{ "meta": { "attribution": "Anchor Terminal (https://www.anchorterminal.com)", "attributionRequired": true, "credit": "CC BY 4.0, credit Anchor Terminal and link the page or API address you took it from (https://www.anchorterminal.com/terms/#data)", "docs": "https://www.anchorterminal.com/docs/", "generatedAt": "2026-10-31", "license": "CC-BY-4.0", "method": "https://www.anchorterminal.com/benchmark/", "methodology": "0.7", "openapi": "https://www.anchorterminal.com/openapi.json", "preview": false, "run": "2026-10-01", "runLabel": "October 2026 research run", "terms": "https://www.anchorterminal.com/terms/#data" }, "tool": { "slug": "hyperqwen", "name": "HyperQwen", "vendor": "syv.ai (syv-ai on GitHub)", "vendorUrl": "https://syv.ai", "kind": "http-api", "category": "local-ai", "summary": "HyperQwen is syv.ai's open-source serving setup for Qwen3.8-27B on one 24 GB NVIDIA GPU. A patch series against a pinned vLLM, a requantisation pipeline and Docker Compose profiles run vLLM's OpenAI-compatible API on port 18020.", "url": "https://www.anchorterminal.com/tools/hyperqwen", "markdownUrl": "https://www.anchorterminal.com/tools/hyperqwen.md", "slimMarkdownUrl": "https://www.anchorterminal.com/tools/hyperqwen.min.md", "jsonUrl": "https://www.anchorterminal.com/api/v1/tools/hyperqwen.json", "repo": "https://github.com/syv-ai/HyperQwen", "license": "Apache-2.0", "transports": [ "http" ], "packages": [ { "registry": "oci", "name": "ghcr.io/syv-ai/hyperqwen" } ], "auth": "none", "authNotes": "No credential by default. `VLLM_API_KEY` in `.env` (or an `api_key.txt` file) turns on vLLM's Bearer check. With no key, the launcher binds 127.0.0.1 outside a container, and Docker Compose publishes the port on 127.0.0.1 unless `BIND` is changed. With a key set, the launcher binds every interface. The key is vLLM's own, one static key with no scopes, which vLLM checks only under `/v1`, `/v2`, `/inference` and `/cohere` (per the vLLM listing's check of 9 October 2026, https://github.com/vllm-project/vllm/blob/main/docs/usage/security.md).", "pricing": "free", "pricingNotes": "Free under Apache-2.0 with no account, key or card. Nothing is sold. The README links a Ko-fi page for donations towards GPU time. The owner pays for a 24 GB NVIDIA GPU, about 40 GB of disk and electricity.", "priceSummary": "Free · OSS", "where": "local", "x402": { "level": "no", "evidence": "No x402, MPP or L402 in the README, the docs or the launch scripts (checked 2026-10-30). Nothing is sold.", "endpoints": [] }, "toolCount": null, "popularity": { "githubStars": 2400, "npmWeekly": null, "pypiWeekly": null, "asOf": "2026-10-30" }, "docsUrl": "https://github.com/syv-ai/HyperQwen#readme", "capabilities": [ "inference.local", "inference.open-weights" ], "tags": [ "open-source", "local", "self-hosted", "free", "no-card", "apache-2.0", "openai-compatible", "docker", "tool-calling", "reasoning", "pre-1.0" ], "graded": true, "anchor": { "graded": true, "score": 56.9, "grade": "C", "agentReady": false, "rank": 738, "ranked": true, "rankOf": 1239, "categoryRank": 11, "methodology": "0.7", "run": "2026-10-01", "scores": { "ergonomics": 65, "maintenance": 68, "payments": 60, "reliability": 59, "schema": 56, "security": 41, "transparency": 55 }, "pending": [ "performance", "tasks" ], "breakdown": [ { "key": "reliability", "name": "Reliability", "weight": 16, "effectiveWeight": 20, "score": 59, "points": 11.8, "reason": "Read with the local-software lines, since HyperQwen runs on the owner's GPU with no hosted service. A prebuilt image on ghcr.io with an immutable tag per commit, and a venv route, with the GPU, driver, Python and disk needs stated and Volta ruled out. The Compose route still needs a clone of the repository, and first start downloads an unpinned third-party checkpoint (18 of 20). Two GitHub workflows run about 40 GPU-less checks, and the image build applies every patch and runs the build-time tests. Both badges read passing on 30 October 2026 per shields.io. Nothing in CI runs on a GPU, where the serving path runs (20 of 25). shields.io counts 0 open and 146 closed issues and 1 open pull request. The README itself still describes an open speculation fault on sm80 cards (#98, #72), and the threads weren't read because github.com wasn't fetched (18 of 25). No tags, releases or changelog. The per-pin notes (docs/vllm-0.29.md to vllm-0.32.md) and PATCHES.md record what changed with each vLLM upgrade (3 of 15). No version number and nothing declared stable (0).", "checked": "2026-10-30" }, { "key": "performance", "name": "Performance", "weight": 