v1.5.1-native-vl.4 — portable native vision-language serving

What changed

This is the first portable native AMD395 release for the fixed Qwen3.6-35B-A3B model with complete image, video, and mixed-media serving in the same resident native process as text generation.

• native media admission and decoding, the 27-block vision stack, merger, M-RoPE injection, language prefill/decode, SSE, and function tools;
• seeded positive-temperature, top-p, full-vocabulary BF16 sampling in addition to the certified greedy path;
• batch size 1 and the 262,144-token total context window;
• no Python, PyTorch, vLLM, Triton, Transformers, or host ROCm userspace at runtime;
• portable minimal FFmpeg and curl/c-ares stacks with fail-closed media allowlists and bounded caches.

Qualification

All G1-G5 release gates passed, including frozen VL capability and correctness, strict v1.5.1 text non-regression, paired VL performance, isolated portable bundle checks on two AMD395 hosts, a one-hour 240-request resident soak, rollback, security scanning, and sealed release evidence.

Release commit: 6a6310c843eab9496ca37d19ba216856c86ba476
Qualified native source: bd012874027defa528279a357609b713e9069df4
Native engine SHA-256: fb5cae0ca5ffaa4bc3d418d5fb1630d822eae9d60f639ba6cc143e427c0cd1e9
Portable archive SHA-256: e55fb584827c7f30bd31a77ab91b9cb2d6f4e0b02bb29965dc1caed8bade0662
Public evidence SHA-256: e3daf526a63bcfa597848c120fff1c8bee2e11056858cc1407270f37b771ba6b

Download verification

sha256sum -c aima-engine-native-portable-dc4b9e19c3e5.tar.zst.sha256
sha256sum -c aima-engine-v1.5.1-native-vl.4-public-evidence.tar.zst.sha256

Model weights remain an external pinned Hugging Face Safetensors input.

Post-tag source status

The default branch also contains additive evidence plus source fixes for Approaching-AI issues #6 and #7. Those post-tag fixes do not alter this immutable .4 archive and will enter the next qualified binary release.

v1.5.1 — variable-length resident AOT prefill

What changed

v1.5.1 fixes Approaching-AI issue #5.
Cold variable-length prompts and prefix extensions now compose the resident
q1024/q2048/q4096/q8192 AOT prefill schedules. Only the final segment is
padded when needed; unmatched prompt tokens no longer fall back to serial
decode.

The release also preserves the v1.5.0 resident prefix LRU, live SSE streaming,
OpenAI function tools, disconnect cancellation, and the portable userspace
runtime.

Qualification

• all 19 batch-1 context/output performance cells passed;
• all nine full-vocabulary correctness contexts passed, with maximum KLD
  0.002174 under the frozen KLD < 0.005 and top-1 gate;
• exact q8192 128-token output identity passed;
• q8192 measured 1660 prefill tok/s and 32.30 decode tok/s for output 512;
• frozen MMLU-256 scored 218/256 (85.16%), versus 216/256 for the GB10
  reference;
• startup, exact-prefix reuse, resident HTTP, variable prompt lengths, SSE,
  tools, and the isolated archive smokes at q1024/q16384/q65536 passed.

Release commit: 6f3e669ac897eaabfeceb7f193a5e02708a4d95e
Qualified native source: 65c198415709dad6d046c247acab3dc9df2a95a0
Native engine SHA-256: a9f18771175757af080c8a1d8d7e3fb3906c9aa41b43a496686103b626f80262

Download verification

sha256sum -c aima-engine-native-portable-c12eb036ad77.tar.zst.sha256

Archive SHA-256:
4e38f90fce3feb7bccf1965d87a3ec2bebddc439ce62e75fe1bc797c6ce1a5bc.

Independent-host status

The full primary AMD395 qualification and isolated archive qualification are
complete. The independent AMD395 host was unavailable at publication, so this
release does not claim a fresh v1.5.1 second-host result. That sanitized result
will be added as append-only evidence without moving the immutable tag.

v1.5.0 — resident short-context serving

What changed

• A q8192 resident service now keeps q1024/q2048/q4096/q8192 AOT prefill buckets loaded and selects the largest fitting bucket per request.
• Exact request-prefix reuse is now a capacity-bounded LRU: four entries at q8192, two at medium long contexts and one at the largest windows.
• Health, CLI and request metrics expose resident buckets, cache capacity and executed AOT-prefill tokens.
• Release qualification now includes a privacy-safe, frozen MMLU-256 scorecard paired against identical GB10 vLLM prompt-token hashes.

