Insights DB

avlp12 / Kimi-K3-Alis-MLX-Dynamic-2.10bpw

Last scanned: Never Scanned

Report an issue

Versions

Version	Commit message	Commit author	Last scanned
7d066b	Retire the 2.71 bpw (v2) build: branch removed, kept only as the historical baseline row	avlp12	Aug 17, 2026
b61f06	Fix stale decode summary: 2.3 -> 8.8 plain / 10.0-10.6 spec (3.8-4.6x), add TP+MMA to lever list	avlp12
7a5c48	DSpark drafter r4 (chat-aligned, 2.25B bf16) for opt-in speculative decoding	avlp12
e9367e	Speculative decoding opt-in: 10.0-10.6 tok/s on code (+21%) — first double-digit decode; adaptive accept-gate, drafter weights included	avlp12
6eb466	Korean speed-neutral re-tiering: KL -20/-30%, PPL -3.9%, decode unchanged	avlp12
c8a9d6	v7 MMA fused-codebook prefill: 2k prefill 42.6->24.8s under TP (beats non-TP), KL-certified	avlp12
784e02	Shared-expert TP lands: 8.8 tok/s default, 4-window KL parity certification, zero extra collectives	avlp12
9600e9	Finalize TP campaign: 8.3 tok/s default config, parity+greedy-identity certification, long-prompt stall fixed via chunked prefill	avlp12
8ebd61	decode: 8.1 tok/s with attention tensor-parallel (K3_ATTN_TP=1); TP long-prompt caveat under jaccl documented	avlp12
c9824a	speed: 6.6 -> 7.0 tok/s (jaccl RDMA collectives on TB5, bit-identical transcripts)	avlp12
c8af97	speed: 6.0 -> 6.6 tok/s (dummy-row elision + GLU fusion, bit-exact; MLX 0.32 re-cert)	avlp12
46c8aa	card: install & serve guide + AI-agent deploy note + why-this-matters; runtime refresh (fusion kernels, web chat, generic 2-box launcher)	avlp12
c099e7	serving: 5.6 -> 6.0 tok/s (fused MoE router + SiTU-at-store + shared-experts packing)	avlp12
2432d3	serving: 5.4 -> 5.6 tok/s (KDA projection packing + fused glue kernels + local sampling)	avlp12
6edc75	card: decode now ~5.4 tok/s (kernel vectorization + lazy collectives)	avlp12
b605c7	card: decode 2.5 -> 4.5 tok/s (decode-specialized codebook kernel, 2.65x expert GLU)	avlp12
1773ff	card: cross-quant H2H reference (unsloth UD-IQ2_XXS, harness caveat)	avlp12
1cde5f	Update README.md	avlp12
f6e222	card: measured decode (2.5 tok/s lazy) + multimodal verification (image-to-CSS reproduction)	avlp12
8f4ed8	Add files using upload-large-folder tool	avlp12

‌
‌
‌
‌
‌
‌
‌
‌
‌
‌

Prisma AIRS detects threats in model files

With Prisma AIRS you can scan models for threats before ML developers download them for use, and apply policies based on your risk tolerance.

Learn more

Found a new threat?

Earn bounties and get recognition for your discoveries by submitting them through Huntr.