10, "effectiveWeight": 0, "pending": true, "points": 0, "reason": "Pending. Latency is measured per call by our probes, which haven't run yet, so this run doesn't score it. Its weight is shared across the assessed categories until the first probe window closes." }, { "key": "schema", "name": "Schema \u0026 documentation", "weight": 13, "effectiveWeight": 16.25, "score": 56, "points": 9.1, "reason": "Read with the API lines, for the local server an agent calls, which is vLLM 0.32.0's OpenAI-compatible server. vLLM's FastAPI app serves its own OpenAPI document at runtime. HyperQwen publishes none and documents only the base URL and the served model name (12 of 25). No llms.txt. The docs are Markdown files in the repository, readable raw (5 of 10). The README's decision tree and the knob tables say which setup suits which workload and when not to use one (for example `CTX=huge` only when the request wouldn't otherwise fit, `SPEC_ATTN=0` past 64k). The routes themselves are left to vLLM's docs (14 of 20). vLLM's typed request models, tool argument schemas enforced by default (`TOOL_STRICT=parameter`), and launchers that refuse invalid knob values before boot (11 of 15). An OpenCode client config, benchmark commands and a 64-item gotchas file with failure signatures such as the 400 for `tool_choice: auto` without a parser. No curl example and no error reference (9 of 15). No versions or changelog, a fixed `/v1` prefix and model name, and PATCHES.md with an upstream reference for every patch (5 of 15).", "checked": "2026-10-30" }, { "key": "ergonomics", "name": "Agent ergonomics", "weight": 13, "effectiveWeight": 16.25, "score": 65, "points": 10.56, "reason": "Read with the API lines. `max_tokens`, structured outputs, a reasoning parser that separates thinking from the answer, and a template step that maps OpenAI `reasoning_effort` values the model's template doesn't know. Context is 65,536 tokens by default, 131,072 or 245,760 in the long setups (18 of 25). No list routes worth paging, `/metrics` always and per-request timing with `REQ_METRICS=1`, and cached prompt tokens reported on every request (12 of 20). OpenAI-shaped errors from vLLM. The launcher's own refusals print a reason, and the gotchas file pairs error messages with fixes, with no list of codes (12 of 20). Generation is stateless and safe to repeat, and an SSE keep-alive patch keeps long prefills from tripping proxy timeouts. No retry guidance found (11 of 20). One `docker compose up` with tool calling, the right parser for this model and a loopback bind preset, and OpenAI clients work unchanged. First start can exceed 30 minutes and there is no SDK of its own (12 of 15).", "checked": "2026-10-30" }, { "key": "security", "name": "Security \u0026 auth", "weight": 14, "effectiveWeight": 17.5, "score": 41, "points": 7.18, "reason": "Read with the tool checklist. vLLM's optional static key, set by `VLLM_API_KEY` or `api_key.txt`, with no scopes and checked only under `/v1`, `/v2`, `/inference` and `/cohere` (per the vLLM listing's check of 9 October 2026). HyperQwen's launcher binds 127.0.0.1 when no key is set, Compose publishes on 127.0.0.1 by default, a CI check tests the no-key bind, and `.env.example` says the key is required before exposing the port (15 of 30). Inference only, but vLLM's control routes such as `/pause` answer without the key, and the container runs as root (6 of 20). The server returns model output and fetches no outside content with the default `VISION=0`. No prompt-injection guidance (9 of 15). vLLM's access logs and Prometheus `/metrics`, with optional per-request timing, and no per-key record (8 of 15). No SECURITY.md, disclosure address or advisories, and no security.txt on syv.ai. The image is rebuilt weekly for base-image fixes, but the base weights come from a third-party repository (dbirks/Qwen3.8-27B-W4A16-AutoRound) downloaded with no pinned revision or checksum of the project's own (3 of 20).", "checked": "2026-10-30" }, { "key": "payments", "name": "Payments \u0026 pricing", "weight": 10, "effectiveWeight": 12.5, "score": 60, "points": 7.5, "reason": "Read with the self-hosted rule. No x402, MPP or L402, and nothing is sold. The Ko-fi link takes donations and buys nothing (0). Free under Apache-2.0 with no account, key or card, so 20, 20 and 20 on the last three lines.", "checked": "2026-10-30" }, { "key": "tasks", "name": "Task success", "weight": 10, "effectiveWeight": 0, "pending": true, "points": 0, "reason": "Pending. Task success needs the category task