRouter live. Provider apply is not submitted while the GPU pool can fail to place.
Not a full MMLU-Pro card. first-n 2743/3447 (79.6%) of 12032.
Not GDN registers / HRNN. Flash has no gated-delta recurrent state on this bridge.

Capability / sticky tables below are the overlay lab (Vast 47588973, then RunPod). The 99M-word row is the sidecar on the live door.

License

MIT, same as DeepSeek-V4-Flash-0731. Overlay patches are MIT. Cite DeepSeek-AI for the base:

markdown
@misc{deepseekai2026deepseekv4,
      title={DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence},
      author={DeepSeek-AI},
      year={2026},
}

Retrieval quality, measured against dense baselines (2026-08-15)

BEIR, leCore BM25 vs embedding models, hit@10 — we lose at equal k:

Table with columns: task, leCore BM25, bge-base, e5-large
task	leCore BM25	bge-base	e5-large
SciFact	0.820	0.883	0.853
NFCorpus	0.681	0.746	0.762
ArguAna	0.685	0.884	0.706

Parity needs 3.2–6.4x more chunks (crossover k: NFCorpus 32, ArguAna 48, SciFact 64). Adversarially, precision@1 falls 1.00 -> 0.19 with two decoys carrying the query's vocabulary, and recall@8 breaks between 10 and 25 competing passages.

So the honest claim is not retrieval accuracy. It is cost and operations: no embedding model, no GPU for retrieval, no re-embed when the corpus changes, deterministic and content-addressable results. Where lexical genuinely wins is identifiers and cross-file code questions (hit 0.90-1.00 by k=16 at zero embed cost). Where it degrades is paraphrase at scale: 1.00 at 47k tokens -> 0.30-0.40 @64 at 4.3M, the fidelity-vs-corpus-size curve the HRR literature predicts.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

staccs

Model Tree

Base

deepseek-ai/DeepSeek-V4-Flash-0731

Quantized

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

Live API (2026-08-13)

Public door: https://lecore-front.fly.dev/v1 · model id deepseek-v4-flash · OpenAI-compatible.

Table with columns: Surface, Number, Meaning
Surface	Number	Meaning
`GET /v1/models` `context_length`	128000000	Client-usable: what an SDK may send or persist via spill / `POST /v1/hrr/bind`. Not attention.
One HTTP body	~9.8M tokens	RunPod edge ~32 MiB. Bigger corpora are multiple bind calls.
vLLM `max_model_len`	131072	Tokens the transformer attends in one forward pass.
Wikipedia BIND + ASK	99M words, 33 items	Planted fact returned, `usage.gpu_tokens=938`, 9.6s. Those words were never in the model.
1B-word BIND	not done	Died around 12M words. Do not cite.

Regular-user chat test (OpenRouter-shaped, no attach): hf-user-test.md. No live pass files until recall returns a non-HTML hit.

bash
# Cursor / any OpenAI client
# Override OpenAI Base URL = https://lecore-front.fly.dev/v1
# Model = deepseek-v4-flash
# Authorization: Bearer <issued key>

# Power path (optional): bind a book, then ask
curl -s https://lecore-front.fly.dev/v1/hrr/bind \
  -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d '{"text":"...corpus chunk..."}'
# → context_id
curl -s https://lecore-front.fly.dev/v1/chat/completions \
  -H "Authorization: Bearer $KEY" -H "X-HRR-Context: $CID" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"..."}]}'

Or dump a huge messages body. The gateway auto-spills. Cursor does not send X-HRR-Context; spill still runs.

For degens & non-ML folks (read this first)

Stock Flash is a whiteboard. Empty board → it forgot. You paste the book again, or you pay for the book again.

The 64 embed rows in this repo are a small in-weight address book from an earlier install. They do not hold Wikipedia. The 99M-word ASK used the sidecar.

Table with columns: What you care about, Stock Flash vibe, This overlay (Flash + in-weight HRR)
What you care about	Stock Flash vibe	This overlay (Flash + in-weight HRR)
“Will it remember a nonce I said last turn if I don’t paste the chat?”	No. Empty board → UNKNOWN / “give me context.”	Yes, if sticky is on. Lab: 0/5 vs 5/5.
“Three facts across a thread, then quiz with no history in `messages`?”	0/3	3/3
“98k-token prompt with a needle past 32k?”	Vanilla Flash 400s / refuses over 32k.	Recalled. ~98k auto-spills to HRR; vLLM saw 2090 tokens; needle `NEEDLE_KV_SPILL_9f3c` recalled. `max_model_len` still 32768.
“Do I have to send ?”

Side-by-side vs vanilla Flash

Memory is measured both arms on this box. Sticky-off is vanilla Flash behavior (empty messages → forget). Sticky-on is ours. Same overlay, same night, SIGD-6EF1FF5EAF.

Table with columns: Meter, Vanilla Flash, Ours, Beat?
Meter	Vanilla Flash	Ours	Beat?
T2 nonce cite (empty next `messages`)	0/5 (sticky off = stock)	5/5	yes
Multi-turn 3-cite	0/3	3/3	yes
Re-prompts	paste the secrets back	1 ask, no paste	yes
98k-token needle (over 32k)	400s / refuses

KV spill prove (98k needle)

Gateway auto-spills a ~98k-token prompt into HRR so vLLM never sees the blob. Flash max_model_len stayed 32768. Vanilla Flash 400s / refuses over 32k.

Receipt: /workspace/lecore/gateway/kv_spill_prove.json

Table with columns: Field, Value
Field	Value
`http_status`	200
`needle_recalled`	true
`answer_excerpt`	`NEEDLE_KV_SPILL_9f3c`
`kv_spill`	true
`spill_tokens_est`	120035
(vLLM saw)

How to use

Weights path (this repo): overlay only. Do not serve these files as a full 48-shard model.