Qualified result

• All 19 batch-1 performance cells passed; minimum frozen-baseline retention was 1.013x prefill and 0.9858x decode.
• Nine-context full-vocabulary correctness passed with top-1 agreement everywhere and worst KLD 0.002174 < 0.005.
• The exact q8192 128-token completion matched.
• MMLU-256 scored 216/256 (84.375%), equal to the frozen GB10 vLLM reference, with zero invalid answers and all 256 prompt-token hashes matched.
• SSE streaming, OpenAI function tools, disconnect recovery, variable prompts, multi-entry LRU, startup and isolated bundle checks passed.
• The checksum-identical archive reproduced on a second AMD395: q8192 cold prefill 1722 tok/s, decode512 32.36 tok/s, exact completion and resident HTTP all passed.

Runtime boundary

The archive is self-contained at userspace level: no host Python, PyTorch, vLLM, Triton, Transformers or ROCm userspace is required. The host still needs Linux x86-64, AMDGPU/KFD/render access, Radeon 8060S (gfx1151), the documented 96 GiB GTT setup and the separately obtained pinned model checkpoint. Model weights are not included.

sha256sum -c aima-engine-native-portable-86e806e8bc5d.tar.zst.sha256
tar --zstd -xf aima-engine-native-portable-86e806e8bc5d.tar.zst
cd aima-engine-native-portable-86e806e8bc5d
./bin/aima-engine doctor --model-dir /srv/models/Qwen3.6-35B-A3B --json
./bin/aima-engine serve --model-dir /srv/models/Qwen3.6-35B-A3B --context-tokens 8192

Portable archive SHA-256: 5ca97a234c1132ec0e715f463107faa7bdce6dab6ab053918cf1465b8d2ea62b
Public evidence SHA-256: d47c49b874d8c421fdd270a0eaa698f6938348726576a20b2c5f0931eaac5e36

The immutable release tag points to d82e6943bc50d821011ce79e95afee06f6b12a36; the additive evidence inventory is on main at 483260b0a41933bc85c35f2e364692efc66bd3db.

v1.4.1 — variable-length prompts and multi-turn fix

What changed

v1.4.1 fixes variable-length cold prompts and ordinary multi-turn conversations in the native resident engine. Every positive prompt that fits the configured cache capacity is now admitted. Independent cache misses restart from clean resident state; exact and append-prefix hits remain latency optimizations rather than admission requirements.

The fast fixed-context AOT paths and provider policies are unchanged. For transparency, a cold prompt shorter than its selected specialization, or a tail beyond it, executes that unmatched portion through the correct resident token path and therefore runs at decode rather than AOT-prefill throughput.

AMD395 qualification

The exact published binary was requalified on Radeon 8060S / gfx1151:

• all 19 batch-1 performance cells passed their frozen 0.97x floor;
• worst median changes versus v1.4.0 were -2.283% prefill and -0.1143% decode, both within the 3% measurement protocol band;
• nine full-vocabulary correctness contexts passed with maximum KLD 0.002174 < 0.005 and matching top-1 everywhere;
• the frozen q8192 completion matched all 128 token IDs exactly;
• q8192 command-to-ready median was 47.36 s against the 51.41 s ceiling;
• q32768 exact-prefix cache reduced TTFT by 2624x with 1.0002 decode-throughput retention;
• a 16-token cold prompt and exact replay, a normal 36-token next-user turn, and an unrelated short request after long-context work all returned HTTP 200 with clean cache-miss isolation;
• live SSE, function tools, disconnect recovery, one-load residency, clean shutdown and isolated portable-bundle provider smokes all passed.

The immutable release tag points to ba45639c178061f9bdadd22c86744f6924f5bf44. The qualified native executable embeds source commit 4536dbaeb6d1d013232db8150fbb6f7c3100b20a and has SHA-256 90c55cb3185d37589c5b2f8afead9ecd4fb53759beae99a95ec763143f9ad7b0.

Install

The archive is self-contained at runtime: it does not load Python, PyTorch, vLLM, Triton, Transformers or a host ROCm userspace. The host still needs Linux x86-64, AMDGPU/KFD/render nodes, a gfx1151 GPU, the documented 96 GiB GTT configuration and separately obtained Qwen3.6-35B-A3B BF16 model weights.

sha256sum -c aima-engine-native-portable-b98b7bc698ae.tar.zst.sha256
tar --zstd -xf aima-engine-native-portable-b98b7bc698ae.tar.zst
cd aima-engine-native-portable-b98b7bc698ae

./bin/aima-engine serve \
  --model-dir /srv/models/Qwen3.6-35B-A3B \
  --context-tokens 8192 \
  --host 127.0.0.1 \
  --port 8000

Portable archive SHA-256: f75562537277af8b3a0e1a92fb012761a1522b7021f3014bc1f5b8355f650d1b.