suites run through each tool, which haven't run yet, so this run doesn't score it. Its weight is shared across the assessed categories until then. A data provider's data-quality score is published on its listing now and becomes half of this category when it's scored." }, { "key": "maintenance", "name": "Maintenance \u0026 community", "weight": 7, "effectiveWeight": 8.75, "score": 68, "points": 5.95, "reason": "No versioned release. Every push to main that touches code publishes `latest` and an immutable `sha-\u003c7\u003e` image from CI, and shields.io reports the last commit on 30 October 2026 (25 of 30). No releases or dated changelog entries. Four vLLM pins (0.29.0 to 0.32.0) have dated upgrade notes since August, and shields.io counts 207 commits in the last month (10 of 20). 0 open of 146 issues, 1 open of 277 pull requests and 31 contributors, with community field reports written into docs/reproductions with credit. Reply times weren't read (20 of 25). Not an MCP server and no SDK of its own. OpenAI clients work against it (5 of 15). Pinned to vLLM 0.32.0 with the CUDA compiler set pinned to match, a weekly image rebuild and CI that fails when PATCHES.md is out of date (8 of 10).", "checked": "2026-10-30" }, { "key": "transparency", "name": "Transparency \u0026 trust", "weight": 7, "effectiveWeight": 8.75, "score": 55, "points": 4.81, "note": "editorial 60, provenance 49", "reason": "The editorial half. Apache-2.0 for the repository, the vendored pth-flash-attn source and the fa2 plugin (with FlashAttention's BSD-3-Clause notices kept), and Apache-2.0 on the base model and every checkpoint it downloads (30). No privacy or data statement for the software. The Docker image turns off vLLM's usage statistics, but the venv launchers don't, and nothing in the README or docs says so. Outbound calls are downloads from Hugging Face and ghcr.io (12 of 30). PATCHES.md says what retires each patch, and upgrade notes say what changed per vLLM pin. Knobs and setups change without a deprecation notice or dates (8 of 20). The telemetry switch is set in the Dockerfile and not documented for venv installs, which inherit vLLM's default (10 of 20).", "checked": "2026-10-30" } ], "assessment": { "date": "2026-10-30", "basis": "public evidence", "confidence": "medium", "notes": { "ergonomics": "Read with the API lines. `max_tokens`, structured outputs, a reasoning parser that separates thinking from the answer, and a template step that maps OpenAI `reasoning_effort` values the model's template doesn't know. Context is 65,536 tokens by default, 131,072 or 245,760 in the long setups (18 of 25). No list routes worth paging, `/metrics` always and per-request timing with `REQ_METRICS=1`, and cached prompt tokens reported on every request (12 of 20). OpenAI-shaped errors from vLLM. The launcher's own refusals print a reason, and the gotchas file pairs error messages with fixes, with no list of codes (12 of 20). Generation is stateless and safe to repeat, and an SSE keep-alive patch keeps long prefills from tripping proxy timeouts. No retry guidance found (11 of 20). One `docker compose up` with tool calling, the right parser for this model and a loopback bind preset, and OpenAI clients work unchanged. First start can exceed 30 minutes and there is no SDK of its own (12 of 15).", "maintenance": "No versioned release. Every push to main that touches code publishes `latest` and an immutable `sha-\u003c7\u003e` image from CI, and shields.io reports the last commit on 30 October 2026 (25 of 30). No releases or dated changelog entries. Four vLLM pins (0.29.0 to 0.32.0) have dated upgrade notes since August, and shields.io counts 207 commits in the last month (10 of 20). 0 open of 146 issues, 1 open of 277 pull requests and 31 contributors, with community field reports written into docs/reproductions with credit. Reply times weren't read (20 of 25). Not an MCP server and no SDK of its own. OpenAI clients work against it (5 of 15). Pinned to vLLM 0.32.0 with the CUDA compiler set pinned to match, a weekly image rebuild and CI that fails when PATCHES.md is out of date (8 of 10).", "payments": "Read with the self-hosted rule. No x402, MPP or L402, and nothing is sold. The Ko-fi link takes donations and buys nothing (0). Free under Apache-2.0 with no account, key or card, so 20, 20 and 20 on the last three lines.", "reliability": "Read with the local-software lines, since HyperQwen runs on the owner's GPU with no hosted service. A prebuilt image on ghcr.io with an immutable tag per commit, and a venv route, with