Get stock Flash (MIT, 48 shards) from deepseek-ai/DeepSeek-V4-Flash-0731.
Drop this overlay’s patched shard on top (see below).
Sit behind a leCore gateway if you want spill / bind / sticky. Overlay rows alone do not retrieve 99M words.
Optional X-HRR-Thread isolates lab runs. X-HRR-Sticky: off is the amnesia control.

Apply the overlay (do not download 48 shards twice)

bash
# stock Flash
huggingface-cli download deepseek-ai/DeepSeek-V4-Flash-0731 \
  --local-dir ./DeepSeek-V4-Flash-0731

# this overlay (~1.0 GiB)
huggingface-cli download staccs/lecore-deepseek-v4-flash-hrr \
  --local-dir ./lecore-deepseek-v4-flash-hrr

# replace ONLY shard 00001 (embed.weight). keep the other 47.
cp ./lecore-deepseek-v4-flash-hrr/model-00001-of-00048.safetensors \
   ./DeepSeek-V4-Flash-0731/model-00001-of-00048.safetensors
cp ./lecore-deepseek-v4-flash-hrr/lecore.json ./DeepSeek-V4-Flash-0731/lecore.json

# serve the stock dir (now with patched embed)
vllm serve ./DeepSeek-V4-Flash-0731 \
  --trust-remote-code \
  --tensor-parallel-size 2 \
  --enable-expert-parallel

Hopper note: official DeepGEMM MegaMoE / FP4 indexer flags are Blackwell. This overlay was served without --moe-backend deep_gemm_mega_moe and without use_fp4_indexer_cache.

What’s actually in the weights

Table with columns: Item, Fact
Item	Fact
Base	`deepseek-ai/DeepSeek-V4-Flash-0731` (MIT, 284B MoE / 13B active, 48 shards, ~156–167G)
This repo	Overlay only (~1022M): patched shard-00001 + `lecore.json` + embed-space router npz + tokenizer/config copies
`lecore.json`	`in_weight: 1`. Installed: `architecture`, `memory_index`, `router`.
`memory_index`	, , , rows (tokenizer / unused tail). 64 short passages written into those rows.

Engine: staccDOTsol/leCore.

Side-by-side (tech receipt)

Table with columns: Official Flash-0731, This overlay, Sidecar-only (in_weight=0)
	Official Flash-0731	This overlay	Sidecar-only (`in_weight=0`)
License	MIT	MIT	—
On-disk product	~156G / 48 shards	~1.0G overlay (+ stock 47 shards)	small npz
`memory_index`	none	in-weight rows 128000–128063	sidecar search, not vocab rows
Sticky SIG (headerless)

Provenance

Base: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731
Live door: https://lecore-front.fly.dev/v1 (Fly responder → RunPod 7kams7pp6wy6n9, 2×H200 / 2×RTX PRO 6000 Blackwell, volume 1kah2lfyvc CA-MTL-3)
Overlay built + served: Vast instance 47588973 (lecore-flash-serve, 2×H200), 2026-08-12
99M-word Wikipedia BIND+ASK: context ctx_01KZXPPMVRM2RAJJW2V9NX3C6X, gpu_tokens=938, 2026-08-13
SIG DIFF: /workspace/tooling-post-api/flash-memory-sigdiff-20260812.md (run SIGD-6EF1FF5EAF)

Not claims

Not a rewritten Flash attention kernel. The 99M-word ASK is spill + lexical (BM25) rank + inject on stock vLLM.
Not “128 million tokens in the transformer.” GET /v1/models advertises client-usable context. GPU attend is 131072.
Not a 1-billion-word prove. That bind died around 12M words.
Not “the 64 embed rows retrieved Wikipedia.” They did not.
Not SWE-bench / Terminal-Bench / DeepSWE / OSWorld. Those need sandboxes we did not run.
Not OpenRouter live. Provider apply is not submitted while the GPU pool can fail to place.
Not a full MMLU-Pro card. first-n 2743/3447 (79.6%) of 12032.
Not GDN registers / HRNN. Flash has no gated-delta recurrent state on this bridge.

Capability / sticky tables below are the overlay lab (Vast 47588973, then RunPod). The 99M-word row is the sidecar on the live door.

License

MIT, same as DeepSeek-V4-Flash-0731. Overlay patches are MIT. Cite DeepSeek-AI for the base:

markdown
@misc{deepseekai2026deepseekv4,
      title={DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence},
      author={DeepSeek-AI},
      year={2026},
}

Retrieval quality, measured against dense baselines (2026-08-15)

BEIR, leCore BM25 vs embedding models, hit@10 — we lose at equal k:

Table with columns: task, leCore BM25, bge-base, e5-large
task	leCore BM25	bge-base	e5-large
SciFact	0.820	0.883	0.853
NFCorpus	0.681	0.746	0.762
ArguAna	0.685	0.884	0.706

lecore-deepseek-v4-flash-hrr

README

Live API (2026-08-13)

For degens & non-ML folks (read this first)

Side-by-side vs vanilla Flash

KV spill prove (98k needle)

How to use

Apply the overlay (do not download 48 shards twice)

What’s actually in the weights

Side-by-side (tech receipt)

Provenance

Not claims

License

Retrieval quality, measured against dense baselines (2026-08-15)

Explore FriendliAI today

README

Live API (2026-08-13)

For degens & non-ML folks (read this first)

Side-by-side vs vanilla Flash

KV spill prove (98k needle)

How to use

Apply the overlay (do not download 48 shards twice)

What’s actually in the weights

Side-by-side (tech receipt)

Provenance

Not claims

License

Retrieval quality, measured against dense baselines (2026-08-15)