Report your finding

Version	Commit message	Commit author	Last scanned
7d066b	Retire the 2.71 bpw (v2) build: branch removed, kept only as the historical baseline row	avlp12	Aug 17, 2026
b61f06	Fix stale decode summary: 2.3 -> 8.8 plain / 10.0-10.6 spec (3.8-4.6x), add TP+MMA to lever list	avlp12
7a5c48	DSpark drafter r4 (chat-aligned, 2.25B bf16) for opt-in speculative decoding	avlp12
e9367e	Speculative decoding opt-in: 10.0-10.6 tok/s on code (+21%) — first double-digit decode; adaptive accept-gate, drafter weights included	avlp12
6eb466	Korean speed-neutral re-tiering: KL -20/-30%, PPL -3.9%, decode unchanged	avlp12
c8a9d6	v7 MMA fused-codebook prefill: 2k prefill 42.6->24.8s under TP (beats non-TP), KL-certified	avlp12
784e02	Shared-expert TP lands: 8.8 tok/s default, 4-window KL parity certification, zero extra collectives	avlp12
9600e9	Finalize TP campaign: 8.3 tok/s default config, parity+greedy-identity certification, long-prompt stall fixed via chunked prefill	avlp12
8ebd61	decode: 8.1 tok/s with attention tensor-parallel (K3_ATTN_TP=1); TP long-prompt caveat under jaccl documented	avlp12
c9824a	speed: 6.6 -> 7.0 tok/s (jaccl RDMA collectives on TB5, bit-identical transcripts)	avlp12
c8af97	speed: 6.0 -> 6.6 tok/s (dummy-row elision + GLU fusion, bit-exact; MLX 0.32 re-cert)	avlp12
46c8aa	card: install & serve guide + AI-agent deploy note + why-this-matters; runtime refresh (fusion kernels, web chat, generic 2-box launcher)	avlp12
c099e7	serving: 5.6 -> 6.0 tok/s (fused MoE router + SiTU-at-store + shared-experts packing)	avlp12
2432d3	serving: 5.4 -> 5.6 tok/s (KDA projection packing + fused glue kernels + local sampling)	avlp12
6edc75	card: decode now ~5.4 tok/s (kernel vectorization + lazy collectives)	avlp12
b605c7	card: decode 2.5 -> 4.5 tok/s (decode-specialized codebook kernel, 2.65x expert GLU)	avlp12
1773ff	card: cross-quant H2H reference (unsloth UD-IQ2_XXS, harness caveat)	avlp12
1cde5f	Update README.md	avlp12
f6e222	card: measured decode (2.5 tok/s lazy) + multimodal verification (image-to-CSS reproduction)	avlp12
8f4ed8	Add files using upload-large-folder tool	avlp12

Version

Commit message

Commit author

Last scanned

7d066b

Retire the 2.71 bpw (v2) build: branch removed, kept only as the historical baseline row

avlp12

Aug 17, 2026

b61f06

Fix stale decode summary: 2.3 -> 8.8 plain / 10.0-10.6 spec (3.8-4.6x), add TP+MMA to lever list

avlp12

7a5c48

DSpark drafter r4 (chat-aligned, 2.25B bf16) for opt-in speculative decoding

avlp12

e9367e

Speculative decoding opt-in: 10.0-10.6 tok/s on code (+21%) — first double-digit decode; adaptive accept-gate, drafter weights included

avlp12

6eb466

Korean speed-neutral re-tiering: KL -20/-30%, PPL -3.9%, decode unchanged

avlp12

c8a9d6

v7 MMA fused-codebook prefill: 2k prefill 42.6->24.8s under TP (beats non-TP), KL-certified

avlp12

784e02

Shared-expert TP lands: 8.8 tok/s default, 4-window KL parity certification, zero extra collectives

avlp12

9600e9

Finalize TP campaign: 8.3 tok/s default config, parity+greedy-identity certification, long-prompt stall fixed via chunked prefill

avlp12

8ebd61

decode: 8.1 tok/s with attention tensor-parallel (K3_ATTN_TP=1); TP long-prompt caveat under jaccl documented

avlp12

c9824a

speed: 6.6 -> 7.0 tok/s (jaccl RDMA collectives on TB5, bit-identical transcripts)

avlp12

c8af97

speed: 6.0 -> 6.6 tok/s (dummy-row elision + GLU fusion, bit-exact; MLX 0.32 re-cert)

avlp12

46c8aa

card: install & serve guide + AI-agent deploy note + why-this-matters; runtime refresh (fusion kernels, web chat, generic 2-box launcher)

avlp12

c099e7

serving: 5.6 -> 6.0 tok/s (fused MoE router + SiTU-at-store + shared-experts packing)

avlp12

2432d3

serving: 5.4 -> 5.6 tok/s (KDA projection packing + fused glue kernels + local sampling)

avlp12

6edc75

card: decode now ~5.4 tok/s (kernel vectorization + lazy collectives)

avlp12

b605c7

card: decode 2.5 -> 4.5 tok/s (decode-specialized codebook kernel, 2.65x expert GLU)

avlp12

1773ff

card: cross-quant H2H reference (unsloth UD-IQ2_XXS, harness caveat)

avlp12

1cde5f

Update README.md

avlp12

f6e222

card: measured decode (2.5 tok/s lazy) + multimodal verification (image-to-CSS reproduction)

avlp12

8f4ed8

Add files using upload-large-folder tool

avlp12

avlp12 / Kimi-K3-Alis-MLX-Dynamic-2.10bpw

Versions

Prisma AIRS detects threats in model files

Found a new threat?

Products and Services

Company

Popular Links