the GPU, driver, Python and disk needs stated and Volta ruled out. The Compose route still needs a clone of the repository, and first start downloads an unpinned third-party checkpoint (18 of 20). Two GitHub workflows run about 40 GPU-less checks, and the image build applies every patch and runs the build-time tests. Both badges read passing on 30 October 2026 per shields.io. Nothing in CI runs on a GPU, where the serving path runs (20 of 25). shields.io counts 0 open and 146 closed issues and 1 open pull request. The README itself still describes an open speculation fault on sm80 cards (#98, #72), and the threads weren't read because github.com wasn't fetched (18 of 25). No tags, releases or changelog. The per-pin notes (docs/vllm-0.29.md to vllm-0.32.md) and PATCHES.md record what changed with each vLLM upgrade (3 of 15). No version number and nothing declared stable (0).", "schema": "Read with the API lines, for the local server an agent calls, which is vLLM 0.32.0's OpenAI-compatible server. vLLM's FastAPI app serves its own OpenAPI document at runtime. HyperQwen publishes none and documents only the base URL and the served model name (12 of 25). No llms.txt. The docs are Markdown files in the repository, readable raw (5 of 10). The README's decision tree and the knob tables say which setup suits which workload and when not to use one (for example `CTX=huge` only when the request wouldn't otherwise fit, `SPEC_ATTN=0` past 64k). The routes themselves are left to vLLM's docs (14 of 20). vLLM's typed request models, tool argument schemas enforced by default (`TOOL_STRICT=parameter`), and launchers that refuse invalid knob values before boot (11 of 15). An OpenCode client config, benchmark commands and a 64-item gotchas file with failure signatures such as the 400 for `tool_choice: auto` without a parser. No curl example and no error reference (9 of 15). No versions or changelog, a fixed `/v1` prefix and model name, and PATCHES.md with an upstream reference for every patch (5 of 15).", "security": "Read with the tool checklist. vLLM's optional static key, set by `VLLM_API_KEY` or `api_key.txt`, with no scopes and checked only under `/v1`, `/v2`, `/inference` and `/cohere` (per the vLLM listing's check of 9 October 2026). HyperQwen's launcher binds 127.0.0.1 when no key is set, Compose publishes on 127.0.0.1 by default, a CI check tests the no-key bind, and `.env.example` says the key is required before exposing the port (15 of 30). Inference only, but vLLM's control routes such as `/pause` answer without the key, and the container runs as root (6 of 20). The server returns model output and fetches no outside content with the default `VISION=0`. No prompt-injection guidance (9 of 15). vLLM's access logs and Prometheus `/metrics`, with optional per-request timing, and no per-key record (8 of 15). No SECURITY.md, disclosure address or advisories, and no security.txt on syv.ai. The image is rebuilt weekly for base-image fixes, but the base weights come from a third-party repository (dbirks/Qwen3.8-27B-W4A16-AutoRound) downloaded with no pinned revision or checksum of the project's own (3 of 20).", "transparency": "The editorial half. Apache-2.0 for the repository, the vendored pth-flash-attn source and the fa2 plugin (with FlashAttention's BSD-3-Clause notices kept), and Apache-2.0 on the base model and every checkpoint it downloads (30). No privacy or data statement for the software. The Docker image turns off vLLM's usage statistics, but the venv launchers don't, and nothing in the README or docs says so. Outbound calls are downloads from Hugging Face and ghcr.io (12 of 30). PATCHES.md says what retires each patch, and upgrade notes say what changed per vLLM pin. Knobs and setups change without a deprecation notice or dates (8 of 20). The telemetry switch is set in the Dockerfile and not documented for venv installs, which inherit vLLM's default (10 of 20)." }, "sources": [ { "what": "README (setups, headline figures, roadmap, field tests)", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/README.md", "seen": "2026-10-30" }, { "what": "PATCHES.md (60 patches against vLLM 0.32.0)", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/PATCHES.md", "seen": "2026-10-30" }, { "what": "docker-compose.yml (profiles, 127.0.0.1 publish, healthcheck)", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/docker-compose.yml", "seen": "2026-10-30" }, { "what": ".env.example (key and BIND guidance)", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/.env.example", "seen": "2026-10-30" }, { "what": "Dockerfile (vLLM 0.32.0, build-time tests, usage statistics off)", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/Dockerfile", "seen": "2026-10-30" }, { "what": "single-user launcher (served model name, tool parser, strict level)", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/single-user/start_qwen.sh", "seen": "2026-10-30" }, { "what": "bind-host and key resolution", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/launcher_common.sh", "seen": "2026-10-30" }, { "what": "key resolver", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/resolve_api_key.sh", "seen": "2026-10-30" }, { "what": "model preparation run on every container start", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/docker/prepare.sh", "seen": "2026-10-30" }, { "what": "CI workflow, patch integrity", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/.github/workflows/patch-integrity.yml", "seen": "2026-10-30" }, { "what": "CI workflow, docker image", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/.github/workflows/docker-image.yml", "seen": "2026-10-30" }, { "what": "bare-metal install", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/docs/install.md", "seen": "2026-10-30" }, { "what": "Docker and WSL2 notes", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/docs/docker.md", "seen": "2026-10-30" }, { "what": "client configuration", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/docs/clients.md", "seen": "2026-10-30" }, { "what": "benchmarks and comparison with ninfer-3090", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/docs/benchmarks.md", "seen": "2026-10-30" }, { "what": "quality per configuration", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/docs/quality.md", "seen": "2026-10-30" }, { "what": "community reproductions", "url": "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/docs/reproductions/README.md", "seen": "2026-10-30" }, { "what": "repository counts (stars, issues, pull requests, contributors, commits, CI status, created month)", "url": "https://img.shields.io/github/created-at/syv-ai/HyperQwen.json", "seen": "2026-10-30" }, { "what": "companion fast-variant files, first commit 18 August 2026", "url": "https://huggingface.co/api/models/syvai/qwen3.8-27b-3090-fast-variant/commits/main", "seen": "2026-10-30" }, { "what": "base W4A16 checkpoint metadata", "url": "https://huggingface.co/api/models/dbirks/Qwen3.8-27B-W4A16-AutoRound", "seen": "2026-10-30" }, { "what": "Qwen3.8-27B model metadata (Apache-2.0, created 5 August 2026)", "url": "https://huggingface.co/api/models/Qwen/Qwen3.8-27B", "seen": "2026-10-30" }, { "what": "syv.ai organisation on Hugging Face (links syv.ai and github.com/syv-ai)", "url": "https://huggingface.co/syvai", "seen": "2026-10-30" }, { "what": "syv.ai contact page (address, CVR)", "url": "https://syv.ai/kontakt", "seen": "2026-10-30" }, { "what": "AlphaSignal item indexed on hermes-ai.net, 19 September 2026 (press lead)", "url": "https://hermes-ai.net/news/hyperqwen-runs-qwen3-8-27b-on-a-single-rtx-3090-at-1-035-tok-s/", "seen": "2026-10-30" } ], "openQuestions": [ "The lead matched. HyperQwen is runnable software (Compose profiles, a prebuilt image, 60 vLLM patches, launchers and CI), not only a benchmark write-up.", "firstReleased is 18 August 2026, the first commit of the companion Hugging Face files that name the repository under its earlier name, syv-ai/qwen38-27b-rtx3090, and the date of the earliest batch measurement in batch/README.md. shields.io gives only the creation month, August 2026. The date of the rename to HyperQwen wasn't established first-hand.", "unchecked: github.com and api.github.com. Our proxy refused api.github.com for this repository and github.com/robots.txt couldn't be read, so github.com wasn't fetched. Repository counts and CI status come from shields.io, files from raw.githubusercontent.com. Issue threads, reply times and the contributor list weren't read.", "The headline figures (127 tokens a second for one stream, about 1,035 aggregate at 64 concurrent) are the project's claims. The community reproductions in docs/reproductions come from GitHub issues we didn't read. The AlphaSignal item repeats the README.", "Whether vLLM 0.32.0 carries the fixes for the vLLM advisories recorded on the vllm listing (including GHSA-h3rc-6mm3-gc2m of 6 October 2026) wasn't checked.", "Who maintains HyperQwen at syv.ai isn't stated first-hand. The README's Ko-fi link is ko-fi.com/mhenrichsen, the handle of syv.ai's founder Mads Henrichsen, and ko-fi.com wasn't read." ] }, "negative": 