The product result, portable-bundle audit and deterministic public-evidence archive are attached for independent verification. Raw sanitized qualification reports are also committed on main at c22aee94fff6bed34384642f3a1f4e358a2043a3.

v1.4.0 — portable native AMD395 engine

AIMA AMD395 Qwen3.6 native engine v1.4.0

This is the first hardened portable release of the batch-1 native engine for
Qwen3.6-35B-A3B-BF16 on Ryzen AI Max+ 395 / Radeon 8060S (gfx1151). Model
weights are not included.

What ships

• A 105,712,311-byte relocatable native archive with a static launcher and a
  complete userspace ELF closure. Runtime Python, PyTorch, vLLM, Triton,
  Transformers and host ROCm userspace are not required.
• Resident OpenAI-compatible HTTP serving, live chunked SSE streaming,
  function tools and tool history, exact-prefix reuse, bearer authentication,
  bounded socket I/O, disconnect cancellation, graceful shutdown and a
  hardened systemd template.
• A no-load doctor command, machine-readable --build-info, recursive
  package manifest, exact component hashes and deterministic public evidence.
• A one-reader direct checkpoint-loader default that avoids same-NVMe O_DIRECT
  contention while retaining an explicit --workers override.

Exact qualification

• Release/tag commit: db54224cfcb9dae60607ccf6481e412e5c3a991e
• Qualified native source commit: 4d6c3ce48fd419f404b1a80d8e2ac6202972cec9
• Native engine SHA-256:
  a9095e5889304381ae8d716ebb262141998f72a254fc7235c290e127b85da872
• CI: CPU-safe release checks passed.
• Performance: all 19 cells passed the per-cell >= 0.97 retained-performance
  gate using two runs within 3%, or a third run and median when required.
• Correctness: all nine full-vocabulary contexts passed KLD < 0.005 and top-1
  identity. Maximum measured KLD was 0.002174; the exact 128-token completion
  hash was aa910692fd03ed4a8e89c04497751e3a28eee36c6148237f7e97c74a6dd68201.
• Startup: three fresh resident processes measured 50.25, 48.59 and
  48.07 seconds; median 48.59 seconds passed the 51.41-second ceiling.
• Prefix cache at q32768/output512: TTFT fell from 23,998 ms to 9.190 ms
  (2611x) with 99.995% decode retention and byte-identical output tokens.
• Portable isolation: checksum, 363 manifest entries, ELF closure and real
  q1024/q16384/q65536 provider smokes all passed with only HOME, LANG and
  PATH supplied.

Selected median batch-1 throughput:

Input / output	Cold prefill tok/s	Decode tok/s
1024 / 512	1633	34.05
8192 / 512	1654	32.23
16384 / 512	1438	30.68
32768 / 512	1362	28.15
65536 / 512	1175	24.62
131072 / 512	873.2	19.52
261632 / 512	562.5	13.60
261120 / 1024	561.0	13.61

The machine-readable product result and deterministic evidence archive contain
the complete matrix, per-run measurements, correctness reports and HTTP/tool
qualification.

Install

sha256sum -c aima-engine-native-portable-a15b2774e3ab.tar.zst.sha256
tar --zstd -xf aima-engine-native-portable-a15b2774e3ab.tar.zst

ENGINE=./aima-engine-native-portable-a15b2774e3ab/bin/aima-engine
MODEL=/srv/models/Qwen3.6-35B-A3B

"$ENGINE" doctor --model-dir "$MODEL" --json
"$ENGINE" serve --model-dir "$MODEL" --context-tokens 8192

See the bundled README.md, docs/INSTALL.md, docs/API.md and systemd
templates for memory configuration, authentication and lifecycle commands.

Asset integrity

• Portable archive SHA-256:
  749a2acb8b8d49b3979e1dbb9785ce3a305bb24129175747fef6330579d2f0f2
• Release-evidence archive SHA-256:
  82b1bedeceaf36ae4a4dfe8fc55a3248041b6eadb98d348ef651a87705ecc27a
• Product-result JSON SHA-256:
  bf28f94ec4b0c4bc45905e38488b5076d4110058123f97577308de698ca96c7f
• Portable-bundle result JSON SHA-256:
  be7acd68bab0bc2a20b083982cf4e418851d39822e4ebbd832a628d6510baf53

Security note: the v1.4 public tree and release evidence passed the repository
hygiene scanner, but scanning is not Git-history sanitization. Treat any secret
that ever entered history as compromised and rotate it; see SECURITY.md.

v1.3.0 — Live Streaming and Function Tools

Highlights

• True token-time HTTP/1.1 chunked SSE with incremental UTF-8, optional usage chunks, [DONE], and client-disconnect decode cancellation.
• OpenAI-compatible function tools, tool_choice, parallel_tool_calls, structured stream/non-stream tool_calls, and assistant/tool history.
• Byte-exact native Qwen tool/chat template parity without Transformers at runtime.
• Self-contained AMD395 / gfx1151 native package with one static launcher and bundle-local ROCm userspace; no host Python, PyTorch, vLLM, Triton, Transformers, or ROCm userspace installation is required.