0, "verdict": "HyperQwen packages a patched vLLM 0.32.0 and requantised Qwen3.8-27B weights into one container with tool calling preset and a loopback bind when no key is set. It runs one model family, its speed figures come from one RTX 3090, and it has no versioned releases, changelog or security policy.", "bestFor": "An operator with one 24 GB NVIDIA card who wants Qwen3.8-27B behind an OpenAI-compatible endpoint for a coding or tool-calling agent, with long context.", "strengths": [ "Prebuilt image on ghcr.io with an immutable tag per commit, built by CI that applies all 60 patches and runs the build-time tests", "Tool calling on by default with the `qwen3_coder` parser and argument schemas enforced (`TOOL_STRICT=parameter`)", "With no API key the server listens on 127.0.0.1 only, and a CI check tests that default", "Five documented setups from 64k to 240k context, with the measured cost of each knob", "Apache-2.0 throughout, with vendored kernels and patches carrying their upstream references" ], "weaknesses": [ "Serves Qwen3.8-27B and checkpoints derived from it only. Other models are on the roadmap", "No tags, releases or changelog. Changes reach users through `latest` on every push to main", "No SECURITY.md or disclosure route, and the base weights come from a third-party repository with no pinned revision", "vLLM's key covers only `/v1`-style prefixes, so control routes such as `/pause` answer without it", "First start pulls a 9.5 GB image and about 20 GB of weights and can take over 30 minutes" ], "agentNotes": [ "Call http://127.0.0.1:18020/v1 with model `qwen3.8-27b` and the `VLLM_API_KEY` value, or any string if no key was set", "Poll `/health` before the first request. A first start downloads and prepares the model and compiles kernels, which can take over 30 minutes", "Declare the context window your setup uses (65,536 by default, 131,072 for `CTX=long`, 245,760 for `CTX=huge`) so the client doesn't truncate", "Ask the owner to set `VLLM_API_KEY` before setting `BIND=0.0.0.0`, and to front the port with a proxy that allowlists routes", "Treat the README's tokens-per-second figures as the project's measurements on one RTX 3090 at 250 W" ], "metrics": { "kind": "local", "measured": false }, "reviewCount": 0, "avgRating": 0, "history": [ { "basis": "public evidence", "confidence": "medium", "grade": "C", "methodology": "0.7", "pending": [ "performance", "tasks" ], "run": "2026-10-01", "runLabel": "October 2026 research run", "score": 56.9 } ], "editorialScores": { "ergonomics": 65, "maintenance": 68, "payments": 60, "reliability": 59, "schema": 56, "security": 41, "transparency": 60 }, "provenanceScore": 49, "checked": { "categories": { "ergonomics": "2026-10-30", "maintenance": "2026-10-30", "payments": "2026-10-30", "reliability": "2026-10-30", "schema": "2026-10-30", "security": "2026-10-30", "transparency": "2026-10-30" }, "last": "2026-10-30", "oldest": "2026-10-30", "stale": false, "staleAfterDays": 90 }, "scoreHistory": [ { "date": "2026-10-30", "grade": "C", "kind": "listed", "method": "0.7", "score": 56.9 } ] }, "connect": { "install": "git clone https://github.com/syv-ai/HyperQwen \u0026\u0026 cd HyperQwen\ncp .env.example .env\n\ndocker compose --profile single up -d # one or a few people chatting\ndocker compose --profile batch up -d # API backend, many concurrent requests" }, "letme": { "capability": "https://letme.dev/inference.local", "tool": "https://letme.dev/hyperqwen" }, "notable": [ "The repository has no tags and no GitHub releases, and shields.io gives its creation month as August 2026. The companion Hugging Face repository syvai/qwen3.8-27b-3090-fast-variant was first committed on 18 August 2026 and names the project's earlier repository, syv-ai/qwen38-27b-rtx3090 (https://huggingface.co/api/models/syvai/qwen3.8-27b-3090-fast-variant/commits/main, https://img.shields.io/github/created-at/syv-ai/HyperQwen.json)", "The headline figures are the project's own, from one RTX 3090 capped at 250 W that the README says is shared with training jobs. 127 tokens a second for one stream, about 1,035 aggregate decode at 64 concurrent requests (128 in, 512 out) and 381 while an answer quotes its prompt (https://raw.githubusercontent.com/syv-ai/HyperQwen/main/README.md)", "docs/reproductions lists runs by other people on other hardware, taken from GitHub issues, such as 135.5 tokens a second for one stream on an RTX 4090 and 182.8 on two RTX 3090s at TP=2. These are relayed by the project and were not read at their source (https://raw.githubusercontent.com/syv-ai/HyperQwen/main/docs/reproductions/README.md)", "PATCHES.md lists 60 patches against vLLM 0.32.0 (27 fixes, 19 additions vLLM lacks, 6 local changes, 5 fixes to the project's own additions and 3 backports), with 10 marked upstreamed and 16 as open upstream pull requests (https://raw.githubusercontent.com/syv-ai/HyperQwen/main/PATCHES.md)", "The base weights are a third-party W4A16 quantisation, dbirks/Qwen3.8-27B-W4A16-AutoRound (created 14 August 2026), which docker/prepare.sh downloads with no pinned revision and then rewrites in place (https://raw.githubusercontent.com/syv-ai/HyperQwen/main/docker/prepare.sh, https://huggingface.co/api/models/dbirks/Qwen3.8-27B-W4A16-AutoRound)", "The Docker image sets `VLLM_NO_USAGE_STATS=1` and `DO_NOT_TRACK=1`. The bare-metal launchers set neither, so a venv install keeps vLLM's default usage statistics (https://raw.githubusercontent.com/syv-ai/HyperQwen/main/Dockerfile)", "An AlphaSignal item indexed on hermes-ai.net on 19 September 2026 used the HyperQwen name and repeated the README's figures. It is a press lead, not a measurement (https://hermes-ai.net/news/hyperqwen-runs-qwen3-8-27b-on-a-single-rtx-3090-at-1-035-tok-s/)" ], "area": "models", "details": [ { "label": "Surface listed", "value": "The HyperQwen repository at main on 30 October 2026 (README, docs, PATCHES.md, Dockerfile, docker-compose.yml, .env.example, the launchers and the CI workflows), read from raw.githubusercontent.com. Nothing was installed or run" }, { "label": "Interfaces", "value": "vLLM 0.32.0's OpenAI-compatible HTTP server on port 18020 (`/v1`, `/health`, `/metrics`), served model name `qwen3.8-27b`. Docker Compose profiles `single`, `batch`, `solo` and `prepare`, or a bare-metal venv install. No MCP server and no SDK of its own" }, { "label": "Setups", "value": "A batch (about 1,035 tokens a second aggregate at 64 concurrent), B single-user default (127 a second, 64k context), C reproduction (381 a second while quoting, 56k), D long context (150k, about 95 to 100 a second), E huge context (240k, 67 a second mixed). Project figures on one RTX 3090 at 250 W" }, { "label": "Model files", "value": "dbirks/Qwen3.8-27B-W4A16-AutoRound (about 19.5 GB) with lm_head, embeddings and the MTP module requantised to int8 in place, plus syvai/qwen3.8-27b-3090-fast-variant and syvai/Qwen3.8-27B-DFlash2-W4A16 (about 1 GB each), all Apache-2.0. Qwen/Qwen3.8-27B is the base model" }, { "label": "Hardware", "value": "A 24 GB Ampere or newer NVIDIA GPU, a CUDA 13 driver (580 or later), about 40 GB of disk. Volta and older can't run it. Measured on compute capability 8.6, with community runs on sm80, sm89 and sm120 and at TP=2 and TP=4" }, { "label": "Tool calling", "value": "`--enable-auto-tool-choice --tool-call-parser qwen3_coder --tool-strict-level parameter` by default (`TOOLS=1`), `--reasoning-parser qwen3`, and a source build of xgrammar for bounded strings in tool schemas" }, { "label": "Network", "value": "Downloads from Hugging Face on first start and the image pull from ghcr.io. The Docker image turns off vLLM's usage statistics. The venv launchers leave vLLM's default on" }, { "label": "CI", "value": "Two GitHub workflows, `patch integrity` (about 40 GPU-less checks) and `docker image` (builds, applies every patch, runs build-time tests, pushes `latest` and `sha-\u003c7\u003e`, rebuilds weekly). Both badges read passing on 30 October 2026" }, { "label": "Repository activity", "value": "646 commits, 31 contributors, 0 open and 146 closed issues, 1 open and 276 closed pull requests, per shields.io on 30 October 2026. No tags, releases or CHANGELOG" } ], "provenance": { "legalEntity": "syv.ai (Copenhagen, CVR DK44671247)", "domain": "syv.ai", "domainRegistered": "", "endpointOnVendorDomain": null, "terms": "", "privacy": "https://syv.ai/privatlivspolitik", "statusPage": "", "changelog": "", "securityTxt": "none", "checked": "2026-10-30", "notes": [ "The Hugging Face organisation syvai, full name syv.ai, links https://syv.ai and github.com/syv-ai, and hosts the drafter files HyperQwen downloads. syv.ai's own pages don't mention HyperQwen.", "syv.ai's contact page gives the address in Copenhagen and CVR DK44671247. The Danish CVR register answered with a bot check and wasn't read.", "No terms, SECURITY.md or changelog for HyperQwen. The Apache-2.0 licence stands in for terms. The privacy policy listed is syv.ai's website policy and doesn't cover the software, which runs on the owner's machine.", "syv.ai answers 404 at /.well-known/security.txt and /security.txt." ], "score": 49, "checks": [ { "check": "Legal entity named", "value": "syv.ai (Copenhagen, CVR DK44671247)", "points": 20, "max": 20, "state": "ok" }, { "check": "Domain age", "value": "syv.ai, no registry record we could read", "points": 0, "max": 15, "state": "no" }, { "check": "Endpoint on the vendor's domain", "value": "no hosted endpoint", "points": 0, "max": 0, "state": "na" }, { "check": "Terms of service", "value": "nothing hosted, so the Apache-2.0 licence stands in", "points": 10, "max": 10, "state": "ok" }, { "check": "Privacy policy", "value": "read, states 0 of the 8 things a reader expects", "points": 4, "max": 10, "state": "part" }, { "check": "Status page", "value": "not found", "points": 0, "max": 10, "state": "no" }, { "check": "Changelog", "value": "not found", "points": 0, "max": 10, "state": "no" }, { "check": "security.txt", "value": "not found", "points": 0, "max": 10, "state": "no" }, { "check": "People named", "value": "2 named with their role, 2 with a public profile", "points": 10, "max": 10, "state": "ok" }, { "check": "Support route", "value": "help pages, a community", "points": 5, "max": 5, "state": "ok" } ], "policies": [ { "kind": "terms", "url": "", "state": "none-found", "points": 0, "max": 10 }, { "kind": "privacy", "url": "https://syv.ai/privatlivspolitik", "state": "read", "readAt": "2026-10-30", "words": 600, "points": 4, "max": 10, "expected": [ { "key": "privacy.date", "label": "Gives the date it was last updated", "found": false }, { "key": "privacy.collected", "label": "Says what personal data is collected", "found": false }, { "key": "privacy.retention", "label": "Says how long data is kept", "found": false }, { "key": "privacy.processors", "label": "Says who else receives the data", "found": false }, { "key": "privacy.sale", "label": "Says whether personal data is sold or shared for advertising", "found": false }, { "key": "privacy.rights", "label": "Says what rights people have over their data", "found": false }, { "key": "privacy.contact", "label": "Gives a privacy contact", "found": false }, { "key": "privacy.transfers", "label": "Says where data is transferred or stored", "found": false } ] } ] }, "humans": { "slug": "hyperqwen", "checked": "2026-10-30", "company": "syv.ai", "people": [ { "name": "Mads Henrichsen", "role": "Founder and partner, syv.ai", "founder": true, "linkedin": "https://www.linkedin.com/in/mhenrichsen/", "github": "https://github.com/mhenrichsen", "page": "/people/mads-henrichsen/" }, { "name": "Søren Søndergaard Poulsen", "role": "Partner, syv.ai", "linkedin": "https://www.linkedin.com/in/s%C3%B8ren-christian-s%C3%B8ndergaard-poulsen/", "page": "/people/soren-sondergaard-poulsen/" } ], "peopleNote": "HyperQwen sits in the syv-ai GitHub organisation. The Hugging Face organisation syvai, full name syv.ai, links https://syv.ai and github.com/syv-ai and hosts the drafter files HyperQwen downloads. syv.ai's team and contact pages (read 30 October 2026) name Mads Henrichsen as founder and partner and Søren Søndergaard Poulsen as partner, with their LinkedIn links. The GitHub link for Mads Henrichsen is the one on his own Hugging Face profile (mhenrichsen), which uploaded the syvai organisation's avatar. syv.ai's pages don't mention HyperQwen, and no first-hand page names its maintainers. The README's donation link is ko-fi.com/mhenrichsen, which wasn't read. shields.io counts 31 contributors, whose names weren't read because github.com wasn't fetched. Work email addresses and phone numbers on syv.ai are left out. No investors are listed. syv.ai's site and a web search found no funding round or named investor, and the Danish CVR register answered with a bot check.", "companySocial": { "github": "https://github.com/syv-ai" }, "support": { "urls": [ "https://github.com/syv-ai/HyperQwen/issues" ], "community": [ { "label": "Field report issue form", "url": "https://github.com/syv-ai/HyperQwen/issues/new?template=field-report.yml", "kind": "github" } ], "contactUrls": [ "https://syv.ai/kontakt" ] }, "sources": [ "https://raw.githubusercontent.com/syv-ai/HyperQwen/main/README.md", "https://syv.ai/team", "https://syv.ai/kontakt", "https://huggingface.co/syvai", "https://huggingface.co/api/organizations/syvai/overview", "https://huggingface.co/mhenrichsen", "https://huggingface.co/api/users/mhenrichsen/overview" ], "ownership": { "status": "private", "checked": "2026-10-30", "sources": [ "https://syv.ai/kontakt", "https://syv.ai/team" ] } }, "pageJsonUrl": "https://www.anchorterminal.com/tools/hyperqwen.json" } }