Qualification

• 19/19 batch-1 performance cells passed the frozen >=0.97x gate. Minimum prefill retention: 1.0136x; minimum decode retention: 0.9839x.
• Against v1.2, worst decode ratio was 0.9990x; worst prefill ratio was 0.9833x at the maximum-window 261632/512 cell.
• 9/9 full-vocabulary correctness contexts passed; maximum KLD 0.002174 < 0.005, with top-1 agreement everywhere and exact 128-token identity at 8K.
• Median q8192 command-to-ready startup: 44.69 s versus the frozen 51.41 s ceiling.
• q32768 exact-prefix TTFT: 9.180 ms versus 23.98 s cold (2612x), with 0.99997x decode retention and identical output hashes.
• SSE first content arrived at 633.3 ms; stream/non-stream text and token hashes matched exactly.
• Stream and non-stream tool calls both produced get_weather({"city":"Paris"}); tool-result history completed correctly, and forced client disconnect left the one-load resident server healthy.
• Fresh isolated extracted-bundle provider smokes passed at q1024, q16384, and q65536 with a complete relocatable ELF closure.

Runtime boundary

Model weights are not included. The runtime requires Linux x86-64, the AMDGPU/KFD kernel driver and render nodes, an AMD gfx1151 GPU, and a separately obtained Qwen3.6-35B-A3B-BF16 checkpoint.

Cold HTTP requests remain exact-static-context specialized. Tool definitions and assistant/tool history count toward that context; longer requests are accepted only when they extend the resident cached token prefix.

Integrity

Archive SHA-256: a60800b573915e804ee00180ec66ebce14b10780e380ecb5aaf35e6c9d768ced

Native engine SHA-256: 7b8ae090a80b90a066c39539313ebed92c899d370fc2ccf7021cda03cec0a81d

sha256sum -c aima-engine-native-portable-9442591da597.tar.zst.sha256

v1.2.0 — Portable Native AMD395 Runtime

Highlights

• Self-contained portable native runtime for AMD Ryzen AI Max+ 395 / gfx1151.
• One static aima-engine entry point with a bundle-local ROCm userspace closure.
• No host Python, PyTorch, vLLM, Triton, Transformers, or system ROCm userspace installation is required.
• Resident OpenAI-compatible HTTP service, exact-prefix cache reuse, native tokenizer, direct Safetensors loading, and graceful shutdown.

Qualification

• 19/19 batch-1 performance cells passed the frozen >=0.97x retention gate.
• Minimum cold-prefill retention: 1.0109x; minimum decode retention: 0.9740x.
• 9/9 full-vocabulary correctness contexts passed; maximum KLD 0.002174 < 0.005, with top-1 agreement in every context.
• Exact 128-token identity passed at 8K context.
• Median command-to-ready startup: 42.52 s.
• 32K exact-prefix hit TTFT: 9.195 ms versus 24.22 s cold (2634x), with 1.0006x decode retention.
• Fresh extracted-bundle full-model smokes passed at 1K, 16K, and 64K.

Runtime boundary

The model weights are not included. The runtime requires Linux x86-64, the AMDGPU/KFD kernel driver and render nodes, an AMD gfx1151 GPU, and a separately obtained Qwen3.6-35B-A3B-BF16 checkpoint.

Integrity

Archive SHA-256: 3bfe25c49b7edb29853a821267220e9c4f7025685aa4e423605ee43e8b240672

sha256sum -c aima-engine-native-portable-805b34f5fe6f.tar.zst.sha256

AIMA AMD395 Qwen3.6 35B Linux Engine v1.0.0

Changelog

All notable changes are documented here. This project follows Semantic
Versioning.

1.0.0 - 2026-07-21

• First public, production-qualified AMD395 release.
• Added resident batch-1 BF16 engine with an OpenAI-compatible HTTP subset.
• Added exact-prefix cache reuse for one entry up to 32,768 tokens.
• Added portable startup-image preparation and registration commands.
• Added release integrity checks, environment diagnostics and native rebuild
  sources.
• Closed output-one, raw-token usage, EOS stopping and strict/exact prefix-cache
  conformance.